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 )
56 %************************************************************************
58 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
60 %************************************************************************
62 These notes describe how we implement specialisation to eliminate
65 The specialisation pass works on Core
66 syntax, complete with all the explicit dictionary application,
67 abstraction and construction as added by the type checker. The
68 existing type checker remains largely as it is.
70 One important thought: the {\em types} passed to an overloaded
71 function, and the {\em dictionaries} passed are mutually redundant.
72 If the same function is applied to the same type(s) then it is sure to
73 be applied to the same dictionary(s)---or rather to the same {\em
74 values}. (The arguments might look different but they will evaluate
77 Second important thought: we know that we can make progress by
78 treating dictionary arguments as static and worth specialising on. So
79 we can do without binding-time analysis, and instead specialise on
80 dictionary arguments and no others.
89 and suppose f is overloaded.
91 STEP 1: CALL-INSTANCE COLLECTION
93 We traverse <body>, accumulating all applications of f to types and
96 (Might there be partial applications, to just some of its types and
97 dictionaries? In principle yes, but in practice the type checker only
98 builds applications of f to all its types and dictionaries, so partial
99 applications could only arise as a result of transformation, and even
100 then I think it's unlikely. In any case, we simply don't accumulate such
101 partial applications.)
106 So now we have a collection of calls to f:
110 Notice that f may take several type arguments. To avoid ambiguity, we
111 say that f is called at type t1/t2 and t3/t4.
113 We take equivalence classes using equality of the *types* (ignoring
114 the dictionary args, which as mentioned previously are redundant).
116 STEP 3: SPECIALISATION
118 For each equivalence class, choose a representative (f t1 t2 d1 d2),
119 and create a local instance of f, defined thus:
121 f@t1/t2 = <f_rhs> t1 t2 d1 d2
123 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
124 of simplification will now result. However we don't actually *do* that
125 simplification. Rather, we leave it for the simplifier to do. If we
126 *did* do it, though, we'd get more call instances from the specialised
127 RHS. We can work out what they are by instantiating the call-instance
128 set from f's RHS with the types t1, t2.
130 Add this new id to f's IdInfo, to record that f has a specialised version.
132 Before doing any of this, check that f's IdInfo doesn't already
133 tell us about an existing instance of f at the required type/s.
134 (This might happen if specialisation was applied more than once, or
135 it might arise from user SPECIALIZE pragmas.)
139 Wait a minute! What if f is recursive? Then we can't just plug in
140 its right-hand side, can we?
142 But it's ok. The type checker *always* creates non-recursive definitions
143 for overloaded recursive functions. For example:
145 f x = f (x+x) -- Yes I know its silly
149 f a (d::Num a) = let p = +.sel a d
151 letrec fl (y::a) = fl (p y y)
155 We still have recusion for non-overloaded functions which we
156 speciailise, but the recursive call should get specialised to the
157 same recursive version.
163 All this is crystal clear when the function is applied to *constant
164 types*; that is, types which have no type variables inside. But what if
165 it is applied to non-constant types? Suppose we find a call of f at type
166 t1/t2. There are two possibilities:
168 (a) The free type variables of t1, t2 are in scope at the definition point
169 of f. In this case there's no problem, we proceed just as before. A common
170 example is as follows. Here's the Haskell:
175 After typechecking we have
177 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
178 in +.sel a d (f a d y) (f a d y)
180 Notice that the call to f is at type type "a"; a non-constant type.
181 Both calls to f are at the same type, so we can specialise to give:
183 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
184 in +.sel a d (f@a y) (f@a y)
187 (b) The other case is when the type variables in the instance types
188 are *not* in scope at the definition point of f. The example we are
189 working with above is a good case. There are two instances of (+.sel a d),
190 but "a" is not in scope at the definition of +.sel. Can we do anything?
191 Yes, we can "common them up", a sort of limited common sub-expression deal.
194 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
195 f@a (x::a) = +.sel@a x x
196 in +.sel@a (f@a y) (f@a y)
198 This can save work, and can't be spotted by the type checker, because
199 the two instances of +.sel weren't originally at the same type.
203 * There are quite a few variations here. For example, the defn of
204 +.sel could be floated ouside the \y, to attempt to gain laziness.
205 It certainly mustn't be floated outside the \d because the d has to
208 * We don't want to inline f_rhs in this case, because
209 that will duplicate code. Just commoning up the call is the point.
211 * Nothing gets added to +.sel's IdInfo.
213 * Don't bother unless the equivalence class has more than one item!
215 Not clear whether this is all worth it. It is of course OK to
216 simply discard call-instances when passing a big lambda.
218 Polymorphism 2 -- Overloading
220 Consider a function whose most general type is
222 f :: forall a b. Ord a => [a] -> b -> b
224 There is really no point in making a version of g at Int/Int and another
225 at Int/Bool, because it's only instancing the type variable "a" which
226 buys us any efficiency. Since g is completely polymorphic in b there
227 ain't much point in making separate versions of g for the different
230 That suggests that we should identify which of g's type variables
231 are constrained (like "a") and which are unconstrained (like "b").
232 Then when taking equivalence classes in STEP 2, we ignore the type args
233 corresponding to unconstrained type variable. In STEP 3 we make
234 polymorphic versions. Thus:
236 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
245 f a (d::Num a) = let g = ...
247 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
249 Here, g is only called at one type, but the dictionary isn't in scope at the
250 definition point for g. Usually the type checker would build a
251 definition for d1 which enclosed g, but the transformation system
252 might have moved d1's defn inward. Solution: float dictionary bindings
253 outwards along with call instances.
257 f x = let g p q = p==q
263 Before specialisation, leaving out type abstractions we have
265 f df x = let g :: Eq a => a -> a -> Bool
267 h :: Num a => a -> a -> (a, Bool)
268 h dh r s = let deq = eqFromNum dh
269 in (+ dh r s, g deq r s)
273 After specialising h we get a specialised version of h, like this:
275 h' r s = let deq = eqFromNum df
276 in (+ df r s, g deq r s)
278 But we can't naively make an instance for g from this, because deq is not in scope
279 at the defn of g. Instead, we have to float out the (new) defn of deq
280 to widen its scope. Notice that this floating can't be done in advance -- it only
281 shows up when specialisation is done.
283 User SPECIALIZE pragmas
284 ~~~~~~~~~~~~~~~~~~~~~~~
285 Specialisation pragmas can be digested by the type checker, and implemented
286 by adding extra definitions along with that of f, in the same way as before
288 f@t1/t2 = <f_rhs> t1 t2 d1 d2
290 Indeed the pragmas *have* to be dealt with by the type checker, because
291 only it knows how to build the dictionaries d1 and d2! For example
293 g :: Ord a => [a] -> [a]
294 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
296 Here, the specialised version of g is an application of g's rhs to the
297 Ord dictionary for (Tree Int), which only the type checker can conjure
298 up. There might not even *be* one, if (Tree Int) is not an instance of
299 Ord! (All the other specialision has suitable dictionaries to hand
302 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
303 it is buried in a complex (as-yet-un-desugared) binding group.
306 f@t1/t2 = f* t1 t2 d1 d2
308 where f* is the Id f with an IdInfo which says "inline me regardless!".
309 Indeed all the specialisation could be done in this way.
310 That in turn means that the simplifier has to be prepared to inline absolutely
311 any in-scope let-bound thing.
314 Again, the pragma should permit polymorphism in unconstrained variables:
316 h :: Ord a => [a] -> b -> b
317 {-# SPECIALIZE h :: [Int] -> b -> b #-}
319 We *insist* that all overloaded type variables are specialised to ground types,
320 (and hence there can be no context inside a SPECIALIZE pragma).
321 We *permit* unconstrained type variables to be specialised to
323 - or left as a polymorphic type variable
324 but nothing in between. So
326 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
328 is *illegal*. (It can be handled, but it adds complication, and gains the
332 SPECIALISING INSTANCE DECLARATIONS
333 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
336 instance Foo a => Foo [a] where
338 {-# SPECIALIZE instance Foo [Int] #-}
340 The original instance decl creates a dictionary-function
343 dfun.Foo.List :: forall a. Foo a -> Foo [a]
345 The SPECIALIZE pragma just makes a specialised copy, just as for
346 ordinary function definitions:
348 dfun.Foo.List@Int :: Foo [Int]
349 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
351 The information about what instance of the dfun exist gets added to
352 the dfun's IdInfo in the same way as a user-defined function too.
355 Automatic instance decl specialisation?
356 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
357 Can instance decls be specialised automatically? It's tricky.
358 We could collect call-instance information for each dfun, but
359 then when we specialised their bodies we'd get new call-instances
360 for ordinary functions; and when we specialised their bodies, we might get
361 new call-instances of the dfuns, and so on. This all arises because of
362 the unrestricted mutual recursion between instance decls and value decls.
364 Still, there's no actual problem; it just means that we may not do all
365 the specialisation we could theoretically do.
367 Furthermore, instance decls are usually exported and used non-locally,
368 so we'll want to compile enough to get those specialisations done.
370 Lastly, there's no such thing as a local instance decl, so we can
371 survive solely by spitting out *usage* information, and then reading that
372 back in as a pragma when next compiling the file. So for now,
373 we only specialise instance decls in response to pragmas.
376 SPITTING OUT USAGE INFORMATION
377 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
379 To spit out usage information we need to traverse the code collecting
380 call-instance information for all imported (non-prelude?) functions
381 and data types. Then we equivalence-class it and spit it out.
383 This is done at the top-level when all the call instances which escape
384 must be for imported functions and data types.
386 *** Not currently done ***
389 Partial specialisation by pragmas
390 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
391 What about partial specialisation:
393 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
394 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
398 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
400 Seems quite reasonable. Similar things could be done with instance decls:
402 instance (Foo a, Foo b) => Foo (a,b) where
404 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
405 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
407 Ho hum. Things are complex enough without this. I pass.
410 Requirements for the simplifer
411 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
412 The simplifier has to be able to take advantage of the specialisation.
414 * When the simplifier finds an application of a polymorphic f, it looks in
415 f's IdInfo in case there is a suitable instance to call instead. This converts
417 f t1 t2 d1 d2 ===> f_t1_t2
419 Note that the dictionaries get eaten up too!
421 * Dictionary selection operations on constant dictionaries must be
424 +.sel Int d ===> +Int
426 The obvious way to do this is in the same way as other specialised
427 calls: +.sel has inside it some IdInfo which tells that if it's applied
428 to the type Int then it should eat a dictionary and transform to +Int.
430 In short, dictionary selectors need IdInfo inside them for constant
433 * Exactly the same applies if a superclass dictionary is being
436 Eq.sel Int d ===> dEqInt
438 * Something similar applies to dictionary construction too. Suppose
439 dfun.Eq.List is the function taking a dictionary for (Eq a) to
440 one for (Eq [a]). Then we want
442 dfun.Eq.List Int d ===> dEq.List_Int
444 Where does the Eq [Int] dictionary come from? It is built in
445 response to a SPECIALIZE pragma on the Eq [a] instance decl.
447 In short, dfun Ids need IdInfo with a specialisation for each
448 constant instance of their instance declaration.
450 All this uses a single mechanism: the SpecEnv inside an Id
453 What does the specialisation IdInfo look like?
454 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
456 The SpecEnv of an Id maps a list of types (the template) to an expression
460 For example, if f has this SpecInfo:
462 [Int, a] -> \d:Ord Int. f' a
464 it means that we can replace the call
466 f Int t ===> (\d. f' t)
468 This chucks one dictionary away and proceeds with the
469 specialised version of f, namely f'.
472 What can't be done this way?
473 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
474 There is no way, post-typechecker, to get a dictionary for (say)
475 Eq a from a dictionary for Eq [a]. So if we find
479 we can't transform to
484 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
486 Of course, we currently have no way to automatically derive
487 eqList, nor to connect it to the Eq [a] instance decl, but you
488 can imagine that it might somehow be possible. Taking advantage
489 of this is permanently ruled out.
491 Still, this is no great hardship, because we intend to eliminate
492 overloading altogether anyway!
496 A note about non-tyvar dictionaries
497 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
498 Some Ids have types like
500 forall a,b,c. Eq a -> Ord [a] -> tau
502 This seems curious at first, because we usually only have dictionary
503 args whose types are of the form (C a) where a is a type variable.
504 But this doesn't hold for the functions arising from instance decls,
505 which sometimes get arguements with types of form (C (T a)) for some
508 Should we specialise wrt this compound-type dictionary? We used to say
510 "This is a heuristic judgement, as indeed is the fact that we
511 specialise wrt only dictionaries. We choose *not* to specialise
512 wrt compound dictionaries because at the moment the only place
513 they show up is in instance decls, where they are simply plugged
514 into a returned dictionary. So nothing is gained by specialising
517 But it is simpler and more uniform to specialise wrt these dicts too;
518 and in future GHC is likely to support full fledged type signatures
520 f ;: Eq [(a,b)] => ...
523 %************************************************************************
525 \subsubsection{The new specialiser}
527 %************************************************************************
529 Our basic game plan is this. For let(rec) bound function
530 f :: (C a, D c) => (a,b,c,d) -> Bool
532 * Find any specialised calls of f, (f ts ds), where
533 ts are the type arguments t1 .. t4, and
534 ds are the dictionary arguments d1 .. d2.
536 * Add a new definition for f1 (say):
538 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
540 Note that we abstract over the unconstrained type arguments.
544 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
546 to the specialisations of f. This will be used by the
547 simplifier to replace calls
548 (f t1 t2 t3 t4) da db
550 (\d1 d1 -> f1 t2 t4) da db
552 All the stuff about how many dictionaries to discard, and what types
553 to apply the specialised function to, are handled by the fact that the
554 SpecEnv contains a template for the result of the specialisation.
556 We don't build *partial* specialisations for f. For example:
558 f :: Eq a => a -> a -> Bool
559 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
561 Here, little is gained by making a specialised copy of f.
562 There's a distinct danger that the specialised version would
563 first build a dictionary for (Eq b, Eq c), and then select the (==)
564 method from it! Even if it didn't, not a great deal is saved.
566 We do, however, generate polymorphic, but not overloaded, specialisations:
568 f :: Eq a => [a] -> b -> b -> b
569 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
571 Hence, the invariant is this:
573 *** no specialised version is overloaded ***
576 %************************************************************************
578 \subsubsection{The exported function}
580 %************************************************************************
583 specProgram :: DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind]
584 specProgram dflags us binds = do
586 showPass dflags "Specialise"
588 let binds' = initSM us (do (binds', uds') <- go binds
589 return (dumpAllDictBinds uds' binds'))
591 endPass dflags "Specialise" Opt_D_dump_spec binds'
593 dumpIfSet_dyn dflags Opt_D_dump_rules "Top-level specialisations"
594 (pprRules (tidyRules emptyTidyEnv (rulesOfBinds binds')))
598 -- We need to start with a Subst that knows all the things
599 -- that are in scope, so that the substitution engine doesn't
600 -- accidentally re-use a unique that's already in use
601 -- Easiest thing is to do it all at once, as if all the top-level
602 -- decls were mutually recursive
603 top_subst = mkEmptySubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
605 go [] = return ([], emptyUDs)
606 go (bind:binds) = do (binds', uds) <- go binds
607 (bind', uds') <- specBind top_subst bind uds
608 return (bind' ++ binds', uds')
611 %************************************************************************
613 \subsubsection{@specExpr@: the main function}
615 %************************************************************************
618 specVar :: Subst -> Id -> CoreExpr
619 specVar subst v = lookupIdSubst subst v
621 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
622 -- We carry a substitution down:
623 -- a) we must clone any binding that might flaot outwards,
624 -- to avoid name clashes
625 -- b) we carry a type substitution to use when analysing
626 -- the RHS of specialised bindings (no type-let!)
628 ---------------- First the easy cases --------------------
629 specExpr subst (Type ty) = return (Type (substTy subst ty), emptyUDs)
630 specExpr subst (Var v) = return (specVar subst v, emptyUDs)
631 specExpr subst (Lit lit) = return (Lit lit, emptyUDs)
632 specExpr subst (Cast e co) = do
633 (e', uds) <- specExpr subst e
634 return ((Cast e' (substTy subst co)), uds)
635 specExpr subst (Note note body) = do
636 (body', uds) <- specExpr subst body
637 return (Note (specNote subst note) body', uds)
640 ---------------- Applications might generate a call instance --------------------
641 specExpr subst expr@(App fun arg)
644 go (App fun arg) args = do (arg', uds_arg) <- specExpr subst arg
645 (fun', uds_app) <- go fun (arg':args)
646 return (App fun' arg', uds_arg `plusUDs` uds_app)
648 go (Var f) args = case specVar subst f of
649 Var f' -> return (Var f', mkCallUDs subst f' args)
650 e' -> return (e', emptyUDs) -- I don't expect this!
651 go other args = specExpr subst other
653 ---------------- Lambda/case require dumping of usage details --------------------
654 specExpr subst e@(Lam _ _) = do
655 (body', uds) <- specExpr subst' body
656 let (filtered_uds, body'') = dumpUDs bndrs' uds body'
657 return (mkLams bndrs' body'', filtered_uds)
659 (bndrs, body) = collectBinders e
660 (subst', bndrs') = substBndrs subst bndrs
661 -- More efficient to collect a group of binders together all at once
662 -- and we don't want to split a lambda group with dumped bindings
664 specExpr subst (Case scrut case_bndr ty alts) = do
665 (scrut', uds_scrut) <- specExpr subst scrut
666 (alts', uds_alts) <- mapAndCombineSM spec_alt alts
667 return (Case scrut' case_bndr' (substTy subst ty) alts', uds_scrut `plusUDs` uds_alts)
669 (subst_alt, case_bndr') = substBndr subst case_bndr
670 -- No need to clone case binder; it can't float like a let(rec)
672 spec_alt (con, args, rhs) = do
673 (rhs', uds) <- specExpr subst_rhs rhs
674 let (uds', rhs'') = dumpUDs args uds rhs'
675 return ((con, args', rhs''), uds')
677 (subst_rhs, args') = substBndrs subst_alt args
679 ---------------- Finally, let is the interesting case --------------------
680 specExpr subst (Let bind body) = do
682 (rhs_subst, body_subst, bind') <- cloneBindSM subst bind
684 -- Deal with the body
685 (body', body_uds) <- specExpr body_subst body
687 -- Deal with the bindings
688 (binds', uds) <- specBind rhs_subst bind' body_uds
691 return (foldr Let body' binds', uds)
693 -- Must apply the type substitution to coerceions
694 specNote subst note = note
697 %************************************************************************
699 \subsubsection{Dealing with a binding}
701 %************************************************************************
704 specBind :: Subst -- Use this for RHSs
706 -> UsageDetails -- Info on how the scope of the binding
707 -> SpecM ([CoreBind], -- New bindings
708 UsageDetails) -- And info to pass upstream
710 specBind rhs_subst bind body_uds = do
711 (bind', bind_uds) <- specBindItself rhs_subst bind (calls body_uds)
713 bndrs = bindersOf bind
714 all_uds = zapCalls bndrs (body_uds `plusUDs` bind_uds)
715 -- It's important that the `plusUDs` is this way round,
716 -- because body_uds may bind dictionaries that are
717 -- used in the calls passed to specDefn. So the
718 -- dictionary bindings in bind_uds may mention
719 -- dictionaries bound in body_uds.
720 case splitUDs bndrs all_uds of
722 (_, ([],[])) -- This binding doesn't bind anything needed
723 -- in the UDs, so put the binding here
724 -- This is the case for most non-dict bindings, except
725 -- for the few that are mentioned in a dict binding
726 -- that is floating upwards in body_uds
727 -> return ([bind'], all_uds)
729 (float_uds, (dict_binds, calls)) -- This binding is needed in the UDs, so float it out
730 -> return ([], float_uds `plusUDs` mkBigUD bind' dict_binds calls)
733 -- A truly gruesome function
734 mkBigUD bind@(NonRec _ _) dbs calls
735 = -- Common case: non-recursive and no specialisations
736 -- (if there were any specialistions it would have been made recursive)
737 MkUD { dict_binds = listToBag (mkDB bind : dbs),
738 calls = listToCallDetails calls }
740 mkBigUD bind dbs calls
742 MkUD { dict_binds = unitBag (mkDB (Rec (bind_prs bind ++ dbsToPairs dbs))),
744 calls = listToCallDetails calls }
746 bind_prs (NonRec b r) = [(b,r)]
747 bind_prs (Rec prs) = prs
750 dbsToPairs ((bind,_):dbs) = bind_prs bind ++ dbsToPairs dbs
752 -- specBindItself deals with the RHS, specialising it according
753 -- to the calls found in the body (if any)
754 specBindItself rhs_subst (NonRec bndr rhs) call_info = do
755 ((bndr',rhs'), spec_defns, spec_uds) <- specDefn rhs_subst call_info (bndr,rhs)
757 new_bind | null spec_defns = NonRec bndr' rhs'
758 | otherwise = Rec ((bndr',rhs'):spec_defns)
759 -- bndr' mentions the spec_defns in its SpecEnv
760 -- Not sure why we couln't just put the spec_defns first
761 return (new_bind, spec_uds)
763 specBindItself rhs_subst (Rec pairs) call_info = do
764 stuff <- mapM (specDefn rhs_subst call_info) pairs
766 (pairs', spec_defns_s, spec_uds_s) = unzip3 stuff
767 spec_defns = concat spec_defns_s
768 spec_uds = plusUDList spec_uds_s
769 new_bind = Rec (spec_defns ++ pairs')
770 return (new_bind, spec_uds)
773 specDefn :: Subst -- Subst to use for RHS
774 -> CallDetails -- Info on how it is used in its scope
775 -> (Id, CoreExpr) -- The thing being bound and its un-processed RHS
776 -> SpecM ((Id, CoreExpr), -- The thing and its processed RHS
777 -- the Id may now have specialisations attached
778 [(Id,CoreExpr)], -- Extra, specialised bindings
779 UsageDetails -- Stuff to fling upwards from the RHS and its
780 ) -- specialised versions
782 specDefn subst calls (fn, rhs)
783 -- The first case is the interesting one
784 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
785 && rhs_ids `lengthAtLeast` n_dicts -- and enough dict args
786 && notNull calls_for_me -- And there are some calls to specialise
788 -- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
789 -- See Note [Inline specialisation] for why we do not
790 -- switch off specialisation for inline functions = do
792 -- Specialise the body of the function
793 (rhs', rhs_uds) <- specExpr subst rhs
795 -- Make a specialised version for each call in calls_for_me
796 stuff <- mapM spec_call calls_for_me
798 (spec_defns, spec_uds, spec_rules) = unzip3 stuff
800 fn' = addIdSpecialisations fn spec_rules
804 rhs_uds `plusUDs` plusUDList spec_uds)
806 | otherwise -- No calls or RHS doesn't fit our preconceptions
807 = WARN( notNull calls_for_me, ptext SLIT("Missed specialisation opportunity for") <+> ppr fn )
808 -- Note [Specialisation shape]
809 (do { (rhs', rhs_uds) <- specExpr subst rhs
810 ; return ((fn, rhs'), [], rhs_uds) })
814 (tyvars, theta, _) = tcSplitSigmaTy fn_type
815 n_tyvars = length tyvars
816 n_dicts = length theta
817 inline_prag = idInlinePragma fn
819 -- It's important that we "see past" any INLINE pragma
820 -- else we'll fail to specialise an INLINE thing
821 (inline_rhs, rhs_inside) = dropInline rhs
822 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs_inside
824 rhs_dicts = take n_dicts rhs_ids
825 rhs_bndrs = rhs_tyvars ++ rhs_dicts
826 body = mkLams (drop n_dicts rhs_ids) rhs_body
827 -- Glue back on the non-dict lambdas
829 calls_for_me = case lookupFM calls fn of
831 Just cs -> fmToList cs
833 ----------------------------------------------------------
834 -- Specialise to one particular call pattern
835 spec_call :: (CallKey, ([DictExpr], VarSet)) -- Call instance
836 -> SpecM ((Id,CoreExpr), -- Specialised definition
837 UsageDetails, -- Usage details from specialised body
838 CoreRule) -- Info for the Id's SpecEnv
839 spec_call (CallKey call_ts, (call_ds, call_fvs))
840 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts ) do
841 -- Calls are only recorded for properly-saturated applications
843 -- Suppose f's defn is f = /\ a b c d -> \ d1 d2 -> rhs
844 -- Supppose the call is for f [Just t1, Nothing, Just t3, Nothing] [dx1, dx2]
846 -- Construct the new binding
847 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b d -> rhs)
848 -- PLUS the usage-details
849 -- { d1' = dx1; d2' = dx2 }
850 -- where d1', d2' are cloned versions of d1,d2, with the type substitution applied.
852 -- Note that the substitution is applied to the whole thing.
853 -- This is convenient, but just slightly fragile. Notably:
854 -- * There had better be no name clashes in a/b/c/d
857 -- poly_tyvars = [b,d] in the example above
858 -- spec_tyvars = [a,c]
859 -- ty_args = [t1,b,t3,d]
860 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
861 spec_tyvars = [tv | (tv, Just _) <- rhs_tyvars `zip` call_ts]
862 ty_args = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
864 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
865 mk_ty_arg rhs_tyvar (Just ty) = Type ty
866 rhs_subst = extendTvSubstList subst (spec_tyvars `zip` [ty | Just ty <- call_ts])
868 (rhs_subst', rhs_dicts') <- cloneBinders rhs_subst rhs_dicts
870 inst_args = ty_args ++ map Var rhs_dicts'
872 -- Figure out the type of the specialised function
873 body_ty = applyTypeToArgs rhs fn_type inst_args
874 (lam_args, app_args) -- Add a dummy argument if body_ty is unlifted
875 | isUnLiftedType body_ty -- C.f. WwLib.mkWorkerArgs
876 = (poly_tyvars ++ [voidArgId], poly_tyvars ++ [realWorldPrimId])
877 | otherwise = (poly_tyvars, poly_tyvars)
878 spec_id_ty = mkPiTypes lam_args body_ty
880 spec_f <- newIdSM fn spec_id_ty
881 (spec_rhs, rhs_uds) <- specExpr rhs_subst' (mkLams lam_args body)
883 -- The rule to put in the function's specialisation is:
884 -- forall b,d, d1',d2'. f t1 b t3 d d1' d2' = f1 b d
885 spec_env_rule = mkLocalRule (mkFastString ("SPEC " ++ showSDoc (ppr fn)))
886 inline_prag -- Note [Auto-specialisation and RULES]
888 (poly_tyvars ++ rhs_dicts')
890 (mkVarApps (Var spec_f) app_args)
892 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
893 final_uds = foldr addDictBind rhs_uds (my_zipEqual "spec_call" rhs_dicts' call_ds)
895 spec_pr | inline_rhs = (spec_f `setInlinePragma` inline_prag, Note InlineMe spec_rhs)
896 | otherwise = (spec_f, spec_rhs)
898 return (spec_pr, final_uds, spec_env_rule)
901 my_zipEqual doc xs ys
902 | debugIsOn && not (equalLength xs ys)
903 = pprPanic "my_zipEqual" (vcat
905 , ppr fn <+> ppr call_ts
906 , ppr (idType fn), ppr theta
907 , ppr n_dicts, ppr rhs_dicts
909 | otherwise = zipEqual doc xs ys
912 Note [Auto-specialisation and RULES]
913 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
918 f :: (Int -> Int) -> Int
922 Suppose that auto-specialisation makes a specialised version of
923 g::Int->Int That version won't appear in the LHS of the RULE for f.
924 So if the specialisation rule fires too early, the rule for f may
927 It might be possible to add new rules, to "complete" the rewrite system.
929 RULE forall d. g Int d = g_spec
933 But that's a bit complicated. For now we ask the programmer's help,
934 by *copying the INLINE activation pragma* to the auto-specialised rule.
935 So if g says {-# NOINLINE[2] g #-}, then the auto-spec rule will also
936 not be active until phase 2.
939 Note [Specialisation shape]
940 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
941 We only specialise a function if it has visible top-level lambdas
942 corresponding to its overloading. E.g. if
943 f :: forall a. Eq a => ....
944 then its body must look like
947 Reason: when specialising the body for a call (f ty dexp), we want to
948 substitute dexp for d, and pick up specialised calls in the body of f.
950 This doesn't always work. One example I came across was htis:
951 newtype Gen a = MkGen{ unGen :: Int -> a }
953 choose :: Eq a => a -> Gen a
954 choose n = MkGen (\r -> n)
956 oneof = choose (1::Int)
958 It's a silly exapmle, but we get
959 choose = /\a. g `cast` co
960 where choose doesn't have any dict arguments. Thus far I have not
961 tried to fix this (wait till there's a real example).
964 Note [Inline specialisations]
965 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
966 We transfer to the specialised function any INLINE stuff from the
967 original. This means (a) the Activation in the IdInfo, and (b) any
970 This is a change (Jun06). Previously the idea is that the point of
971 inlining was precisely to specialise the function at its call site,
972 and that's not so important for the specialised copies. But
973 *pragma-directed* specialisation now takes place in the
974 typechecker/desugarer, with manually specified INLINEs. The
975 specialiation here is automatic. It'd be very odd if a function
976 marked INLINE was specialised (because of some local use), and then
977 forever after (including importing modules) the specialised version
978 wasn't INLINEd. After all, the programmer said INLINE!
980 You might wonder why we don't just not specialise INLINE functions.
981 It's because even INLINE functions are sometimes not inlined, when
982 they aren't applied to interesting arguments. But perhaps the type
983 arguments alone are enough to specialise (even though the args are too
984 boring to trigger inlining), and it's certainly better to call the
987 A case in point is dictionary functions, which are current marked
988 INLINE, but which are worth specialising.
991 dropInline :: CoreExpr -> (Bool, CoreExpr)
992 dropInline (Note InlineMe rhs) = (True, rhs)
993 dropInline rhs = (False, rhs)
996 %************************************************************************
998 \subsubsection{UsageDetails and suchlike}
1000 %************************************************************************
1005 dict_binds :: !(Bag DictBind),
1006 -- Floated dictionary bindings
1007 -- The order is important;
1008 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
1009 -- (Remember, Bags preserve order in GHC.)
1011 calls :: !CallDetails
1014 type DictBind = (CoreBind, VarSet)
1015 -- The set is the free vars of the binding
1016 -- both tyvars and dicts
1018 type DictExpr = CoreExpr
1020 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM }
1022 type ProtoUsageDetails = ([DictBind],
1023 [(Id, CallKey, ([DictExpr], VarSet))]
1026 ------------------------------------------------------------
1027 type CallDetails = FiniteMap Id CallInfo
1028 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1029 type CallInfo = FiniteMap CallKey
1030 ([DictExpr], VarSet) -- Dict args and the vars of the whole
1031 -- call (including tyvars)
1032 -- [*not* include the main id itself, of course]
1033 -- The finite maps eliminate duplicates
1034 -- The list of types and dictionaries is guaranteed to
1035 -- match the type of f
1037 -- Type isn't an instance of Ord, so that we can control which
1038 -- instance we use. That's tiresome here. Oh well
1039 instance Eq CallKey where
1040 k1 == k2 = case k1 `compare` k2 of { EQ -> True; other -> False }
1042 instance Ord CallKey where
1043 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1045 cmp Nothing Nothing = EQ
1046 cmp Nothing (Just t2) = LT
1047 cmp (Just t1) Nothing = GT
1048 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1050 unionCalls :: CallDetails -> CallDetails -> CallDetails
1051 unionCalls c1 c2 = plusFM_C plusFM c1 c2
1053 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> CallDetails
1054 singleCall id tys dicts
1055 = unitFM id (unitFM (CallKey tys) (dicts, call_fvs))
1057 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1058 tys_fvs = tyVarsOfTypes (catMaybes tys)
1059 -- The type args (tys) are guaranteed to be part of the dictionary
1060 -- types, because they are just the constrained types,
1061 -- and the dictionary is therefore sure to be bound
1062 -- inside the binding for any type variables free in the type;
1063 -- hence it's safe to neglect tyvars free in tys when making
1064 -- the free-var set for this call
1065 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1067 -- We don't include the 'id' itself.
1069 listToCallDetails calls
1070 = foldr (unionCalls . mk_call) emptyFM calls
1072 mk_call (id, tys, dicts_w_fvs) = unitFM id (unitFM tys dicts_w_fvs)
1073 -- NB: the free vars of the call are provided
1075 callDetailsToList calls = [ (id,tys,dicts)
1076 | (id,fm) <- fmToList calls,
1077 (tys, dicts) <- fmToList fm
1080 mkCallUDs subst f args
1082 || not (all isClassPred theta)
1083 -- Only specialise if all overloading is on class params.
1084 -- In ptic, with implicit params, the type args
1085 -- *don't* say what the value of the implicit param is!
1086 || not (spec_tys `lengthIs` n_tyvars)
1087 || not ( dicts `lengthIs` n_dicts)
1088 || maybeToBool (lookupRule (\act -> True) (substInScope subst) emptyRuleBase f args)
1089 -- There's already a rule covering this call. A typical case
1090 -- is where there's an explicit user-provided rule. Then
1091 -- we don't want to create a specialised version
1092 -- of the function that overlaps.
1093 = emptyUDs -- Not overloaded, or no specialisation wanted
1096 = MkUD {dict_binds = emptyBag,
1097 calls = singleCall f spec_tys dicts
1100 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1101 constrained_tyvars = tyVarsOfTheta theta
1102 n_tyvars = length tyvars
1103 n_dicts = length theta
1105 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1106 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1109 | tyvar `elemVarSet` constrained_tyvars = Just ty
1110 | otherwise = Nothing
1112 ------------------------------------------------------------
1113 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1114 plusUDs (MkUD {dict_binds = db1, calls = calls1})
1115 (MkUD {dict_binds = db2, calls = calls2})
1116 = MkUD {dict_binds = d, calls = c}
1118 d = db1 `unionBags` db2
1119 c = calls1 `unionCalls` calls2
1121 plusUDList = foldr plusUDs emptyUDs
1123 -- zapCalls deletes calls to ids from uds
1124 zapCalls ids uds = uds {calls = delListFromFM (calls uds) ids}
1126 mkDB bind = (bind, bind_fvs bind)
1128 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1129 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1132 rhs_fvs = unionVarSets (map pair_fvs prs)
1134 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1135 -- Don't forget variables mentioned in the
1136 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1137 -- Also tyvars mentioned in its type; they may not appear in the RHS
1141 addDictBind (dict,rhs) uds = uds { dict_binds = mkDB (NonRec dict rhs) `consBag` dict_binds uds }
1143 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
1144 = foldrBag add binds dbs
1146 add (bind,_) binds = bind : binds
1148 dumpUDs :: [CoreBndr]
1149 -> UsageDetails -> CoreExpr
1150 -> (UsageDetails, CoreExpr)
1151 dumpUDs bndrs uds body
1152 = (free_uds, foldr add_let body dict_binds)
1154 (free_uds, (dict_binds, _)) = splitUDs bndrs uds
1155 add_let (bind,_) body = Let bind body
1157 splitUDs :: [CoreBndr]
1159 -> (UsageDetails, -- These don't mention the binders
1160 ProtoUsageDetails) -- These do
1162 splitUDs bndrs uds@(MkUD {dict_binds = orig_dbs,
1163 calls = orig_calls})
1165 = if isEmptyBag dump_dbs && null dump_calls then
1166 -- Common case: binder doesn't affect floats
1170 -- Binders bind some of the fvs of the floats
1171 (MkUD {dict_binds = free_dbs,
1172 calls = listToCallDetails free_calls},
1173 (bagToList dump_dbs, dump_calls)
1177 bndr_set = mkVarSet bndrs
1179 (free_dbs, dump_dbs, dump_idset)
1180 = foldlBag dump_db (emptyBag, emptyBag, bndr_set) orig_dbs
1181 -- Important that it's foldl not foldr;
1182 -- we're accumulating the set of dumped ids in dump_set
1184 -- Filter out any calls that mention things that are being dumped
1185 orig_call_list = callDetailsToList orig_calls
1186 (dump_calls, free_calls) = partition captured orig_call_list
1187 captured (id,tys,(dicts, fvs)) = fvs `intersectsVarSet` dump_idset
1188 || id `elemVarSet` dump_idset
1190 dump_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1191 | dump_idset `intersectsVarSet` fvs -- Dump it
1192 = (free_dbs, dump_dbs `snocBag` db,
1193 extendVarSetList dump_idset (bindersOf bind))
1195 | otherwise -- Don't dump it
1196 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1200 %************************************************************************
1202 \subsubsection{Boring helper functions}
1204 %************************************************************************
1207 type SpecM a = UniqSM a
1211 mapAndCombineSM f [] = return ([], emptyUDs)
1212 mapAndCombineSM f (x:xs) = do (y, uds1) <- f x
1213 (ys, uds2) <- mapAndCombineSM f xs
1214 return (y:ys, uds1 `plusUDs` uds2)
1216 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1217 -- Clone the binders of the bind; return new bind with the cloned binders
1218 -- Return the substitution to use for RHSs, and the one to use for the body
1219 cloneBindSM subst (NonRec bndr rhs) = do
1220 us <- getUniqueSupplyM
1221 let (subst', bndr') = cloneIdBndr subst us bndr
1222 return (subst, subst', NonRec bndr' rhs)
1224 cloneBindSM subst (Rec pairs) = do
1225 us <- getUniqueSupplyM
1226 let (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1227 return (subst', subst', Rec (bndrs' `zip` map snd pairs))
1229 cloneBinders subst bndrs = do
1230 us <- getUniqueSupplyM
1231 return (cloneIdBndrs subst us bndrs)
1233 newIdSM old_id new_ty = do
1236 -- Give the new Id a similar occurrence name to the old one
1237 name = idName old_id
1238 new_id = mkUserLocal (mkSpecOcc (nameOccName name)) uniq new_ty (getSrcSpan name)
1243 Old (but interesting) stuff about unboxed bindings
1244 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1246 What should we do when a value is specialised to a *strict* unboxed value?
1248 map_*_* f (x:xs) = let h = f x
1252 Could convert let to case:
1254 map_*_Int# f (x:xs) = case f x of h# ->
1258 This may be undesirable since it forces evaluation here, but the value
1259 may not be used in all branches of the body. In the general case this
1260 transformation is impossible since the mutual recursion in a letrec
1261 cannot be expressed as a case.
1263 There is also a problem with top-level unboxed values, since our
1264 implementation cannot handle unboxed values at the top level.
1266 Solution: Lift the binding of the unboxed value and extract it when it
1269 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1274 Now give it to the simplifier and the _Lifting will be optimised away.
1276 The benfit is that we have given the specialised "unboxed" values a
1277 very simplep lifted semantics and then leave it up to the simplifier to
1278 optimise it --- knowing that the overheads will be removed in nearly
1281 In particular, the value will only be evaluted in the branches of the
1282 program which use it, rather than being forced at the point where the
1283 value is bound. For example:
1285 filtermap_*_* p f (x:xs)
1292 filtermap_*_Int# p f (x:xs)
1293 = let h = case (f x) of h# -> _Lift h#
1296 True -> case h of _Lift h#
1300 The binding for h can still be inlined in the one branch and the
1301 _Lifting eliminated.
1304 Question: When won't the _Lifting be eliminated?
1306 Answer: When they at the top-level (where it is necessary) or when
1307 inlining would duplicate work (or possibly code depending on
1308 options). However, the _Lifting will still be eliminated if the
1309 strictness analyser deems the lifted binding strict.