2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1996
4 \section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
7 module Specialise ( specProgram ) where
9 #include "HsVersions.h"
11 import MkId ( mkUserLocal )
12 import Id ( Id, DictVar, idType, mkTemplateLocals,
14 getIdSpecialisation, setIdSpecialisation, isSpecPragmaId,
16 IdSet, mkIdSet, addOneToIdSet, intersectIdSets, isEmptyIdSet,
17 emptyIdSet, unionIdSets, minusIdSet, unitIdSet, elementOfIdSet,
19 IdEnv, mkIdEnv, lookupIdEnv, addOneToIdEnv, delOneFromIdEnv
22 import Type ( Type, mkTyVarTy, splitSigmaTy, instantiateTy, isDictTy,
23 tyVarsOfType, tyVarsOfTypes, applyTys, mkForAllTys
25 import TyCon ( TyCon )
26 import TyVar ( TyVar, mkTyVar, mkSysTyVar,
27 TyVarSet, mkTyVarSet, isEmptyTyVarSet, intersectTyVarSets,
28 elementOfTyVarSet, unionTyVarSets, emptyTyVarSet,
30 TyVarEnv, mkTyVarEnv, delFromTyVarEnv
32 import Kind ( mkBoxedTypeKind )
34 import FreeVars ( exprFreeVars )
35 import PprCore () -- Instances
36 import Name ( NamedThing(..), getSrcLoc, mkSysLocalName, isLocallyDefined )
37 import SrcLoc ( noSrcLoc )
38 import SpecEnv ( addToSpecEnv, lookupSpecEnv, specEnvValues )
40 import UniqSupply ( UniqSupply,
41 UniqSM, initUs, thenUs, returnUs, getUnique, mapUs
43 import Unique ( mkAlphaTyVarUnique )
45 import Maybes ( MaybeErr(..), maybeToBool, catMaybes )
47 import List ( partition )
48 import Util ( zipEqual )
55 %************************************************************************
57 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
59 %************************************************************************
61 These notes describe how we implement specialisation to eliminate
64 The specialisation pass works on Core
65 syntax, complete with all the explicit dictionary application,
66 abstraction and construction as added by the type checker. The
67 existing type checker remains largely as it is.
69 One important thought: the {\em types} passed to an overloaded
70 function, and the {\em dictionaries} passed are mutually redundant.
71 If the same function is applied to the same type(s) then it is sure to
72 be applied to the same dictionary(s)---or rather to the same {\em
73 values}. (The arguments might look different but they will evaluate
76 Second important thought: we know that we can make progress by
77 treating dictionary arguments as static and worth specialising on. So
78 we can do without binding-time analysis, and instead specialise on
79 dictionary arguments and no others.
88 and suppose f is overloaded.
90 STEP 1: CALL-INSTANCE COLLECTION
92 We traverse <body>, accumulating all applications of f to types and
95 (Might there be partial applications, to just some of its types and
96 dictionaries? In principle yes, but in practice the type checker only
97 builds applications of f to all its types and dictionaries, so partial
98 applications could only arise as a result of transformation, and even
99 then I think it's unlikely. In any case, we simply don't accumulate such
100 partial applications.)
102 There's a choice of whether to collect details of all *polymorphic* functions
103 or simply all *overloaded* ones. How to sort this out?
104 Pass in a predicate on the function to say if it is "interesting"?
105 This is dependent on the user flags: SpecialiseOverloaded
111 So now we have a collection of calls to f:
115 Notice that f may take several type arguments. To avoid ambiguity, we
116 say that f is called at type t1/t2 and t3/t4.
118 We take equivalence classes using equality of the *types* (ignoring
119 the dictionary args, which as mentioned previously are redundant).
121 STEP 3: SPECIALISATION
123 For each equivalence class, choose a representative (f t1 t2 d1 d2),
124 and create a local instance of f, defined thus:
126 f@t1/t2 = <f_rhs> t1 t2 d1 d2
128 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
129 of simplification will now result. However we don't actually *do* that
130 simplification. Rather, we leave it for the simplifier to do. If we
131 *did* do it, though, we'd get more call instances from the specialised
132 RHS. We can work out what they are by instantiating the call-instance
133 set from f's RHS with the types t1, t2.
135 Add this new id to f's IdInfo, to record that f has a specialised version.
137 Before doing any of this, check that f's IdInfo doesn't already
138 tell us about an existing instance of f at the required type/s.
139 (This might happen if specialisation was applied more than once, or
140 it might arise from user SPECIALIZE pragmas.)
144 Wait a minute! What if f is recursive? Then we can't just plug in
145 its right-hand side, can we?
147 But it's ok. The type checker *always* creates non-recursive definitions
148 for overloaded recursive functions. For example:
150 f x = f (x+x) -- Yes I know its silly
154 f a (d::Num a) = let p = +.sel a d
156 letrec fl (y::a) = fl (p y y)
160 We still have recusion for non-overloaded functions which we
161 speciailise, but the recursive call should get specialised to the
162 same recursive version.
168 All this is crystal clear when the function is applied to *constant
169 types*; that is, types which have no type variables inside. But what if
170 it is applied to non-constant types? Suppose we find a call of f at type
171 t1/t2. There are two possibilities:
173 (a) The free type variables of t1, t2 are in scope at the definition point
174 of f. In this case there's no problem, we proceed just as before. A common
175 example is as follows. Here's the Haskell:
180 After typechecking we have
182 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
183 in +.sel a d (f a d y) (f a d y)
185 Notice that the call to f is at type type "a"; a non-constant type.
186 Both calls to f are at the same type, so we can specialise to give:
188 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
189 in +.sel a d (f@a y) (f@a y)
192 (b) The other case is when the type variables in the instance types
193 are *not* in scope at the definition point of f. The example we are
194 working with above is a good case. There are two instances of (+.sel a d),
195 but "a" is not in scope at the definition of +.sel. Can we do anything?
196 Yes, we can "common them up", a sort of limited common sub-expression deal.
199 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
200 f@a (x::a) = +.sel@a x x
201 in +.sel@a (f@a y) (f@a y)
203 This can save work, and can't be spotted by the type checker, because
204 the two instances of +.sel weren't originally at the same type.
208 * There are quite a few variations here. For example, the defn of
209 +.sel could be floated ouside the \y, to attempt to gain laziness.
210 It certainly mustn't be floated outside the \d because the d has to
213 * We don't want to inline f_rhs in this case, because
214 that will duplicate code. Just commoning up the call is the point.
216 * Nothing gets added to +.sel's IdInfo.
218 * Don't bother unless the equivalence class has more than one item!
220 Not clear whether this is all worth it. It is of course OK to
221 simply discard call-instances when passing a big lambda.
223 Polymorphism 2 -- Overloading
225 Consider a function whose most general type is
227 f :: forall a b. Ord a => [a] -> b -> b
229 There is really no point in making a version of g at Int/Int and another
230 at Int/Bool, because it's only instancing the type variable "a" which
231 buys us any efficiency. Since g is completely polymorphic in b there
232 ain't much point in making separate versions of g for the different
235 That suggests that we should identify which of g's type variables
236 are constrained (like "a") and which are unconstrained (like "b").
237 Then when taking equivalence classes in STEP 2, we ignore the type args
238 corresponding to unconstrained type variable. In STEP 3 we make
239 polymorphic versions. Thus:
241 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
250 f a (d::Num a) = let g = ...
252 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
254 Here, g is only called at one type, but the dictionary isn't in scope at the
255 definition point for g. Usually the type checker would build a
256 definition for d1 which enclosed g, but the transformation system
257 might have moved d1's defn inward. Solution: float dictionary bindings
258 outwards along with call instances.
262 f x = let g p q = p==q
268 Before specialisation, leaving out type abstractions we have
270 f df x = let g :: Eq a => a -> a -> Bool
272 h :: Num a => a -> a -> (a, Bool)
273 h dh r s = let deq = eqFromNum dh
274 in (+ dh r s, g deq r s)
278 After specialising h we get a specialised version of h, like this:
280 h' r s = let deq = eqFromNum df
281 in (+ df r s, g deq r s)
283 But we can't naively make an instance for g from this, because deq is not in scope
284 at the defn of g. Instead, we have to float out the (new) defn of deq
285 to widen its scope. Notice that this floating can't be done in advance -- it only
286 shows up when specialisation is done.
288 User SPECIALIZE pragmas
289 ~~~~~~~~~~~~~~~~~~~~~~~
290 Specialisation pragmas can be digested by the type checker, and implemented
291 by adding extra definitions along with that of f, in the same way as before
293 f@t1/t2 = <f_rhs> t1 t2 d1 d2
295 Indeed the pragmas *have* to be dealt with by the type checker, because
296 only it knows how to build the dictionaries d1 and d2! For example
298 g :: Ord a => [a] -> [a]
299 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
301 Here, the specialised version of g is an application of g's rhs to the
302 Ord dictionary for (Tree Int), which only the type checker can conjure
303 up. There might not even *be* one, if (Tree Int) is not an instance of
304 Ord! (All the other specialision has suitable dictionaries to hand
307 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
308 it is buried in a complex (as-yet-un-desugared) binding group.
311 f@t1/t2 = f* t1 t2 d1 d2
313 where f* is the Id f with an IdInfo which says "inline me regardless!".
314 Indeed all the specialisation could be done in this way.
315 That in turn means that the simplifier has to be prepared to inline absolutely
316 any in-scope let-bound thing.
319 Again, the pragma should permit polymorphism in unconstrained variables:
321 h :: Ord a => [a] -> b -> b
322 {-# SPECIALIZE h :: [Int] -> b -> b #-}
324 We *insist* that all overloaded type variables are specialised to ground types,
325 (and hence there can be no context inside a SPECIALIZE pragma).
326 We *permit* unconstrained type variables to be specialised to
328 - or left as a polymorphic type variable
329 but nothing in between. So
331 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
333 is *illegal*. (It can be handled, but it adds complication, and gains the
337 SPECIALISING INSTANCE DECLARATIONS
338 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
341 instance Foo a => Foo [a] where
343 {-# SPECIALIZE instance Foo [Int] #-}
345 The original instance decl creates a dictionary-function
348 dfun.Foo.List :: forall a. Foo a -> Foo [a]
350 The SPECIALIZE pragma just makes a specialised copy, just as for
351 ordinary function definitions:
353 dfun.Foo.List@Int :: Foo [Int]
354 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
356 The information about what instance of the dfun exist gets added to
357 the dfun's IdInfo in the same way as a user-defined function too.
360 Automatic instance decl specialisation?
361 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
362 Can instance decls be specialised automatically? It's tricky.
363 We could collect call-instance information for each dfun, but
364 then when we specialised their bodies we'd get new call-instances
365 for ordinary functions; and when we specialised their bodies, we might get
366 new call-instances of the dfuns, and so on. This all arises because of
367 the unrestricted mutual recursion between instance decls and value decls.
369 Still, there's no actual problem; it just means that we may not do all
370 the specialisation we could theoretically do.
372 Furthermore, instance decls are usually exported and used non-locally,
373 so we'll want to compile enough to get those specialisations done.
375 Lastly, there's no such thing as a local instance decl, so we can
376 survive solely by spitting out *usage* information, and then reading that
377 back in as a pragma when next compiling the file. So for now,
378 we only specialise instance decls in response to pragmas.
381 SPITTING OUT USAGE INFORMATION
382 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
384 To spit out usage information we need to traverse the code collecting
385 call-instance information for all imported (non-prelude?) functions
386 and data types. Then we equivalence-class it and spit it out.
388 This is done at the top-level when all the call instances which escape
389 must be for imported functions and data types.
391 *** Not currently done ***
394 Partial specialisation by pragmas
395 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
396 What about partial specialisation:
398 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
399 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
403 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
405 Seems quite reasonable. Similar things could be done with instance decls:
407 instance (Foo a, Foo b) => Foo (a,b) where
409 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
410 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
412 Ho hum. Things are complex enough without this. I pass.
415 Requirements for the simplifer
416 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
417 The simplifier has to be able to take advantage of the specialisation.
419 * When the simplifier finds an application of a polymorphic f, it looks in
420 f's IdInfo in case there is a suitable instance to call instead. This converts
422 f t1 t2 d1 d2 ===> f_t1_t2
424 Note that the dictionaries get eaten up too!
426 * Dictionary selection operations on constant dictionaries must be
429 +.sel Int d ===> +Int
431 The obvious way to do this is in the same way as other specialised
432 calls: +.sel has inside it some IdInfo which tells that if it's applied
433 to the type Int then it should eat a dictionary and transform to +Int.
435 In short, dictionary selectors need IdInfo inside them for constant
438 * Exactly the same applies if a superclass dictionary is being
441 Eq.sel Int d ===> dEqInt
443 * Something similar applies to dictionary construction too. Suppose
444 dfun.Eq.List is the function taking a dictionary for (Eq a) to
445 one for (Eq [a]). Then we want
447 dfun.Eq.List Int d ===> dEq.List_Int
449 Where does the Eq [Int] dictionary come from? It is built in
450 response to a SPECIALIZE pragma on the Eq [a] instance decl.
452 In short, dfun Ids need IdInfo with a specialisation for each
453 constant instance of their instance declaration.
455 All this uses a single mechanism: the SpecEnv inside an Id
458 What does the specialisation IdInfo look like?
459 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
461 The SpecEnv of an Id maps a list of types (the template) to an expression
465 For example, if f has this SpecInfo:
467 [Int, a] -> \d:Ord Int. f' a
469 it means that we can replace the call
471 f Int t ===> (\d. f' t)
473 This chucks one dictionary away and proceeds with the
474 specialised version of f, namely f'.
477 What can't be done this way?
478 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
479 There is no way, post-typechecker, to get a dictionary for (say)
480 Eq a from a dictionary for Eq [a]. So if we find
484 we can't transform to
489 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
491 Of course, we currently have no way to automatically derive
492 eqList, nor to connect it to the Eq [a] instance decl, but you
493 can imagine that it might somehow be possible. Taking advantage
494 of this is permanently ruled out.
496 Still, this is no great hardship, because we intend to eliminate
497 overloading altogether anyway!
501 A note about non-tyvar dictionaries
502 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
503 Some Ids have types like
505 forall a,b,c. Eq a -> Ord [a] -> tau
507 This seems curious at first, because we usually only have dictionary
508 args whose types are of the form (C a) where a is a type variable.
509 But this doesn't hold for the functions arising from instance decls,
510 which sometimes get arguements with types of form (C (T a)) for some
513 Should we specialise wrt this compound-type dictionary? We used to say
515 "This is a heuristic judgement, as indeed is the fact that we
516 specialise wrt only dictionaries. We choose *not* to specialise
517 wrt compound dictionaries because at the moment the only place
518 they show up is in instance decls, where they are simply plugged
519 into a returned dictionary. So nothing is gained by specialising
522 But it is simpler and more uniform to specialise wrt these dicts too;
523 and in future GHC is likely to support full fledged type signatures
525 f ;: Eq [(a,b)] => ...
528 %************************************************************************
530 \subsubsection{The new specialiser}
532 %************************************************************************
534 Our basic game plan is this. For let(rec) bound function
535 f :: (C a, D c) => (a,b,c,d) -> Bool
537 * Find any specialised calls of f, (f ts ds), where
538 ts are the type arguments t1 .. t4, and
539 ds are the dictionary arguments d1 .. d2.
541 * Add a new definition for f1 (say):
543 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
545 Note that we abstract over the unconstrained type arguments.
549 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
551 to the specialisations of f. This will be used by the
552 simplifier to replace calls
553 (f t1 t2 t3 t4) da db
555 (\d1 d1 -> f1 t2 t4) da db
557 All the stuff about how many dictionaries to discard, and what types
558 to apply the specialised function to, are handled by the fact that the
559 SpecEnv contains a template for the result of the specialisation.
561 We don't build *partial* specialisations for f. For example:
563 f :: Eq a => a -> a -> Bool
564 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
566 Here, little is gained by making a specialised copy of f.
567 There's a distinct danger that the specialised version would
568 first build a dictionary for (Eq b, Eq c), and then select the (==)
569 method from it! Even if it didn't, not a great deal is saved.
571 We do, however, generate polymorphic, but not overloaded, specialisations:
573 f :: Eq a => [a] -> b -> b -> b
574 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
576 Hence, the invariant is this:
578 *** no specialised version is overloaded ***
581 %************************************************************************
583 \subsubsection{The exported function}
585 %************************************************************************
588 specProgram :: UniqSupply -> [CoreBinding] -> [CoreBinding]
590 = initSM us (go binds `thenSM` \ (binds', uds') ->
591 returnSM (dumpAllDictBinds uds' binds')
594 go [] = returnSM ([], emptyUDs)
595 go (bind:binds) = go binds `thenSM` \ (binds', uds) ->
596 specBind bind uds `thenSM` \ (bind', uds') ->
597 returnSM (bind' ++ binds', uds')
600 %************************************************************************
602 \subsubsection{@specExpr@: the main function}
604 %************************************************************************
607 specExpr :: CoreExpr -> SpecM (CoreExpr, UsageDetails)
609 ---------------- First the easy cases --------------------
610 specExpr e@(Var _) = returnSM (e, emptyUDs)
611 specExpr e@(Lit _) = returnSM (e, emptyUDs)
612 specExpr e@(Con _ _) = returnSM (e, emptyUDs)
613 specExpr e@(Prim _ _) = returnSM (e, emptyUDs)
615 specExpr (Note note body)
616 = specExpr body `thenSM` \ (body', uds) ->
617 returnSM (Note note body', uds)
620 ---------------- Applications might generate a call instance --------------------
621 specExpr e@(App fun arg)
624 go (App fun arg) args = go fun (arg:args)
625 go (Var f) args = returnSM (e, mkCallUDs f args)
626 go other args = specExpr other `thenSM` \ (e', uds) ->
627 returnSM (foldl App e' args, uds)
629 ---------------- Lambda/case require dumping of usage details --------------------
631 = specExpr body `thenSM` \ (body', uds) ->
633 (filtered_uds, body'') = dumpUDs bndrs uds body'
635 returnSM (foldr Lam body'' bndrs, filtered_uds)
637 (bndrs, body) = go [] e
639 -- More efficient to collect a group of binders together all at once
640 go bndrs (Lam bndr e) = go (bndr:bndrs) e
641 go bndrs e = (reverse bndrs, e)
644 specExpr (Case scrut alts)
645 = specExpr scrut `thenSM` \ (scrut', uds_scrut) ->
646 spec_alts alts `thenSM` \ (alts', uds_alts) ->
647 returnSM (Case scrut' alts', uds_scrut `plusUDs` uds_alts)
649 spec_alts (AlgAlts alts deflt)
650 = mapAndCombineSM spec_alg_alt alts `thenSM` \ (alts', uds1) ->
651 spec_deflt deflt `thenSM` \ (deflt', uds2) ->
652 returnSM (AlgAlts alts' deflt', uds1 `plusUDs` uds2)
654 spec_alts (PrimAlts alts deflt)
655 = mapAndCombineSM spec_prim_alt alts `thenSM` \ (alts', uds1) ->
656 spec_deflt deflt `thenSM` \ (deflt', uds2) ->
657 returnSM (PrimAlts alts' deflt', uds1 `plusUDs` uds2)
659 spec_alg_alt (con, args, rhs)
660 = specExpr rhs `thenSM` \ (rhs', uds) ->
662 (uds', rhs'') = dumpUDs (map ValBinder args) uds rhs'
664 returnSM ((con, args, rhs''), uds')
666 spec_prim_alt (lit, rhs)
667 = specExpr rhs `thenSM` \ (rhs', uds) ->
668 returnSM ((lit, rhs'), uds)
670 spec_deflt NoDefault = returnSM (NoDefault, emptyUDs)
671 spec_deflt (BindDefault arg rhs)
672 = specExpr rhs `thenSM` \ (rhs', uds) ->
674 (uds', rhs'') = dumpUDs [ValBinder arg] uds rhs'
676 returnSM (BindDefault arg rhs'', uds')
678 ---------------- Finally, let is the interesting case --------------------
679 specExpr (Let bind body)
680 = -- Deal with the body
681 specExpr body `thenSM` \ (body', body_uds) ->
683 -- Deal with the bindings
684 specBind bind body_uds `thenSM` \ (binds', uds) ->
687 returnSM (foldr Let body' binds', uds)
690 %************************************************************************
692 \subsubsection{Dealing with a binding}
694 %************************************************************************
697 specBind :: CoreBinding
698 -> UsageDetails -- Info on how the scope of the binding
699 -> SpecM ([CoreBinding], -- New bindings
700 UsageDetails) -- And info to pass upstream
702 specBind (NonRec bndr rhs) body_uds
703 | isDictTy (idType bndr)
704 = -- It's a dictionary binding
705 -- Pick it up and float it outwards.
706 specExpr rhs `thenSM` \ (rhs', rhs_uds) ->
708 all_uds = rhs_uds `plusUDs` addDictBind body_uds bndr rhs'
710 returnSM ([], all_uds)
712 | isSpecPragmaId bndr
713 = specExpr rhs `thenSM` \ (rhs', rhs_uds) ->
714 returnSM ([], rhs_uds `plusUDs` body_uds)
717 = -- Deal with the RHS, specialising it according
718 -- to the calls found in the body
719 specDefn (calls body_uds) (bndr,rhs) `thenSM` \ ((bndr',rhs'), spec_defns, spec_uds) ->
721 (all_uds, (dict_binds, dump_calls))
722 = splitUDs [ValBinder bndr]
723 (body_uds `plusUDs` spec_uds)
724 -- It's important that the `plusUDs` is this way round,
725 -- because body_uds may bind dictionaries that are
726 -- used in the calls passed to specDefn. So the
727 -- dictionary bindings in spec_uds may mention
728 -- dictionaries bound in body_uds.
730 -- If we make specialisations then we Rec the whole lot together
731 -- If not, leave it as a NonRec
732 new_bind | null spec_defns = NonRec bndr' rhs'
733 | otherwise = Rec ((bndr',rhs'):spec_defns)
735 returnSM ( new_bind : mkDictBinds dict_binds, all_uds )
737 specBind (Rec pairs) body_uds
738 = mapSM (specDefn (calls body_uds)) pairs `thenSM` \ stuff ->
740 (pairs', spec_defns_s, spec_uds_s) = unzip3 stuff
741 spec_defns = concat spec_defns_s
742 spec_uds = plusUDList spec_uds_s
744 (all_uds, (dict_binds, dump_calls))
745 = splitUDs (map (ValBinder . fst) pairs)
746 (body_uds `plusUDs` spec_uds)
747 -- See notes for non-rec case
749 new_bind = Rec (spec_defns ++ pairs')
751 returnSM ( new_bind : mkDictBinds dict_binds, all_uds )
753 specDefn :: CallDetails -- Info on how it is used in its scope
754 -> (Id, CoreExpr) -- The thing being bound and its un-processed RHS
755 -> SpecM ((Id, CoreExpr), -- The thing and its processed RHS
756 -- the Id may now have specialisations attached
757 [(Id,CoreExpr)], -- Extra, specialised bindings
758 UsageDetails -- Stuff to fling upwards from the RHS and its
759 ) -- specialised versions
761 specDefn calls (fn, rhs)
762 -- The first case is the interesting one
763 | n_tyvars == length rhs_tyvars -- Rhs of fn's defn has right number of big lambdas
764 && n_dicts <= length rhs_bndrs -- and enough dict args
765 && not (null calls_for_me) -- And there are some calls to specialise
766 = -- Specialise the body of the function
767 specExpr body `thenSM` \ (body', body_uds) ->
769 (float_uds, bound_uds@(dict_binds,_)) = splitUDs rhs_bndrs body_uds
772 -- Make a specialised version for each call in calls_for_me
773 mapSM (spec_call bound_uds) calls_for_me `thenSM` \ stuff ->
775 (spec_defns, spec_uds, spec_env_stuff) = unzip3 stuff
777 fn' = addIdSpecialisations fn spec_env_stuff
778 rhs' = foldr Lam (mkDictLets dict_binds body') rhs_bndrs
780 returnSM ((fn',rhs'),
782 float_uds `plusUDs` plusUDList spec_uds)
784 | otherwise -- No calls or RHS doesn't fit our preconceptions
785 = specExpr rhs `thenSM` \ (rhs', rhs_uds) ->
786 returnSM ((fn, rhs'), [], rhs_uds)
790 (tyvars, theta, tau) = splitSigmaTy fn_type
791 n_tyvars = length tyvars
792 n_dicts = length theta
794 (rhs_tyvars, rhs_ids, rhs_body) = collectBinders rhs
795 rhs_dicts = take n_dicts rhs_ids
796 rhs_bndrs = map TyBinder rhs_tyvars ++ map ValBinder rhs_dicts
797 body = mkValLam (drop n_dicts rhs_ids) rhs_body
798 -- Glue back on the non-dict lambdas
800 calls_for_me = case lookupFM calls fn of
802 Just cs -> fmToList cs
804 ----------------------------------------------------------
805 -- Specialise to one particular call pattern
806 spec_call :: ProtoUsageDetails -- From the original body, captured by
807 -- the dictionary lambdas
808 -> ([Maybe Type], [DictVar]) -- Call instance
809 -> SpecM ((Id,CoreExpr), -- Specialised definition
810 UsageDetails, -- Usage details from specialised body
811 ([TyVar], [Type], CoreExpr)) -- Info for the Id's SpecEnv
812 spec_call bound_uds (call_ts, call_ds)
813 = ASSERT( length call_ts == n_tyvars && length call_ds == n_dicts )
814 -- Calls are only recorded for properly-saturated applications
816 -- Supppose the call is for f [Just t1, Nothing, Just t3, Nothing] [d1, d2]
818 -- Construct the new binding
819 -- f1 = /\ b d -> (..rhs of f..) t1 b t3 d d1 d2
820 -- and the type of this binder
822 mk_spec_ty Nothing = newTyVarSM `thenSM` \ tyvar ->
823 returnSM (Just tyvar, mkTyVarTy tyvar)
824 mk_spec_ty (Just ty) = returnSM (Nothing, ty)
826 mapSM mk_spec_ty call_ts `thenSM` \ stuff ->
828 (maybe_spec_tyvars, spec_tys) = unzip stuff
829 spec_tyvars = catMaybes maybe_spec_tyvars
830 spec_rhs = mkTyLam spec_tyvars $
831 mkGenApp rhs (map TyArg spec_tys ++ map VarArg call_ds)
832 spec_id_ty = mkForAllTys spec_tyvars (instantiateTy ty_env tau)
833 ty_env = mkTyVarEnv (zipEqual "spec_call" tyvars spec_tys)
836 newIdSM fn spec_id_ty `thenSM` \ spec_f ->
839 -- Construct the stuff for f's spec env
840 -- [b,d] [t1,b,t3,d] |-> \d1 d2 -> f1 b d
841 -- The only awkward bit is that d1,d2 might well be global
842 -- dictionaries, so it's tidier to make new local variables
843 -- for the lambdas in the RHS, rather than lambda-bind the
844 -- dictionaries themselves.
846 -- In fact we use the standard template locals, so that the
847 -- they don't need to be "tidied" before putting in interface files
849 arg_ds = mkTemplateLocals (map idType call_ds)
850 spec_env_rhs = mkValLam arg_ds $
851 mkTyApp (Var spec_f) $
852 map mkTyVarTy spec_tyvars
853 spec_env_info = (spec_tyvars, spec_tys, spec_env_rhs)
856 -- Specialise the UDs from f's RHS
858 -- Only the overloaded tyvars should be free in the uds
859 ty_env = [ (rhs_tyvar,ty)
860 | (rhs_tyvar, Just ty) <- zipEqual "specUDs1" rhs_tyvars call_ts
862 dict_env = zipEqual "specUDs2" rhs_dicts call_ds
864 specUDs ty_env dict_env bound_uds `thenSM` \ spec_uds ->
866 returnSM ((spec_f, spec_rhs),
872 %************************************************************************
874 \subsubsection{UsageDetails and suchlike}
876 %************************************************************************
879 type FreeDicts = IdSet
883 dict_binds :: !(Bag DictBind),
884 -- Floated dictionary bindings
885 -- The order is important;
886 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
887 -- (Remember, Bags preserve order in GHC.)
888 -- The FreeDicts is the free vars of the RHS
890 calls :: !CallDetails
893 type DictBind = (DictVar, CoreExpr, TyVarSet, FreeDicts)
894 -- The FreeDicts are the free dictionaries (only)
895 -- of the RHS of the dictionary bindings
896 -- Similarly the TyVarSet
898 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM }
900 type ProtoUsageDetails = ([DictBind],
901 [(Id, [Maybe Type], [DictVar])]
904 ------------------------------------------------------------
905 type CallDetails = FiniteMap Id CallInfo
906 type CallInfo = FiniteMap [Maybe Type] -- Nothing => unconstrained type argument
907 [DictVar] -- Dict args
908 -- The finite maps eliminate duplicates
909 -- The list of types and dictionaries is guaranteed to
910 -- match the type of f
912 callDetailsToList calls = [ (id,tys,dicts)
913 | (id,fm) <- fmToList calls,
914 (tys,dicts) <- fmToList fm
917 listToCallDetails calls = foldr (unionCalls . singleCall) emptyFM calls
919 unionCalls :: CallDetails -> CallDetails -> CallDetails
920 unionCalls c1 c2 = plusFM_C plusFM c1 c2
922 singleCall (id, tys, dicts) = unitFM id (unitFM tys dicts)
926 || length spec_tys /= n_tyvars
927 || length dicts /= n_dicts
928 = emptyUDs -- Not overloaded
931 = MkUD {dict_binds = emptyBag,
932 calls = singleCall (f, spec_tys, dicts)
935 (tyvars, theta, tau) = splitSigmaTy (idType f)
936 constrained_tyvars = foldr (unionTyVarSets . tyVarsOfTypes . snd) emptyTyVarSet theta
937 n_tyvars = length tyvars
938 n_dicts = length theta
940 spec_tys = [mk_spec_ty tv ty | (tv, TyArg ty) <- tyvars `zip` args]
941 dicts = [d | (_, VarArg d) <- theta `zip` (drop n_tyvars args)]
943 mk_spec_ty tyvar ty | tyvar `elementOfTyVarSet` constrained_tyvars
948 ------------------------------------------------------------
949 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
950 plusUDs (MkUD {dict_binds = db1, calls = calls1})
951 (MkUD {dict_binds = db2, calls = calls2})
952 = MkUD {dict_binds, calls}
954 dict_binds = db1 `unionBags` db2
955 calls = calls1 `unionCalls` calls2
957 plusUDList = foldr plusUDs emptyUDs
959 mkDB dict rhs = (dict, rhs, db_ftvs, db_fvs)
961 db_ftvs = tyVarsOfType (idType dict) -- Superset of RHS fvs
962 db_fvs = dictRhsFVs rhs
964 addDictBind uds dict rhs = uds { dict_binds = mkDB dict rhs `consBag` dict_binds uds }
966 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
967 = foldrBag add binds dbs
969 add (dict,rhs,_,_) binds = NonRec dict rhs : binds
971 mkDictBinds :: [DictBind] -> [CoreBinding]
972 mkDictBinds = map (\(d,r,_,_) -> NonRec d r)
974 mkDictLets :: [DictBind] -> CoreExpr -> CoreExpr
975 mkDictLets dbs body = foldr mk body dbs
977 mk (d,r,_,_) e = Let (NonRec d r) e
979 dumpUDs :: [CoreBinder]
980 -> UsageDetails -> CoreExpr
981 -> (UsageDetails, CoreExpr)
982 dumpUDs bndrs uds body
983 = (free_uds, mkDictLets dict_binds body)
985 (free_uds, (dict_binds, _)) = splitUDs bndrs uds
987 splitUDs :: [CoreBinder]
989 -> (UsageDetails, -- These don't mention the binders
990 ProtoUsageDetails) -- These do
992 splitUDs bndrs uds@(MkUD {dict_binds = orig_dbs,
995 = if isEmptyBag dump_dbs && null dump_calls then
996 -- Common case: binder doesn't affect floats
1000 -- Binders bind some of the fvs of the floats
1001 (MkUD {dict_binds = free_dbs,
1002 calls = listToCallDetails free_calls},
1003 (bagToList dump_dbs, dump_calls)
1007 tyvar_set = mkTyVarSet [tv | TyBinder tv <- bndrs]
1008 id_set = mkIdSet [id | ValBinder id <- bndrs]
1010 (free_dbs, dump_dbs, dump_idset)
1011 = foldlBag dump_db (emptyBag, emptyBag, id_set) orig_dbs
1012 -- Important that it's foldl not foldr;
1013 -- we're accumulating the set of dumped ids in dump_set
1015 -- Filter out any calls that mention things that are being dumped
1016 -- Don't need to worry about the tyvars because the dicts will
1017 -- spot the captured ones; any fully polymorphic arguments will
1018 -- be Nothings in the call details
1019 orig_call_list = callDetailsToList orig_calls
1020 (dump_calls, free_calls) = partition captured orig_call_list
1021 captured (id,tys,dicts) = any (`elementOfIdSet` dump_idset) (id:dicts)
1023 dump_db (free_dbs, dump_dbs, dump_idset) db@(dict, rhs, ftvs, fvs)
1024 | isEmptyIdSet (dump_idset `intersectIdSets` fvs)
1025 && isEmptyTyVarSet (tyvar_set `intersectTyVarSets` ftvs)
1026 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1028 | otherwise -- Dump it
1029 = (free_dbs, dump_dbs `snocBag` db,
1030 dump_idset `addOneToIdSet` dict)
1033 Given a type and value substitution, specUDs creates a specialised copy of
1037 specUDs :: [(TyVar,Type)] -> [(DictVar,DictVar)] -> ProtoUsageDetails -> SpecM UsageDetails
1038 specUDs tv_env_list dict_env_list (dbs, calls)
1039 = specDBs dict_env_list dbs `thenSM` \ (dict_env_list', dbs') ->
1041 dict_env = mkIdEnv dict_env_list'
1043 returnSM (MkUD { dict_binds = dbs',
1044 calls = listToCallDetails (map (inst_call dict_env) calls)
1047 bound_tyvars = mkTyVarSet (map fst tv_env_list)
1048 tv_env = mkTyVarEnv tv_env_list -- Doesn't change
1050 inst_call dict_env (id, tys, dicts) = (id, map inst_maybe_ty tys,
1051 map (lookupId dict_env) dicts)
1053 inst_maybe_ty Nothing = Nothing
1054 inst_maybe_ty (Just ty) = Just (instantiateTy tv_env ty)
1057 = returnSM (dict_env, emptyBag)
1058 specDBs dict_env ((dict, rhs, ftvs, fvs) : dbs)
1059 = newIdSM dict (instantiateTy tv_env (idType dict)) `thenSM` \ dict' ->
1061 rhs' = foldl App (foldr Lam rhs (t_bndrs ++ d_bndrs)) (t_args ++ d_args)
1062 (t_bndrs, t_args) = unzip [(TyBinder tv, TyArg ty) | (tv,ty) <- tv_env_list,
1063 tv `elementOfTyVarSet` ftvs]
1064 (d_bndrs, d_args) = unzip [(ValBinder d, VarArg d') | (d,d') <- dict_env,
1065 d `elementOfIdSet` fvs]
1066 dict_env' = (dict,dict') : dict_env
1067 ftvs' = tyVarsOfTypes [ty | TyArg ty <- t_args] `unionTyVarSets`
1068 (ftvs `minusTyVarSet` bound_tyvars)
1069 fvs' = mkIdSet [d | VarArg d <- d_args] `unionIdSets`
1070 (fvs `minusIdSet` mkIdSet [d | ValBinder d <- d_bndrs])
1072 specDBs dict_env' dbs `thenSM` \ (dict_env'', dbs') ->
1073 returnSM ( dict_env'', (dict', rhs', ftvs', fvs') `consBag` dbs' )
1076 %************************************************************************
1078 \subsubsection{Boring helper functions}
1080 %************************************************************************
1083 lookupId:: IdEnv Id -> Id -> Id
1084 lookupId env id = case lookupIdEnv env id of
1088 dictRhsFVs :: CoreExpr -> IdSet
1089 dictRhsFVs e = exprFreeVars isLocallyDefined e
1091 addIdSpecialisations id spec_stuff
1092 = (if not (null errs) then
1093 pprTrace "Duplicate specialisations" (vcat (map ppr errs))
1096 setIdSpecialisation id new_spec_env
1098 (new_spec_env, errs) = foldr add (getIdSpecialisation id, []) spec_stuff
1100 add (tyvars, tys, template) (spec_env, errs)
1101 = case addToSpecEnv True spec_env tyvars tys template of
1102 Succeeded spec_env' -> (spec_env', errs)
1103 Failed err -> (spec_env, err:errs)
1105 ----------------------------------------
1106 type SpecM a = UniqSM a
1110 getUniqSM = getUnique
1114 mapAndCombineSM f [] = returnSM ([], emptyUDs)
1115 mapAndCombineSM f (x:xs) = f x `thenSM` \ (y, uds1) ->
1116 mapAndCombineSM f xs `thenSM` \ (ys, uds2) ->
1117 returnSM (y:ys, uds1 `plusUDs` uds2)
1119 newIdSM old_id new_ty
1120 = getUnique `thenSM` \ uniq ->
1121 returnSM (mkUserLocal (getOccName old_id)
1128 = getUnique `thenSM` \ uniq ->
1129 returnSM (mkSysTyVar uniq mkBoxedTypeKind)
1133 Old (but interesting) stuff about unboxed bindings
1134 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1136 What should we do when a value is specialised to a *strict* unboxed value?
1138 map_*_* f (x:xs) = let h = f x
1142 Could convert let to case:
1144 map_*_Int# f (x:xs) = case f x of h# ->
1148 This may be undesirable since it forces evaluation here, but the value
1149 may not be used in all branches of the body. In the general case this
1150 transformation is impossible since the mutual recursion in a letrec
1151 cannot be expressed as a case.
1153 There is also a problem with top-level unboxed values, since our
1154 implementation cannot handle unboxed values at the top level.
1156 Solution: Lift the binding of the unboxed value and extract it when it
1159 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1164 Now give it to the simplifier and the _Lifting will be optimised away.
1166 The benfit is that we have given the specialised "unboxed" values a
1167 very simplep lifted semantics and then leave it up to the simplifier to
1168 optimise it --- knowing that the overheads will be removed in nearly
1171 In particular, the value will only be evaluted in the branches of the
1172 program which use it, rather than being forced at the point where the
1173 value is bound. For example:
1175 filtermap_*_* p f (x:xs)
1182 filtermap_*_Int# p f (x:xs)
1183 = let h = case (f x) of h# -> _Lift h#
1186 True -> case h of _Lift h#
1190 The binding for h can still be inlined in the one branch and the
1191 _Lifting eliminated.
1194 Question: When won't the _Lifting be eliminated?
1196 Answer: When they at the top-level (where it is necessary) or when
1197 inlining would duplicate work (or possibly code depending on
1198 options). However, the _Lifting will still be eliminated if the
1199 strictness analyser deems the lifted binding strict.