2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1996
4 \section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
12 #include "HsVersions.h"
14 import MkId ( mkUserLocal )
15 import Id ( Id, DictVar, idType,
17 getIdSpecialisation, setIdSpecialisation, isSpecPragmaId,
19 IdSet, mkIdSet, addOneToIdSet, intersectIdSets, isEmptyIdSet,
20 emptyIdSet, unionIdSets, minusIdSet, unitIdSet, elementOfIdSet,
22 IdEnv, mkIdEnv, lookupIdEnv, addOneToIdEnv, delOneFromIdEnv
25 import Type ( Type, mkTyVarTy, splitSigmaTy, instantiateTy, isDictTy,
26 tyVarsOfType, tyVarsOfTypes, applyTys, mkForAllTys
28 import TyCon ( TyCon )
30 TyVarSet, mkTyVarSet, isEmptyTyVarSet, intersectTyVarSets,
31 elementOfTyVarSet, unionTyVarSets, emptyTyVarSet,
32 TyVarEnv, mkTyVarEnv, delFromTyVarEnv
35 import PprCore () -- Instances
36 import Name ( NamedThing(..), getSrcLoc )
37 import SpecEnv ( addToSpecEnv, lookupSpecEnv, specEnvValues )
39 import UniqSupply ( UniqSupply,
40 UniqSM, initUs, thenUs, returnUs, getUnique, mapUs
44 import Maybes ( MaybeErr(..), maybeToBool )
46 import List ( partition )
47 import Util ( zipEqual )
54 %************************************************************************
56 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
58 %************************************************************************
60 These notes describe how we implement specialisation to eliminate
63 The specialisation pass works on Core
64 syntax, complete with all the explicit dictionary application,
65 abstraction and construction as added by the type checker. The
66 existing type checker remains largely as it is.
68 One important thought: the {\em types} passed to an overloaded
69 function, and the {\em dictionaries} passed are mutually redundant.
70 If the same function is applied to the same type(s) then it is sure to
71 be applied to the same dictionary(s)---or rather to the same {\em
72 values}. (The arguments might look different but they will evaluate
75 Second important thought: we know that we can make progress by
76 treating dictionary arguments as static and worth specialising on. So
77 we can do without binding-time analysis, and instead specialise on
78 dictionary arguments and no others.
87 and suppose f is overloaded.
89 STEP 1: CALL-INSTANCE COLLECTION
91 We traverse <body>, accumulating all applications of f to types and
94 (Might there be partial applications, to just some of its types and
95 dictionaries? In principle yes, but in practice the type checker only
96 builds applications of f to all its types and dictionaries, so partial
97 applications could only arise as a result of transformation, and even
98 then I think it's unlikely. In any case, we simply don't accumulate such
99 partial applications.)
101 There's a choice of whether to collect details of all *polymorphic* functions
102 or simply all *overloaded* ones. How to sort this out?
103 Pass in a predicate on the function to say if it is "interesting"?
104 This is dependent on the user flags: SpecialiseOverloaded
110 So now we have a collection of calls to f:
114 Notice that f may take several type arguments. To avoid ambiguity, we
115 say that f is called at type t1/t2 and t3/t4.
117 We take equivalence classes using equality of the *types* (ignoring
118 the dictionary args, which as mentioned previously are redundant).
120 STEP 3: SPECIALISATION
122 For each equivalence class, choose a representative (f t1 t2 d1 d2),
123 and create a local instance of f, defined thus:
125 f@t1/t2 = <f_rhs> t1 t2 d1 d2
127 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
128 of simplification will now result. However we don't actually *do* that
129 simplification. Rather, we leave it for the simplifier to do. If we
130 *did* do it, though, we'd get more call instances from the specialised
131 RHS. We can work out what they are by instantiating the call-instance
132 set from f's RHS with the types t1, t2.
134 Add this new id to f's IdInfo, to record that f has a specialised version.
136 Before doing any of this, check that f's IdInfo doesn't already
137 tell us about an existing instance of f at the required type/s.
138 (This might happen if specialisation was applied more than once, or
139 it might arise from user SPECIALIZE pragmas.)
143 Wait a minute! What if f is recursive? Then we can't just plug in
144 its right-hand side, can we?
146 But it's ok. The type checker *always* creates non-recursive definitions
147 for overloaded recursive functions. For example:
149 f x = f (x+x) -- Yes I know its silly
153 f a (d::Num a) = let p = +.sel a d
155 letrec fl (y::a) = fl (p y y)
159 We still have recusion for non-overloaded functions which we
160 speciailise, but the recursive call should get specialised to the
161 same recursive version.
167 All this is crystal clear when the function is applied to *constant
168 types*; that is, types which have no type variables inside. But what if
169 it is applied to non-constant types? Suppose we find a call of f at type
170 t1/t2. There are two possibilities:
172 (a) The free type variables of t1, t2 are in scope at the definition point
173 of f. In this case there's no problem, we proceed just as before. A common
174 example is as follows. Here's the Haskell:
179 After typechecking we have
181 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
182 in +.sel a d (f a d y) (f a d y)
184 Notice that the call to f is at type type "a"; a non-constant type.
185 Both calls to f are at the same type, so we can specialise to give:
187 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
188 in +.sel a d (f@a y) (f@a y)
191 (b) The other case is when the type variables in the instance types
192 are *not* in scope at the definition point of f. The example we are
193 working with above is a good case. There are two instances of (+.sel a d),
194 but "a" is not in scope at the definition of +.sel. Can we do anything?
195 Yes, we can "common them up", a sort of limited common sub-expression deal.
198 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
199 f@a (x::a) = +.sel@a x x
200 in +.sel@a (f@a y) (f@a y)
202 This can save work, and can't be spotted by the type checker, because
203 the two instances of +.sel weren't originally at the same type.
207 * There are quite a few variations here. For example, the defn of
208 +.sel could be floated ouside the \y, to attempt to gain laziness.
209 It certainly mustn't be floated outside the \d because the d has to
212 * We don't want to inline f_rhs in this case, because
213 that will duplicate code. Just commoning up the call is the point.
215 * Nothing gets added to +.sel's IdInfo.
217 * Don't bother unless the equivalence class has more than one item!
219 Not clear whether this is all worth it. It is of course OK to
220 simply discard call-instances when passing a big lambda.
222 Polymorphism 2 -- Overloading
224 Consider a function whose most general type is
226 f :: forall a b. Ord a => [a] -> b -> b
228 There is really no point in making a version of g at Int/Int and another
229 at Int/Bool, because it's only instancing the type variable "a" which
230 buys us any efficiency. Since g is completely polymorphic in b there
231 ain't much point in making separate versions of g for the different
234 That suggests that we should identify which of g's type variables
235 are constrained (like "a") and which are unconstrained (like "b").
236 Then when taking equivalence classes in STEP 2, we ignore the type args
237 corresponding to unconstrained type variable. In STEP 3 we make
238 polymorphic versions. Thus:
240 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
249 f a (d::Num a) = let g = ...
251 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
253 Here, g is only called at one type, but the dictionary isn't in scope at the
254 definition point for g. Usually the type checker would build a
255 definition for d1 which enclosed g, but the transformation system
256 might have moved d1's defn inward. Solution: float dictionary bindings
257 outwards along with call instances.
261 f x = let g p q = p==q
267 Before specialisation, leaving out type abstractions we have
269 f df x = let g :: Eq a => a -> a -> Bool
271 h :: Num a => a -> a -> (a, Bool)
272 h dh r s = let deq = eqFromNum dh
273 in (+ dh r s, g deq r s)
277 After specialising h we get a specialised version of h, like this:
279 h' r s = let deq = eqFromNum df
280 in (+ df r s, g deq r s)
282 But we can't naively make an instance for g from this, because deq is not in scope
283 at the defn of g. Instead, we have to float out the (new) defn of deq
284 to widen its scope. Notice that this floating can't be done in advance -- it only
285 shows up when specialisation is done.
287 User SPECIALIZE pragmas
288 ~~~~~~~~~~~~~~~~~~~~~~~
289 Specialisation pragmas can be digested by the type checker, and implemented
290 by adding extra definitions along with that of f, in the same way as before
292 f@t1/t2 = <f_rhs> t1 t2 d1 d2
294 Indeed the pragmas *have* to be dealt with by the type checker, because
295 only it knows how to build the dictionaries d1 and d2! For example
297 g :: Ord a => [a] -> [a]
298 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
300 Here, the specialised version of g is an application of g's rhs to the
301 Ord dictionary for (Tree Int), which only the type checker can conjure
302 up. There might not even *be* one, if (Tree Int) is not an instance of
303 Ord! (All the other specialision has suitable dictionaries to hand
306 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
307 it is buried in a complex (as-yet-un-desugared) binding group.
310 f@t1/t2 = f* t1 t2 d1 d2
312 where f* is the Id f with an IdInfo which says "inline me regardless!".
313 Indeed all the specialisation could be done in this way.
314 That in turn means that the simplifier has to be prepared to inline absolutely
315 any in-scope let-bound thing.
318 Again, the pragma should permit polymorphism in unconstrained variables:
320 h :: Ord a => [a] -> b -> b
321 {-# SPECIALIZE h :: [Int] -> b -> b #-}
323 We *insist* that all overloaded type variables are specialised to ground types,
324 (and hence there can be no context inside a SPECIALIZE pragma).
325 We *permit* unconstrained type variables to be specialised to
327 - or left as a polymorphic type variable
328 but nothing in between. So
330 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
332 is *illegal*. (It can be handled, but it adds complication, and gains the
336 SPECIALISING INSTANCE DECLARATIONS
337 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
340 instance Foo a => Foo [a] where
342 {-# SPECIALIZE instance Foo [Int] #-}
344 The original instance decl creates a dictionary-function
347 dfun.Foo.List :: forall a. Foo a -> Foo [a]
349 The SPECIALIZE pragma just makes a specialised copy, just as for
350 ordinary function definitions:
352 dfun.Foo.List@Int :: Foo [Int]
353 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
355 The information about what instance of the dfun exist gets added to
356 the dfun's IdInfo in the same way as a user-defined function too.
359 Automatic instance decl specialisation?
360 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
361 Can instance decls be specialised automatically? It's tricky.
362 We could collect call-instance information for each dfun, but
363 then when we specialised their bodies we'd get new call-instances
364 for ordinary functions; and when we specialised their bodies, we might get
365 new call-instances of the dfuns, and so on. This all arises because of
366 the unrestricted mutual recursion between instance decls and value decls.
368 Still, there's no actual problem; it just means that we may not do all
369 the specialisation we could theoretically do.
371 Furthermore, instance decls are usually exported and used non-locally,
372 so we'll want to compile enough to get those specialisations done.
374 Lastly, there's no such thing as a local instance decl, so we can
375 survive solely by spitting out *usage* information, and then reading that
376 back in as a pragma when next compiling the file. So for now,
377 we only specialise instance decls in response to pragmas.
380 SPITTING OUT USAGE INFORMATION
381 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
383 To spit out usage information we need to traverse the code collecting
384 call-instance information for all imported (non-prelude?) functions
385 and data types. Then we equivalence-class it and spit it out.
387 This is done at the top-level when all the call instances which escape
388 must be for imported functions and data types.
390 *** Not currently done ***
393 Partial specialisation by pragmas
394 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
395 What about partial specialisation:
397 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
398 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
402 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
404 Seems quite reasonable. Similar things could be done with instance decls:
406 instance (Foo a, Foo b) => Foo (a,b) where
408 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
409 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
411 Ho hum. Things are complex enough without this. I pass.
414 Requirements for the simplifer
415 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
416 The simplifier has to be able to take advantage of the specialisation.
418 * When the simplifier finds an application of a polymorphic f, it looks in
419 f's IdInfo in case there is a suitable instance to call instead. This converts
421 f t1 t2 d1 d2 ===> f_t1_t2
423 Note that the dictionaries get eaten up too!
425 * Dictionary selection operations on constant dictionaries must be
428 +.sel Int d ===> +Int
430 The obvious way to do this is in the same way as other specialised
431 calls: +.sel has inside it some IdInfo which tells that if it's applied
432 to the type Int then it should eat a dictionary and transform to +Int.
434 In short, dictionary selectors need IdInfo inside them for constant
437 * Exactly the same applies if a superclass dictionary is being
440 Eq.sel Int d ===> dEqInt
442 * Something similar applies to dictionary construction too. Suppose
443 dfun.Eq.List is the function taking a dictionary for (Eq a) to
444 one for (Eq [a]). Then we want
446 dfun.Eq.List Int d ===> dEq.List_Int
448 Where does the Eq [Int] dictionary come from? It is built in
449 response to a SPECIALIZE pragma on the Eq [a] instance decl.
451 In short, dfun Ids need IdInfo with a specialisation for each
452 constant instance of their instance declaration.
454 All this uses a single mechanism: the SpecEnv inside an Id
457 What does the specialisation IdInfo look like?
458 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
460 The SpecEnv of an Id maps a list of types (the template) to an expression
464 For example, if f has this SpecInfo:
466 [Int, a] -> \d:Ord Int. f' a
468 it means that we can replace the call
470 f Int t ===> (\d. f' t)
472 This chucks one dictionary away and proceeds with the
473 specialised version of f, namely f'.
476 What can't be done this way?
477 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
478 There is no way, post-typechecker, to get a dictionary for (say)
479 Eq a from a dictionary for Eq [a]. So if we find
483 we can't transform to
488 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
490 Of course, we currently have no way to automatically derive
491 eqList, nor to connect it to the Eq [a] instance decl, but you
492 can imagine that it might somehow be possible. Taking advantage
493 of this is permanently ruled out.
495 Still, this is no great hardship, because we intend to eliminate
496 overloading altogether anyway!
500 A note about non-tyvar dictionaries
501 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
502 Some Ids have types like
504 forall a,b,c. Eq a -> Ord [a] -> tau
506 This seems curious at first, because we usually only have dictionary
507 args whose types are of the form (C a) where a is a type variable.
508 But this doesn't hold for the functions arising from instance decls,
509 which sometimes get arguements with types of form (C (T a)) for some
512 Should we specialise wrt this compound-type dictionary? We used to say
514 "This is a heuristic judgement, as indeed is the fact that we
515 specialise wrt only dictionaries. We choose *not* to specialise
516 wrt compound dictionaries because at the moment the only place
517 they show up is in instance decls, where they are simply plugged
518 into a returned dictionary. So nothing is gained by specialising
521 But it is simpler and more uniform to specialise wrt these dicts too;
522 and in future GHC is likely to support full fledged type signatures
524 f ;: Eq [(a,b)] => ...
527 %************************************************************************
529 \subsubsection{The new specialiser}
531 %************************************************************************
533 Our basic game plan is this. For let(rec) bound function
534 f :: (C a, D c) => (a,b,c,d) -> Bool
536 * Find any specialised calls of f, (f ts ds), where
537 ts are the type arguments t1 .. t4, and
538 ds are the dictionary arguments d1 .. d2.
540 * Add a new definition for f1 (say):
542 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
544 Note that we abstract over the unconstrained type arguments.
548 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
550 to the specialisations of f. This will be used by the
551 simplifier to replace calls
552 (f t1 t2 t3 t4) da db
554 (\d1 d1 -> f1 t2 t4) da db
556 All the stuff about how many dictionaries to discard, and what types
557 to apply the specialised function to, are handled by the fact that the
558 SpecEnv contains a template for the result of the specialisation.
560 We don't build *partial* specialisations for f. For example:
562 f :: Eq a => a -> a -> Bool
563 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
565 Here, little is gained by making a specialised copy of f.
566 There's a distinct danger that the specialised version would
567 first build a dictionary for (Eq b, Eq c), and then select the (==)
568 method from it! Even if it didn't, not a great deal is saved.
570 We do, however, generate polymorphic, but not overloaded, specialisations:
572 f :: Eq a => [a] -> b -> b -> b
573 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
575 Hence, the invariant is this:
577 *** no specialised version is overloaded ***
580 %************************************************************************
582 \subsubsection{The exported function}
584 %************************************************************************
587 specProgram :: UniqSupply -> [CoreBinding] -> [CoreBinding]
589 = initSM us (go binds `thenSM` \ (binds', uds') ->
590 returnSM (dumpAllDictBinds uds' binds')
593 go [] = returnSM ([], emptyUDs)
594 go (bind:binds) = go binds `thenSM` \ (binds', uds) ->
595 specBind bind uds `thenSM` \ (bind', uds') ->
596 returnSM (bind' ++ binds', uds')
599 %************************************************************************
601 \subsubsection{@specExpr@: the main function}
603 %************************************************************************
606 specExpr :: CoreExpr -> SpecM (CoreExpr, UsageDetails)
608 ---------------- First the easy cases --------------------
609 specExpr e@(Var _) = returnSM (e, emptyUDs)
610 specExpr e@(Lit _) = returnSM (e, emptyUDs)
611 specExpr e@(Con _ _) = returnSM (e, emptyUDs)
612 specExpr e@(Prim _ _) = returnSM (e, emptyUDs)
614 specExpr (Note note body)
615 = specExpr body `thenSM` \ (body', uds) ->
616 returnSM (Note note body', uds)
619 ---------------- Applications might generate a call instance --------------------
620 specExpr e@(App fun arg)
623 go (App fun arg) args = go fun (arg:args)
624 go (Var f) args = returnSM (e, mkCallUDs f args)
625 go other args = specExpr other `thenSM` \ (e', uds) ->
626 returnSM (foldl App e' args, uds)
628 ---------------- Lambda/case require dumping of usage details --------------------
630 = specExpr body `thenSM` \ (body', uds) ->
632 (filtered_uds, body'') = dumpUDs bndrs uds body'
634 returnSM (foldr Lam body'' bndrs, filtered_uds)
636 (bndrs, body) = go [] e
638 -- More efficient to collect a group of binders together all at once
639 go bndrs (Lam bndr e) = go (bndr:bndrs) e
640 go bndrs e = (reverse bndrs, e)
643 specExpr (Case scrut alts)
644 = specExpr scrut `thenSM` \ (scrut', uds_scrut) ->
645 spec_alts alts `thenSM` \ (alts', uds_alts) ->
646 returnSM (Case scrut' alts', uds_scrut `plusUDs` uds_alts)
648 spec_alts (AlgAlts alts deflt)
649 = mapAndCombineSM spec_alg_alt alts `thenSM` \ (alts', uds1) ->
650 spec_deflt deflt `thenSM` \ (deflt', uds2) ->
651 returnSM (AlgAlts alts' deflt', uds1 `plusUDs` uds2)
653 spec_alts (PrimAlts alts deflt)
654 = mapAndCombineSM spec_prim_alt alts `thenSM` \ (alts', uds1) ->
655 spec_deflt deflt `thenSM` \ (deflt', uds2) ->
656 returnSM (PrimAlts alts' deflt', uds1 `plusUDs` uds2)
658 spec_alg_alt (con, args, rhs)
659 = specExpr rhs `thenSM` \ (rhs', uds) ->
661 (uds', rhs'') = dumpUDs (map ValBinder args) uds rhs'
663 returnSM ((con, args, rhs''), uds')
665 spec_prim_alt (lit, rhs)
666 = specExpr rhs `thenSM` \ (rhs', uds) ->
667 returnSM ((lit, rhs'), uds)
669 spec_deflt NoDefault = returnSM (NoDefault, emptyUDs)
670 spec_deflt (BindDefault arg rhs)
671 = specExpr rhs `thenSM` \ (rhs', uds) ->
673 (uds', rhs'') = dumpUDs [ValBinder arg] uds rhs'
675 returnSM (BindDefault arg rhs'', uds')
677 ---------------- Finally, let is the interesting case --------------------
678 specExpr (Let bind body)
679 = -- Deal with the body
680 specExpr body `thenSM` \ (body', body_uds) ->
682 -- Deal with the bindings
683 specBind bind body_uds `thenSM` \ (binds', uds) ->
686 returnSM (foldr Let body' binds', uds)
689 %************************************************************************
691 \subsubsection{Dealing with a binding}
693 %************************************************************************
696 specBind :: CoreBinding
697 -> UsageDetails -- Info on how the scope of the binding
698 -> SpecM ([CoreBinding], -- New bindings
699 UsageDetails) -- And info to pass upstream
701 specBind (NonRec bndr rhs) body_uds
702 | isDictTy (idType bndr)
703 = -- It's a dictionary binding
704 -- Pick it up and float it outwards.
705 specExpr rhs `thenSM` \ (rhs', rhs_uds) ->
707 all_uds = rhs_uds `plusUDs` addDictBind body_uds bndr rhs'
709 returnSM ([], all_uds)
711 | isSpecPragmaId bndr
712 = specExpr rhs `thenSM` \ (rhs', rhs_uds) ->
713 returnSM ([], rhs_uds)
716 = -- Deal with the RHS, specialising it according
717 -- to the calls found in the body
718 specDefn (calls body_uds) (bndr,rhs) `thenSM` \ ((bndr',rhs'), spec_defns, spec_uds) ->
720 (all_uds, (dict_binds, dump_calls))
721 = splitUDs [ValBinder bndr] (spec_uds `plusUDs` body_uds)
723 -- If we make specialisations then we Rec the whole lot together
724 -- If not, leave it as a NonRec
725 new_bind | null spec_defns = NonRec bndr' rhs'
726 | otherwise = Rec ((bndr',rhs'):spec_defns)
728 returnSM ( new_bind : dict_binds, all_uds )
730 specBind (Rec pairs) body_uds
731 = mapSM (specDefn (calls body_uds)) pairs `thenSM` \ stuff ->
733 (pairs', spec_defns_s, spec_uds_s) = unzip3 stuff
734 spec_defns = concat spec_defns_s
735 spec_uds = plusUDList spec_uds_s
736 (all_uds, (dict_binds, dump_calls))
737 = splitUDs (map (ValBinder . fst) pairs) (spec_uds `plusUDs` body_uds)
738 new_bind = Rec (spec_defns ++ pairs')
740 returnSM ( new_bind : dict_binds, all_uds )
742 specDefn :: CallDetails -- Info on how it is used in its scope
743 -> (Id, CoreExpr) -- The thing being bound and its un-processed RHS
744 -> SpecM ((Id, CoreExpr), -- The thing and its processed RHS
745 -- the Id may now have specialisations attached
746 [(Id,CoreExpr)], -- Extra, specialised bindings
747 UsageDetails -- Stuff to fling upwards from the RHS and its
748 ) -- specialised versions
750 specDefn calls (fn, rhs)
751 -- The first case is the interesting one
752 | n_tyvars == length rhs_tyvars -- Rhs of fn's defn has right number of big lambdas
753 && n_dicts <= length rhs_bndrs -- and enough dict args
754 && not (null calls_for_me) -- And there are some calls to specialise
755 = -- Specialise the body of the function
756 specExpr body `thenSM` \ (body', body_uds) ->
758 (float_uds, bound_uds@(dict_binds,_)) = splitUDs rhs_bndrs body_uds
761 -- Make a specialised version for each call in calls_for_me
762 mapSM (spec_call bound_uds) calls_for_me `thenSM` \ stuff ->
764 (spec_defns, spec_uds, spec_env_stuff) = unzip3 stuff
766 fn' = addIdSpecialisations fn spec_env_stuff
767 rhs' = foldr Lam (foldr Let body' dict_binds) rhs_bndrs
769 returnSM ((fn',rhs'),
771 float_uds `plusUDs` plusUDList spec_uds)
773 | otherwise -- No calls or RHS doesn't fit our preconceptions
774 = specExpr rhs `thenSM` \ (rhs', rhs_uds) ->
775 returnSM ((fn, rhs'), [], rhs_uds)
779 (tyvars, theta, tau) = splitSigmaTy fn_type
780 n_tyvars = length tyvars
781 n_dicts = length theta
782 mk_spec_tys call_ts = zipWith mk_spec_ty call_ts tyvars
784 mk_spec_ty (Just ty) _ = ty
785 mk_spec_ty Nothing tyvar = mkTyVarTy tyvar
787 (rhs_tyvars, rhs_ids, rhs_body) = collectBinders rhs
788 rhs_dicts = take n_dicts rhs_ids
789 rhs_bndrs = map TyBinder rhs_tyvars ++ map ValBinder rhs_dicts
790 body = mkValLam (drop n_dicts rhs_ids) rhs_body
791 -- Glue back on the non-dict lambdas
793 calls_for_me = case lookupFM calls fn of
795 Just cs -> fmToList cs
797 -- Filter out calls for which we already have a specialisation
798 calls_to_spec = filter spec_me calls_for_me
799 spec_me (call_ts, _) = not (maybeToBool (lookupSpecEnv id_spec_env (mk_spec_tys call_ts)))
800 id_spec_env = getIdSpecialisation fn
802 ----------------------------------------------------------
803 -- Specialise to one particular call pattern
804 spec_call :: ProtoUsageDetails -- From the original body, captured by
805 -- the dictionary lambdas
806 -> ([Maybe Type], [DictVar]) -- Call instance
807 -> SpecM ((Id,CoreExpr), -- Specialised definition
808 UsageDetails, -- Usage details from specialised body
809 ([TyVar], [Type], CoreExpr)) -- Info for the Id's SpecEnv
810 spec_call bound_uds (call_ts, call_ds)
811 = ASSERT( length call_ts == n_tyvars && length call_ds == n_dicts )
812 -- Calls are only recorded for properly-saturated applications
814 -- Supppose the call is for f [Just t1, Nothing, Just t3, Nothing] [d1, d2]
816 -- Construct the new binding
817 -- f1 = /\ b d -> (..rhs of f..) t1 b t3 d d1 d2
818 -- and the type of this binder
820 spec_tyvars = [tyvar | (tyvar, Nothing) <- tyvars `zip` call_ts]
821 spec_tys = mk_spec_tys call_ts
822 spec_rhs = mkTyLam spec_tyvars $
823 mkGenApp rhs (map TyArg spec_tys ++ map VarArg call_ds)
824 spec_id_ty = mkForAllTys spec_tyvars (instantiateTy ty_env tau)
825 ty_env = mkTyVarEnv (zipEqual "spec_call" tyvars spec_tys)
827 newIdSM fn spec_id_ty `thenSM` \ spec_f ->
830 -- Construct the stuff for f's spec env
831 -- [b,d] [t1,b,t3,d] |-> \d1 d2 -> f1 b d
832 -- The only awkward bit is that d1,d2 might well be global
833 -- dictionaries, so it's tidier to make new local variables
834 -- for the lambdas in the RHS, rather than lambda-bind the
835 -- dictionaries themselves.
836 mapSM (\d -> newIdSM d (idType d)) call_ds `thenSM` \ arg_ds ->
838 spec_env_rhs = mkValLam arg_ds $
839 mkTyApp (Var spec_f) $
840 map mkTyVarTy spec_tyvars
841 spec_env_info = (spec_tyvars, spec_tys, spec_env_rhs)
844 -- Specialise the UDs from f's RHS
846 -- Only the overloaded tyvars should be free in the uds
847 ty_env = [ (rhs_tyvar,ty)
848 | (rhs_tyvar, Just ty) <- zipEqual "specUDs1" rhs_tyvars call_ts
850 dict_env = zipEqual "specUDs2" rhs_dicts call_ds
852 specUDs ty_env dict_env bound_uds `thenSM` \ spec_uds ->
854 returnSM ((spec_f, spec_rhs),
860 %************************************************************************
862 \subsubsection{UsageDetails and suchlike}
864 %************************************************************************
867 type FreeDicts = IdSet
871 dict_binds :: !(Bag (DictVar, CoreExpr, TyVarSet, FreeDicts)),
872 -- Floated dictionary bindings
873 -- The order is important;
874 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
875 -- (Remember, Bags preserve order in GHC.)
876 -- The FreeDicts is the free vars of the RHS
878 calls :: !CallDetails
881 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM }
883 type ProtoUsageDetails = ([CoreBinding], -- Dict bindings
884 [(Id, [Maybe Type], [DictVar])]
887 ------------------------------------------------------------
888 type CallDetails = FiniteMap Id CallInfo
889 type CallInfo = FiniteMap [Maybe Type] -- Nothing => unconstrained type argument
890 [DictVar] -- Dict args
891 -- The finite maps eliminate duplicates
892 -- The list of types and dictionaries is guaranteed to
893 -- match the type of f
895 callDetailsToList calls = [ (id,tys,dicts)
896 | (id,fm) <- fmToList calls,
897 (tys,dicts) <- fmToList fm
900 listToCallDetails calls = foldr (unionCalls . singleCall) emptyFM calls
902 unionCalls :: CallDetails -> CallDetails -> CallDetails
903 unionCalls c1 c2 = plusFM_C plusFM c1 c2
905 singleCall (id, tys, dicts) = unitFM id (unitFM tys dicts)
909 || length spec_tys /= n_tyvars
910 || length dicts /= n_dicts
911 = emptyUDs -- Not overloaded
914 = MkUD {dict_binds = emptyBag,
915 calls = singleCall (f, spec_tys, dicts)
918 (tyvars, theta, tau) = splitSigmaTy (idType f)
919 constrained_tyvars = foldr (unionTyVarSets . tyVarsOfTypes . snd) emptyTyVarSet theta
920 n_tyvars = length tyvars
921 n_dicts = length theta
923 spec_tys = [mk_spec_ty tv ty | (tv, TyArg ty) <- tyvars `zip` args]
924 dicts = [d | (_, VarArg d) <- theta `zip` (drop n_tyvars args)]
926 mk_spec_ty tyvar ty | tyvar `elementOfTyVarSet` constrained_tyvars
931 ------------------------------------------------------------
932 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
933 plusUDs (MkUD {dict_binds = db1, calls = calls1})
934 (MkUD {dict_binds = db2, calls = calls2})
935 = MkUD {dict_binds, calls}
937 dict_binds = db1 `unionBags` db2
938 calls = calls1 `unionCalls` calls2
940 plusUDList = foldr plusUDs emptyUDs
942 mkDB dict rhs = (dict, rhs, db_ftvs, db_fvs)
944 db_ftvs = tyVarsOfType (idType dict) -- Superset of RHS fvs
945 db_fvs = dictRhsFVs rhs
947 addDictBind uds dict rhs = uds { dict_binds = mkDB dict rhs `consBag` dict_binds uds }
949 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
950 = foldrBag add binds dbs
952 add (dict,rhs,_,_) binds = NonRec dict rhs : binds
954 dumpUDs :: [CoreBinder]
955 -> UsageDetails -> CoreExpr
956 -> (UsageDetails, CoreExpr)
957 dumpUDs bndrs uds body
958 = (free_uds, foldr Let body dict_binds)
960 (free_uds, (dict_binds, _)) = splitUDs bndrs uds
962 splitUDs :: [CoreBinder]
964 -> (UsageDetails, -- These don't mention the binders
965 ProtoUsageDetails) -- These do
967 splitUDs bndrs uds@(MkUD {dict_binds = orig_dbs,
970 = if isEmptyBag dump_dbs && null dump_calls then
971 -- Common case: binder doesn't affect floats
975 -- Binders bind some of the fvs of the floats
976 (MkUD {dict_binds = free_dbs,
977 calls = listToCallDetails free_calls},
978 (bagToList dump_dbs, dump_calls)
982 tyvar_set = mkTyVarSet [tv | TyBinder tv <- bndrs]
983 id_set = mkIdSet [id | ValBinder id <- bndrs]
985 (free_dbs, dump_dbs, dump_idset)
986 = foldlBag dump_db (emptyBag, emptyBag, id_set) orig_dbs
987 -- Important that it's foldl not foldr;
988 -- we're accumulating the set of dumped ids in dump_set
990 -- Filter out any calls that mention things that are being dumped
991 -- Don't need to worry about the tyvars because the dicts will
992 -- spot the captured ones; any fully polymorphic arguments will
993 -- be Nothings in the call details
994 orig_call_list = callDetailsToList orig_calls
995 (dump_calls, free_calls) = partition captured orig_call_list
996 captured (id,tys,dicts) = any (`elementOfIdSet` dump_idset) (id:dicts)
998 dump_db (free_dbs, dump_dbs, dump_idset) db@(dict, rhs, ftvs, fvs)
999 | isEmptyIdSet (dump_idset `intersectIdSets` fvs)
1000 && isEmptyTyVarSet (tyvar_set `intersectTyVarSets` ftvs)
1001 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1003 | otherwise -- Dump it
1004 = (free_dbs, dump_dbs `snocBag` NonRec dict rhs,
1005 dump_idset `addOneToIdSet` dict)
1008 Given a type and value substitution, specUDs creates a specialised copy of
1012 specUDs :: [(TyVar,Type)] -> [(DictVar,DictVar)] -> ProtoUsageDetails -> SpecM UsageDetails
1013 specUDs tv_env_list dict_env_list (dbs, calls)
1014 = specDBs dict_env dbs `thenSM` \ (dict_env', dbs') ->
1015 returnSM (MkUD { dict_binds = dbs',
1016 calls = listToCallDetails (map (inst_call dict_env') calls)
1019 tv_env = mkTyVarEnv tv_env_list
1020 dict_env = mkIdEnv dict_env_list
1022 inst_call dict_env (id, tys, dicts) = (id, map inst_maybe_ty tys,
1023 map (lookupId dict_env) dicts)
1025 inst_maybe_ty Nothing = Nothing
1026 inst_maybe_ty (Just ty) = Just (instantiateTy tv_env ty)
1029 = returnSM (dict_env, emptyBag)
1030 specDBs dict_env (NonRec dict rhs : dbs)
1031 = newIdSM dict (instantiateTy tv_env (idType dict)) `thenSM` \ dict' ->
1033 dict_env' = addOneToIdEnv dict_env dict dict'
1034 rhs' = instantiateDictRhs tv_env dict_env rhs
1036 specDBs dict_env' dbs `thenSM` \ (dict_env'', dbs') ->
1037 returnSM ( dict_env'', mkDB dict' rhs' `consBag` dbs' )
1040 %************************************************************************
1042 \subsubsection{Boring helper functions}
1044 %************************************************************************
1047 lookupId:: IdEnv Id -> Id -> Id
1048 lookupId env id = case lookupIdEnv env id of
1052 instantiateDictRhs :: TyVarEnv Type -> IdEnv Id -> CoreExpr -> CoreExpr
1053 -- Cheapo function for simple RHSs
1054 instantiateDictRhs ty_env id_env rhs
1057 go_arg (VarArg a) = VarArg (lookupId id_env a)
1058 go_arg (TyArg t) = TyArg (instantiateTy ty_env t)
1060 go (App e1 arg) = App (go e1) (go_arg arg)
1061 go (Var v) = Var (lookupId id_env v)
1063 go (Con con args) = Con con (map go_arg args)
1064 go (Note n e) = Note (go_note n) (go e)
1065 go (Case e alts) = Case (go e) alts -- See comment below re alts
1066 go other = pprPanic "instantiateDictRhs" (ppr rhs)
1068 go_note (Coerce t1 t2) = Coerce (instantiateTy ty_env t1) (instantiateTy ty_env t2)
1071 dictRhsFVs :: CoreExpr -> IdSet
1072 -- Cheapo function for simple RHSs
1076 go (App e1 (VarArg a)) = go e1 `addOneToIdSet` a
1077 go (App e1 (TyArg t)) = go e1
1078 go (Var v) = unitIdSet v
1079 go (Lit l) = emptyIdSet
1080 go (Con _ args) = mkIdSet [id | VarArg id <- args]
1081 go (Note _ e) = go e
1083 go (Case e _) = go e -- Claim: no free dictionaries in the alternatives
1084 -- These case expressions are of the form
1085 -- case d of { D a b c -> b }
1087 go other = pprPanic "dictRhsFVs" (ppr e)
1090 addIdSpecialisations id spec_stuff
1091 = (if not (null errs) then
1092 pprTrace "Duplicate specialisations" (vcat (map ppr errs))
1095 setIdSpecialisation id new_spec_env
1097 (new_spec_env, errs) = foldr add (getIdSpecialisation id, []) spec_stuff
1099 add (tyvars, tys, template) (spec_env, errs)
1100 = case addToSpecEnv True spec_env tyvars tys template of
1101 Succeeded spec_env' -> (spec_env', errs)
1102 Failed err -> (spec_env, err:errs)
1104 -- Given an Id, isSpecVars returns all its specialisations.
1105 -- We extract these from its SpecEnv.
1106 -- This is used by the occurrence analyser and free-var finder;
1107 -- we regard an Id's specialisations as free in the Id's definition.
1109 idSpecVars :: Id -> [Id]
1111 = map get_spec (specEnvValues (getIdSpecialisation id))
1113 -- get_spec is another cheapo function like dictRhsFVs
1114 -- It knows what these specialisation temlates look like,
1115 -- and just goes for the jugular
1116 get_spec (App f _) = get_spec f
1117 get_spec (Lam _ b) = get_spec b
1118 get_spec (Var v) = v
1120 ----------------------------------------
1121 type SpecM a = UniqSM a
1125 getUniqSM = getUnique
1129 mapAndCombineSM f [] = returnSM ([], emptyUDs)
1130 mapAndCombineSM f (x:xs) = f x `thenSM` \ (y, uds1) ->
1131 mapAndCombineSM f xs `thenSM` \ (ys, uds2) ->
1132 returnSM (y:ys, uds1 `plusUDs` uds2)
1134 newIdSM old_id new_ty
1135 = getUnique `thenSM` \ uniq ->
1136 returnSM (mkUserLocal (getOccName old_id)
1144 Old (but interesting) stuff about unboxed bindings
1145 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1147 What should we do when a value is specialised to a *strict* unboxed value?
1149 map_*_* f (x:xs) = let h = f x
1153 Could convert let to case:
1155 map_*_Int# f (x:xs) = case f x of h# ->
1159 This may be undesirable since it forces evaluation here, but the value
1160 may not be used in all branches of the body. In the general case this
1161 transformation is impossible since the mutual recursion in a letrec
1162 cannot be expressed as a case.
1164 There is also a problem with top-level unboxed values, since our
1165 implementation cannot handle unboxed values at the top level.
1167 Solution: Lift the binding of the unboxed value and extract it when it
1170 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1175 Now give it to the simplifier and the _Lifting will be optimised away.
1177 The benfit is that we have given the specialised "unboxed" values a
1178 very simplep lifted semantics and then leave it up to the simplifier to
1179 optimise it --- knowing that the overheads will be removed in nearly
1182 In particular, the value will only be evaluted in the branches of the
1183 program which use it, rather than being forced at the point where the
1184 value is bound. For example:
1186 filtermap_*_* p f (x:xs)
1193 filtermap_*_Int# p f (x:xs)
1194 = let h = case (f x) of h# -> _Lift h#
1197 True -> case h of _Lift h#
1201 The binding for h can still be inlined in the one branch and the
1202 _Lifting eliminated.
1205 Question: When won't the _Lifting be eliminated?
1207 Answer: When they at the top-level (where it is necessary) or when
1208 inlining would duplicate work (or possibly code depending on
1209 options). However, the _Lifting will still be eliminated if the
1210 strictness analyser deems the lifted binding strict.