2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
4 \section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
7 module Specialise ( specProgram ) where
9 #include "HsVersions.h"
11 import CmdLineOpts ( opt_D_verbose_core2core, opt_D_dump_spec )
12 import Id ( Id, idName, idType, mkTemplateLocals, mkUserLocal,
13 getIdSpecialisation, setIdSpecialisation,
19 import Type ( Type, TyVarSubst, mkTyVarTy, splitSigmaTy, substTy,
20 fullSubstTy, tyVarsOfType, tyVarsOfTypes,
21 mkForAllTys, boxedTypeKind
23 import Var ( TyVar, mkSysTyVar, setVarUnique )
27 import CoreUtils ( IdSubst, SubstCoreExpr(..), exprFreeVars,
28 substExpr, substId, substIds, coreExprType
30 import CoreLint ( beginPass, endPass )
31 import PprCore () -- Instances
32 import SpecEnv ( addToSpecEnv )
34 import UniqSupply ( UniqSupply,
35 UniqSM, initUs_, thenUs, thenUs_, returnUs, getUniqueUs,
36 getUs, setUs, uniqFromSupply, splitUniqSupply, mapUs
38 import Name ( nameOccName, mkSpecOcc, getSrcLoc )
40 import Maybes ( MaybeErr(..), catMaybes )
42 import List ( partition )
43 import Util ( zipEqual, mapAccumL )
50 %************************************************************************
52 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
54 %************************************************************************
56 These notes describe how we implement specialisation to eliminate
59 The specialisation pass works on Core
60 syntax, complete with all the explicit dictionary application,
61 abstraction and construction as added by the type checker. The
62 existing type checker remains largely as it is.
64 One important thought: the {\em types} passed to an overloaded
65 function, and the {\em dictionaries} passed are mutually redundant.
66 If the same function is applied to the same type(s) then it is sure to
67 be applied to the same dictionary(s)---or rather to the same {\em
68 values}. (The arguments might look different but they will evaluate
71 Second important thought: we know that we can make progress by
72 treating dictionary arguments as static and worth specialising on. So
73 we can do without binding-time analysis, and instead specialise on
74 dictionary arguments and no others.
83 and suppose f is overloaded.
85 STEP 1: CALL-INSTANCE COLLECTION
87 We traverse <body>, accumulating all applications of f to types and
90 (Might there be partial applications, to just some of its types and
91 dictionaries? In principle yes, but in practice the type checker only
92 builds applications of f to all its types and dictionaries, so partial
93 applications could only arise as a result of transformation, and even
94 then I think it's unlikely. In any case, we simply don't accumulate such
95 partial applications.)
100 So now we have a collection of calls to f:
104 Notice that f may take several type arguments. To avoid ambiguity, we
105 say that f is called at type t1/t2 and t3/t4.
107 We take equivalence classes using equality of the *types* (ignoring
108 the dictionary args, which as mentioned previously are redundant).
110 STEP 3: SPECIALISATION
112 For each equivalence class, choose a representative (f t1 t2 d1 d2),
113 and create a local instance of f, defined thus:
115 f@t1/t2 = <f_rhs> t1 t2 d1 d2
117 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
118 of simplification will now result. However we don't actually *do* that
119 simplification. Rather, we leave it for the simplifier to do. If we
120 *did* do it, though, we'd get more call instances from the specialised
121 RHS. We can work out what they are by instantiating the call-instance
122 set from f's RHS with the types t1, t2.
124 Add this new id to f's IdInfo, to record that f has a specialised version.
126 Before doing any of this, check that f's IdInfo doesn't already
127 tell us about an existing instance of f at the required type/s.
128 (This might happen if specialisation was applied more than once, or
129 it might arise from user SPECIALIZE pragmas.)
133 Wait a minute! What if f is recursive? Then we can't just plug in
134 its right-hand side, can we?
136 But it's ok. The type checker *always* creates non-recursive definitions
137 for overloaded recursive functions. For example:
139 f x = f (x+x) -- Yes I know its silly
143 f a (d::Num a) = let p = +.sel a d
145 letrec fl (y::a) = fl (p y y)
149 We still have recusion for non-overloaded functions which we
150 speciailise, but the recursive call should get specialised to the
151 same recursive version.
157 All this is crystal clear when the function is applied to *constant
158 types*; that is, types which have no type variables inside. But what if
159 it is applied to non-constant types? Suppose we find a call of f at type
160 t1/t2. There are two possibilities:
162 (a) The free type variables of t1, t2 are in scope at the definition point
163 of f. In this case there's no problem, we proceed just as before. A common
164 example is as follows. Here's the Haskell:
169 After typechecking we have
171 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
172 in +.sel a d (f a d y) (f a d y)
174 Notice that the call to f is at type type "a"; a non-constant type.
175 Both calls to f are at the same type, so we can specialise to give:
177 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
178 in +.sel a d (f@a y) (f@a y)
181 (b) The other case is when the type variables in the instance types
182 are *not* in scope at the definition point of f. The example we are
183 working with above is a good case. There are two instances of (+.sel a d),
184 but "a" is not in scope at the definition of +.sel. Can we do anything?
185 Yes, we can "common them up", a sort of limited common sub-expression deal.
188 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
189 f@a (x::a) = +.sel@a x x
190 in +.sel@a (f@a y) (f@a y)
192 This can save work, and can't be spotted by the type checker, because
193 the two instances of +.sel weren't originally at the same type.
197 * There are quite a few variations here. For example, the defn of
198 +.sel could be floated ouside the \y, to attempt to gain laziness.
199 It certainly mustn't be floated outside the \d because the d has to
202 * We don't want to inline f_rhs in this case, because
203 that will duplicate code. Just commoning up the call is the point.
205 * Nothing gets added to +.sel's IdInfo.
207 * Don't bother unless the equivalence class has more than one item!
209 Not clear whether this is all worth it. It is of course OK to
210 simply discard call-instances when passing a big lambda.
212 Polymorphism 2 -- Overloading
214 Consider a function whose most general type is
216 f :: forall a b. Ord a => [a] -> b -> b
218 There is really no point in making a version of g at Int/Int and another
219 at Int/Bool, because it's only instancing the type variable "a" which
220 buys us any efficiency. Since g is completely polymorphic in b there
221 ain't much point in making separate versions of g for the different
224 That suggests that we should identify which of g's type variables
225 are constrained (like "a") and which are unconstrained (like "b").
226 Then when taking equivalence classes in STEP 2, we ignore the type args
227 corresponding to unconstrained type variable. In STEP 3 we make
228 polymorphic versions. Thus:
230 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
239 f a (d::Num a) = let g = ...
241 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
243 Here, g is only called at one type, but the dictionary isn't in scope at the
244 definition point for g. Usually the type checker would build a
245 definition for d1 which enclosed g, but the transformation system
246 might have moved d1's defn inward. Solution: float dictionary bindings
247 outwards along with call instances.
251 f x = let g p q = p==q
257 Before specialisation, leaving out type abstractions we have
259 f df x = let g :: Eq a => a -> a -> Bool
261 h :: Num a => a -> a -> (a, Bool)
262 h dh r s = let deq = eqFromNum dh
263 in (+ dh r s, g deq r s)
267 After specialising h we get a specialised version of h, like this:
269 h' r s = let deq = eqFromNum df
270 in (+ df r s, g deq r s)
272 But we can't naively make an instance for g from this, because deq is not in scope
273 at the defn of g. Instead, we have to float out the (new) defn of deq
274 to widen its scope. Notice that this floating can't be done in advance -- it only
275 shows up when specialisation is done.
277 User SPECIALIZE pragmas
278 ~~~~~~~~~~~~~~~~~~~~~~~
279 Specialisation pragmas can be digested by the type checker, and implemented
280 by adding extra definitions along with that of f, in the same way as before
282 f@t1/t2 = <f_rhs> t1 t2 d1 d2
284 Indeed the pragmas *have* to be dealt with by the type checker, because
285 only it knows how to build the dictionaries d1 and d2! For example
287 g :: Ord a => [a] -> [a]
288 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
290 Here, the specialised version of g is an application of g's rhs to the
291 Ord dictionary for (Tree Int), which only the type checker can conjure
292 up. There might not even *be* one, if (Tree Int) is not an instance of
293 Ord! (All the other specialision has suitable dictionaries to hand
296 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
297 it is buried in a complex (as-yet-un-desugared) binding group.
300 f@t1/t2 = f* t1 t2 d1 d2
302 where f* is the Id f with an IdInfo which says "inline me regardless!".
303 Indeed all the specialisation could be done in this way.
304 That in turn means that the simplifier has to be prepared to inline absolutely
305 any in-scope let-bound thing.
308 Again, the pragma should permit polymorphism in unconstrained variables:
310 h :: Ord a => [a] -> b -> b
311 {-# SPECIALIZE h :: [Int] -> b -> b #-}
313 We *insist* that all overloaded type variables are specialised to ground types,
314 (and hence there can be no context inside a SPECIALIZE pragma).
315 We *permit* unconstrained type variables to be specialised to
317 - or left as a polymorphic type variable
318 but nothing in between. So
320 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
322 is *illegal*. (It can be handled, but it adds complication, and gains the
326 SPECIALISING INSTANCE DECLARATIONS
327 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
330 instance Foo a => Foo [a] where
332 {-# SPECIALIZE instance Foo [Int] #-}
334 The original instance decl creates a dictionary-function
337 dfun.Foo.List :: forall a. Foo a -> Foo [a]
339 The SPECIALIZE pragma just makes a specialised copy, just as for
340 ordinary function definitions:
342 dfun.Foo.List@Int :: Foo [Int]
343 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
345 The information about what instance of the dfun exist gets added to
346 the dfun's IdInfo in the same way as a user-defined function too.
349 Automatic instance decl specialisation?
350 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
351 Can instance decls be specialised automatically? It's tricky.
352 We could collect call-instance information for each dfun, but
353 then when we specialised their bodies we'd get new call-instances
354 for ordinary functions; and when we specialised their bodies, we might get
355 new call-instances of the dfuns, and so on. This all arises because of
356 the unrestricted mutual recursion between instance decls and value decls.
358 Still, there's no actual problem; it just means that we may not do all
359 the specialisation we could theoretically do.
361 Furthermore, instance decls are usually exported and used non-locally,
362 so we'll want to compile enough to get those specialisations done.
364 Lastly, there's no such thing as a local instance decl, so we can
365 survive solely by spitting out *usage* information, and then reading that
366 back in as a pragma when next compiling the file. So for now,
367 we only specialise instance decls in response to pragmas.
370 SPITTING OUT USAGE INFORMATION
371 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
373 To spit out usage information we need to traverse the code collecting
374 call-instance information for all imported (non-prelude?) functions
375 and data types. Then we equivalence-class it and spit it out.
377 This is done at the top-level when all the call instances which escape
378 must be for imported functions and data types.
380 *** Not currently done ***
383 Partial specialisation by pragmas
384 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
385 What about partial specialisation:
387 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
388 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
392 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
394 Seems quite reasonable. Similar things could be done with instance decls:
396 instance (Foo a, Foo b) => Foo (a,b) where
398 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
399 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
401 Ho hum. Things are complex enough without this. I pass.
404 Requirements for the simplifer
405 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
406 The simplifier has to be able to take advantage of the specialisation.
408 * When the simplifier finds an application of a polymorphic f, it looks in
409 f's IdInfo in case there is a suitable instance to call instead. This converts
411 f t1 t2 d1 d2 ===> f_t1_t2
413 Note that the dictionaries get eaten up too!
415 * Dictionary selection operations on constant dictionaries must be
418 +.sel Int d ===> +Int
420 The obvious way to do this is in the same way as other specialised
421 calls: +.sel has inside it some IdInfo which tells that if it's applied
422 to the type Int then it should eat a dictionary and transform to +Int.
424 In short, dictionary selectors need IdInfo inside them for constant
427 * Exactly the same applies if a superclass dictionary is being
430 Eq.sel Int d ===> dEqInt
432 * Something similar applies to dictionary construction too. Suppose
433 dfun.Eq.List is the function taking a dictionary for (Eq a) to
434 one for (Eq [a]). Then we want
436 dfun.Eq.List Int d ===> dEq.List_Int
438 Where does the Eq [Int] dictionary come from? It is built in
439 response to a SPECIALIZE pragma on the Eq [a] instance decl.
441 In short, dfun Ids need IdInfo with a specialisation for each
442 constant instance of their instance declaration.
444 All this uses a single mechanism: the SpecEnv inside an Id
447 What does the specialisation IdInfo look like?
448 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
450 The SpecEnv of an Id maps a list of types (the template) to an expression
454 For example, if f has this SpecInfo:
456 [Int, a] -> \d:Ord Int. f' a
458 it means that we can replace the call
460 f Int t ===> (\d. f' t)
462 This chucks one dictionary away and proceeds with the
463 specialised version of f, namely f'.
466 What can't be done this way?
467 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
468 There is no way, post-typechecker, to get a dictionary for (say)
469 Eq a from a dictionary for Eq [a]. So if we find
473 we can't transform to
478 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
480 Of course, we currently have no way to automatically derive
481 eqList, nor to connect it to the Eq [a] instance decl, but you
482 can imagine that it might somehow be possible. Taking advantage
483 of this is permanently ruled out.
485 Still, this is no great hardship, because we intend to eliminate
486 overloading altogether anyway!
490 A note about non-tyvar dictionaries
491 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
492 Some Ids have types like
494 forall a,b,c. Eq a -> Ord [a] -> tau
496 This seems curious at first, because we usually only have dictionary
497 args whose types are of the form (C a) where a is a type variable.
498 But this doesn't hold for the functions arising from instance decls,
499 which sometimes get arguements with types of form (C (T a)) for some
502 Should we specialise wrt this compound-type dictionary? We used to say
504 "This is a heuristic judgement, as indeed is the fact that we
505 specialise wrt only dictionaries. We choose *not* to specialise
506 wrt compound dictionaries because at the moment the only place
507 they show up is in instance decls, where they are simply plugged
508 into a returned dictionary. So nothing is gained by specialising
511 But it is simpler and more uniform to specialise wrt these dicts too;
512 and in future GHC is likely to support full fledged type signatures
514 f ;: Eq [(a,b)] => ...
517 %************************************************************************
519 \subsubsection{The new specialiser}
521 %************************************************************************
523 Our basic game plan is this. For let(rec) bound function
524 f :: (C a, D c) => (a,b,c,d) -> Bool
526 * Find any specialised calls of f, (f ts ds), where
527 ts are the type arguments t1 .. t4, and
528 ds are the dictionary arguments d1 .. d2.
530 * Add a new definition for f1 (say):
532 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
534 Note that we abstract over the unconstrained type arguments.
538 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
540 to the specialisations of f. This will be used by the
541 simplifier to replace calls
542 (f t1 t2 t3 t4) da db
544 (\d1 d1 -> f1 t2 t4) da db
546 All the stuff about how many dictionaries to discard, and what types
547 to apply the specialised function to, are handled by the fact that the
548 SpecEnv contains a template for the result of the specialisation.
550 We don't build *partial* specialisations for f. For example:
552 f :: Eq a => a -> a -> Bool
553 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
555 Here, little is gained by making a specialised copy of f.
556 There's a distinct danger that the specialised version would
557 first build a dictionary for (Eq b, Eq c), and then select the (==)
558 method from it! Even if it didn't, not a great deal is saved.
560 We do, however, generate polymorphic, but not overloaded, specialisations:
562 f :: Eq a => [a] -> b -> b -> b
563 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
565 Hence, the invariant is this:
567 *** no specialised version is overloaded ***
570 %************************************************************************
572 \subsubsection{The exported function}
574 %************************************************************************
577 specProgram :: UniqSupply -> [CoreBind] -> IO [CoreBind]
580 beginPass "Specialise"
582 let binds' = initSM us (go binds `thenSM` \ (binds', uds') ->
583 returnSM (dumpAllDictBinds uds' binds'))
585 endPass "Specialise" (opt_D_dump_spec || opt_D_verbose_core2core) binds'
588 go [] = returnSM ([], emptyUDs)
589 go (bind:binds) = go binds `thenSM` \ (binds', uds) ->
590 specBind bind uds `thenSM` \ (bind', uds') ->
591 returnSM (bind' ++ binds', uds')
594 %************************************************************************
596 \subsubsection{@specExpr@: the main function}
598 %************************************************************************
601 specExpr :: CoreExpr -> SpecM (CoreExpr, UsageDetails)
603 ---------------- First the easy cases --------------------
604 specExpr e@(Type _) = returnSM (e, emptyUDs)
605 specExpr e@(Var _) = returnSM (e, emptyUDs)
607 specExpr e@(Con con args)
608 = mapAndCombineSM specExpr args `thenSM` \ (args', uds) ->
609 returnSM (Con con args', uds)
611 specExpr (Note note body)
612 = specExpr body `thenSM` \ (body', uds) ->
613 returnSM (Note note body', uds)
616 ---------------- Applications might generate a call instance --------------------
617 specExpr expr@(App fun arg)
620 go (App fun arg) args = specExpr arg `thenSM` \ (arg', uds_arg) ->
621 go fun (arg':args) `thenSM` \ (fun', uds_app) ->
622 returnSM (App fun' arg', uds_arg `plusUDs` uds_app)
624 go (Var f) args = returnSM (Var f, mkCallUDs f args)
625 go other args = specExpr other
627 ---------------- Lambda/case require dumping of usage details --------------------
629 = specExpr body `thenSM` \ (body', uds) ->
631 (filtered_uds, body'') = dumpUDs bndrs uds body'
633 returnSM (mkLams bndrs body'', filtered_uds)
635 (bndrs, body) = go [] e
637 -- More efficient to collect a group of binders together all at once
638 -- and we don't want to split a lambda group with dumped bindings
639 go bndrs (Lam bndr e) = go (bndr:bndrs) e
640 go bndrs e = (reverse bndrs, e)
643 specExpr (Case scrut case_bndr alts)
644 = specExpr scrut `thenSM` \ (scrut', uds_scrut) ->
645 mapAndCombineSM spec_alt alts `thenSM` \ (alts', uds_alts) ->
646 returnSM (Case scrut' case_bndr alts', uds_scrut `plusUDs` uds_alts)
648 spec_alt (con, args, rhs)
649 = specExpr rhs `thenSM` \ (rhs', uds) ->
651 (uds', rhs'') = dumpUDs args uds rhs'
653 returnSM ((con, args, rhs''), uds')
655 ---------------- Finally, let is the interesting case --------------------
656 specExpr (Let bind body)
657 = -- Deal with the body
658 specExpr body `thenSM` \ (body', body_uds) ->
660 -- Deal with the bindings
661 specBind bind body_uds `thenSM` \ (binds', uds) ->
664 returnSM (foldr Let body' binds', uds)
667 %************************************************************************
669 \subsubsection{Dealing with a binding}
671 %************************************************************************
675 -> UsageDetails -- Info on how the scope of the binding
676 -> SpecM ([CoreBind], -- New bindings
677 UsageDetails) -- And info to pass upstream
679 specBind bind@(NonRec bndr rhs) body_uds
680 | isSpecPragmaId bndr -- Aha! A spec-pragma Id. Collect UDs from
681 -- its RHS and discard it!
682 = specExpr rhs `thenSM` \ (rhs', rhs_uds) ->
683 returnSM ([], rhs_uds `plusUDs` body_uds)
686 specBind bind body_uds
687 = specBindItself bind (calls body_uds) `thenSM` \ (bind', bind_uds) ->
689 bndrs = bindersOf bind
690 all_uds = zapCalls bndrs (body_uds `plusUDs` bind_uds)
691 -- It's important that the `plusUDs` is this way round,
692 -- because body_uds may bind dictionaries that are
693 -- used in the calls passed to specDefn. So the
694 -- dictionary bindings in bind_uds may mention
695 -- dictionaries bound in body_uds.
697 case splitUDs bndrs all_uds of
699 (_, ([],[])) -- This binding doesn't bind anything needed
700 -- in the UDs, so put the binding here
701 -- This is the case for most non-dict bindings, except
702 -- for the few that are mentioned in a dict binding
703 -- that is floating upwards in body_uds
704 -> returnSM ([bind'], all_uds)
706 (float_uds, (dict_binds, calls)) -- This binding is needed in the UDs, so float it out
707 -> returnSM ([], float_uds `plusUDs` mkBigUD bind' dict_binds calls)
710 -- A truly gruesome function
711 mkBigUD bind@(NonRec _ _) dbs calls
712 = -- Common case: non-recursive and no specialisations
713 -- (if there were any specialistions it would have been made recursive)
714 MkUD { dict_binds = listToBag (mkDB bind : dbs),
715 calls = listToCallDetails calls }
717 mkBigUD bind dbs calls
719 MkUD { dict_binds = unitBag (mkDB (Rec (bind_prs bind ++ dbsToPairs dbs))),
721 calls = listToCallDetails calls }
723 bind_prs (NonRec b r) = [(b,r)]
724 bind_prs (Rec prs) = prs
727 dbsToPairs ((bind,_):dbs) = bind_prs bind ++ dbsToPairs dbs
729 -- specBindItself deals with the RHS, specialising it according
730 -- to the calls found in the body (if any)
731 specBindItself (NonRec bndr rhs) call_info
732 = specDefn call_info (bndr,rhs) `thenSM` \ ((bndr',rhs'), spec_defns, spec_uds) ->
734 new_bind | null spec_defns = NonRec bndr' rhs'
735 | otherwise = Rec ((bndr',rhs'):spec_defns)
736 -- bndr' mentions the spec_defns in its SpecEnv
737 -- Not sure why we couln't just put the spec_defns first
739 returnSM (new_bind, spec_uds)
741 specBindItself (Rec pairs) call_info
742 = mapSM (specDefn call_info) pairs `thenSM` \ stuff ->
744 (pairs', spec_defns_s, spec_uds_s) = unzip3 stuff
745 spec_defns = concat spec_defns_s
746 spec_uds = plusUDList spec_uds_s
747 new_bind = Rec (spec_defns ++ pairs')
749 returnSM (new_bind, spec_uds)
752 specDefn :: CallDetails -- Info on how it is used in its scope
753 -> (Id, CoreExpr) -- The thing being bound and its un-processed RHS
754 -> SpecM ((Id, CoreExpr), -- The thing and its processed RHS
755 -- the Id may now have specialisations attached
756 [(Id,CoreExpr)], -- Extra, specialised bindings
757 UsageDetails -- Stuff to fling upwards from the RHS and its
758 ) -- specialised versions
760 specDefn calls (fn, rhs)
761 -- The first case is the interesting one
762 | n_tyvars == length rhs_tyvars -- Rhs of fn's defn has right number of big lambdas
763 && n_dicts <= length rhs_bndrs -- and enough dict args
764 && not (null calls_for_me) -- And there are some calls to specialise
765 = -- Specialise the body of the function
766 specExpr body `thenSM` \ (body', body_uds) ->
768 (float_uds, bound_uds@(dict_binds,_)) = splitUDs rhs_bndrs body_uds
771 -- Make a specialised version for each call in calls_for_me
772 mapSM (spec_call bound_uds) calls_for_me `thenSM` \ stuff ->
774 (spec_defns, spec_uds, spec_env_stuff) = unzip3 stuff
776 fn' = addIdSpecialisations fn spec_env_stuff
777 rhs' = mkLams rhs_bndrs (mkDictLets dict_binds body')
779 returnSM ((fn',rhs'),
781 float_uds `plusUDs` plusUDList spec_uds)
783 | otherwise -- No calls or RHS doesn't fit our preconceptions
784 = specExpr rhs `thenSM` \ (rhs', rhs_uds) ->
785 returnSM ((fn, rhs'), [], rhs_uds)
789 (tyvars, theta, tau) = splitSigmaTy fn_type
790 n_tyvars = length tyvars
791 n_dicts = length theta
793 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs
794 rhs_dicts = take n_dicts rhs_ids
795 rhs_bndrs = rhs_tyvars ++ rhs_dicts
796 body = mkLams (drop n_dicts rhs_ids) rhs_body
797 -- Glue back on the non-dict lambdas
799 calls_for_me = case lookupFM calls fn of
801 Just cs -> fmToList cs
803 ----------------------------------------------------------
804 -- Specialise to one particular call pattern
805 spec_call :: ProtoUsageDetails -- From the original body, captured by
806 -- the dictionary lambdas
807 -> ([Maybe Type], ([DictExpr], IdOrTyVarSet)) -- Call instance
808 -> SpecM ((Id,CoreExpr), -- Specialised definition
809 UsageDetails, -- Usage details from specialised body
810 ([TyVar], [Type], CoreExpr)) -- Info for the Id's SpecEnv
811 spec_call bound_uds (call_ts, (call_ds, _))
812 = ASSERT( length call_ts == n_tyvars && length call_ds == n_dicts )
813 -- Calls are only recorded for properly-saturated applications
815 -- Supppose the call is for f [Just t1, Nothing, Just t3, Nothing] [d1, d2]
817 -- Construct the new binding
818 -- f1 = /\ b d -> (..rhs of f..) t1 b t3 d d1 d2
819 -- and the type of this binder
821 mk_spec_ty Nothing = newTyVarSM `thenSM` \ tyvar ->
822 returnSM (Just tyvar, mkTyVarTy tyvar)
823 mk_spec_ty (Just ty) = returnSM (Nothing, ty)
825 mapSM mk_spec_ty call_ts `thenSM` \ stuff ->
827 (maybe_spec_tyvars, spec_tys) = unzip stuff
828 spec_tyvars = catMaybes maybe_spec_tyvars
829 spec_id_ty = mkForAllTys spec_tyvars
830 (substTy (zipVarEnv tyvars spec_tys) tau)
831 -- NB When substituting in tau we need a ty_env mentioning tyvars
832 -- but when substituting in UDs we need a ty_evn mentioning rhs_tyvars
833 ud_ty_env = zipVarEnv rhs_tyvars spec_tys
834 ud_dict_env = zipVarEnv rhs_dicts (map Done call_ds)
836 -- Only the overloaded tyvars should be free in the uds
837 ty_env = mkVarEnv [ (rhs_tyvar, ty)
838 | (rhs_tyvar, Just ty) <- zipEqual "specUDs1" rhs_tyvars call_ts
843 -- Specialise the UDs from f's RHS
844 specUDs ud_ty_env ud_dict_env bound_uds `thenSM` \ spec_uds ->
847 -- Construct the stuff for f's spec env
848 -- [b,d] [t1,b,t3,d] |-> \d1 d2 -> f1 b d
849 -- The only awkward bit is that d1,d2 might well be global
850 -- dictionaries, so it's tidier to make new local variables
851 -- for the lambdas in the RHS, rather than lambda-bind the
852 -- dictionaries themselves.
854 -- In fact we use the standard template locals, so that the
855 -- they don't need to be "tidied" before putting in interface files
856 newIdSM fn spec_id_ty `thenSM` \ spec_f ->
858 arg_ds = mkTemplateLocals (map coreExprType call_ds)
859 spec_env_rhs = mkLams arg_ds $
860 mkTyApps (Var spec_f) $
861 map mkTyVarTy spec_tyvars
862 spec_env_info = (spec_tyvars, spec_tys, spec_env_rhs)
865 -- Finally construct f's RHS
866 -- Annoyingly, the specialised UDs may mention some of the *un* specialised
867 -- type variables. Here's a case that came up in nofib/spectral/typech98:
868 -- f = /\m a -> \d:Monad m -> let d':Monad (T m a) = ...a... in ...
869 -- When we try to make a specialised verison of f, from a call pattern
871 -- where ? is the Nothing for an unspecialised position, we must get
872 -- spec_f = /\ a -> let d':Monad (T Maybe a) = ...a... in ....
873 -- If we don't do the splitUDs below, the d' binding floats out too far.
874 -- Sigh. What a mess.
876 (float_uds, (dict_binds,_)) = splitUDs spec_tyvars spec_uds
878 spec_rhs = mkLams spec_tyvars $
879 mkDictLets dict_binds $
880 mkApps rhs (map Type spec_tys ++ call_ds)
882 returnSM ((spec_f, spec_rhs),
888 %************************************************************************
890 \subsubsection{UsageDetails and suchlike}
892 %************************************************************************
895 type FreeDicts = IdSet
899 dict_binds :: !(Bag DictBind),
900 -- Floated dictionary bindings
901 -- The order is important;
902 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
903 -- (Remember, Bags preserve order in GHC.)
904 -- The FreeDicts is the free vars of the RHS
906 calls :: !CallDetails
909 type DictBind = (CoreBind, IdOrTyVarSet)
910 -- The set is the free vars of the binding
911 -- both tyvars and dicts
913 type DictExpr = CoreExpr
915 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM }
917 type ProtoUsageDetails = ([DictBind],
918 [(Id, [Maybe Type], ([DictExpr], IdOrTyVarSet))]
921 ------------------------------------------------------------
922 type CallDetails = FiniteMap Id CallInfo
923 type CallInfo = FiniteMap [Maybe Type] -- Nothing => unconstrained type argument
924 ([DictExpr], IdSet) -- Dict args and the free dicts
925 -- free dicts does *not* include the main id itself
926 -- The finite maps eliminate duplicates
927 -- The list of types and dictionaries is guaranteed to
928 -- match the type of f
930 unionCalls :: CallDetails -> CallDetails -> CallDetails
931 unionCalls c1 c2 = plusFM_C plusFM c1 c2
933 singleCall (id, tys, dicts)
934 = unitFM id (unitFM tys (dicts, dict_fvs))
936 dict_fvs = foldr (unionVarSet . exprFreeVars) emptyVarSet dicts
937 -- The type args (tys) are guaranteed to be part of the dictionary
938 -- types, because they are just the constrained types,
939 -- and the dictionary is therefore sure to be bound
940 -- inside the binding for any type variables free in the type;
941 -- hence it's safe to neglect tyvars free in tys when making
942 -- the free-var set for this call
944 -- We don't include the 'id' itself.
946 listToCallDetails calls
947 = foldr (unionCalls . mk_call) emptyFM calls
949 mk_call (id, tys, dicts_w_fvs) = unitFM id (unitFM tys dicts_w_fvs)
950 -- NB: the free vars of the call are provided
952 callDetailsToList calls = [ (id,tys,dicts)
953 | (id,fm) <- fmToList calls,
954 (tys,dicts) <- fmToList fm
959 || length spec_tys /= n_tyvars
960 || length dicts /= n_dicts
961 = emptyUDs -- Not overloaded
964 = MkUD {dict_binds = emptyBag,
965 calls = singleCall (f, spec_tys, dicts)
968 (tyvars, theta, tau) = splitSigmaTy (idType f)
969 constrained_tyvars = foldr (unionVarSet . tyVarsOfTypes . snd) emptyVarSet theta
970 n_tyvars = length tyvars
971 n_dicts = length theta
973 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
974 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
976 mk_spec_ty tyvar ty | tyvar `elemVarSet` constrained_tyvars
981 ------------------------------------------------------------
982 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
983 plusUDs (MkUD {dict_binds = db1, calls = calls1})
984 (MkUD {dict_binds = db2, calls = calls2})
985 = MkUD {dict_binds = d, calls = c}
987 d = db1 `unionBags` db2
988 c = calls1 `unionCalls` calls2
990 plusUDList = foldr plusUDs emptyUDs
992 -- zapCalls deletes calls to ids from uds
993 zapCalls ids uds = uds {calls = delListFromFM (calls uds) ids}
995 mkDB bind = (bind, bind_fvs bind)
997 bind_fvs (NonRec bndr rhs) = exprFreeVars rhs
998 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs (map fst prs)
1000 rhs_fvs = foldr (unionVarSet . exprFreeVars . snd) emptyVarSet prs
1002 addDictBind uds bind = uds { dict_binds = mkDB bind `consBag` dict_binds uds }
1004 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
1005 = foldrBag add binds dbs
1007 add (bind,_) binds = bind : binds
1009 mkDictBinds :: [DictBind] -> [CoreBind]
1010 mkDictBinds = map fst
1012 mkDictLets :: [DictBind] -> CoreExpr -> CoreExpr
1013 mkDictLets dbs body = foldr mk body dbs
1015 mk (bind,_) e = Let bind e
1017 dumpUDs :: [CoreBndr]
1018 -> UsageDetails -> CoreExpr
1019 -> (UsageDetails, CoreExpr)
1020 dumpUDs bndrs uds body
1021 = (free_uds, mkDictLets dict_binds body)
1023 (free_uds, (dict_binds, _)) = splitUDs bndrs uds
1025 splitUDs :: [CoreBndr]
1027 -> (UsageDetails, -- These don't mention the binders
1028 ProtoUsageDetails) -- These do
1030 splitUDs bndrs uds@(MkUD {dict_binds = orig_dbs,
1031 calls = orig_calls})
1033 = if isEmptyBag dump_dbs && null dump_calls then
1034 -- Common case: binder doesn't affect floats
1038 -- Binders bind some of the fvs of the floats
1039 (MkUD {dict_binds = free_dbs,
1040 calls = listToCallDetails free_calls},
1041 (bagToList dump_dbs, dump_calls)
1045 bndr_set = mkVarSet bndrs
1047 (free_dbs, dump_dbs, dump_idset)
1048 = foldlBag dump_db (emptyBag, emptyBag, bndr_set) orig_dbs
1049 -- Important that it's foldl not foldr;
1050 -- we're accumulating the set of dumped ids in dump_set
1052 -- Filter out any calls that mention things that are being dumped
1053 orig_call_list = callDetailsToList orig_calls
1054 (dump_calls, free_calls) = partition captured orig_call_list
1055 captured (id,tys,(dicts, fvs)) = fvs `intersectsVarSet` dump_idset
1056 || id `elemVarSet` dump_idset
1058 dump_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1059 | dump_idset `intersectsVarSet` fvs -- Dump it
1060 = (free_dbs, dump_dbs `snocBag` db,
1061 dump_idset `unionVarSet` mkVarSet (bindersOf bind))
1063 | otherwise -- Don't dump it
1064 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1067 Given a type and value substitution, specUDs creates a specialised copy of
1071 specUDs :: TyVarSubst -> IdSubst -> ProtoUsageDetails -> SpecM UsageDetails
1072 specUDs tv_env dict_env (dbs, calls)
1073 = getUniqSupplySM `thenSM` \ us ->
1075 ((us', dict_env'), dbs') = mapAccumL specDB (us, dict_env) dbs
1077 setUniqSupplySM us' `thenSM_`
1078 returnSM (MkUD { dict_binds = listToBag dbs',
1079 calls = foldr (unionCalls . singleCall . inst_call dict_env')
1083 inst_call dict_env (id, tys, (dicts,fvs)) = (id, map (inst_maybe_ty fvs) tys,
1084 map (substExpr tv_env dict_env fvs) dicts)
1086 inst_maybe_ty fvs Nothing = Nothing
1087 inst_maybe_ty fvs (Just ty) = Just (fullSubstTy tv_env fvs ty)
1089 specDB (us, dict_env) (NonRec bndr rhs, fvs)
1090 = ((us', dict_env'), mkDB (NonRec bndr' (substExpr tv_env dict_env fvs rhs)))
1092 (dict_env', _, us', bndr') = substId clone_fn tv_env dict_env fvs us bndr
1093 -- Fudge the in_scope set a bit by using the free vars of
1094 -- the binding, and ignoring the one that comes back
1096 specDB (us, dict_env) (Rec prs, fvs)
1097 = ((us', dict_env'), mkDB (Rec (bndrs' `zip` rhss')))
1099 (dict_env', _, us', bndrs') = substIds clone_fn tv_env dict_env fvs us (map fst prs)
1100 rhss' = [substExpr tv_env dict_env' fvs rhs | (_, rhs) <- prs]
1102 clone_fn _ us id = case splitUniqSupply us of
1103 (us1, us2) -> Just (us1, setVarUnique id (uniqFromSupply us2))
1106 %************************************************************************
1108 \subsubsection{Boring helper functions}
1110 %************************************************************************
1113 lookupId:: IdEnv Id -> Id -> Id
1114 lookupId env id = case lookupVarEnv env id of
1118 addIdSpecialisations id spec_stuff
1119 = (if not (null errs) then
1120 pprTrace "Duplicate specialisations" (vcat (map ppr errs))
1123 setIdSpecialisation id new_spec_env
1125 (new_spec_env, errs) = foldr add (getIdSpecialisation id, []) spec_stuff
1127 add (tyvars, tys, template) (spec_env, errs)
1128 = case addToSpecEnv True spec_env tyvars tys template of
1129 Succeeded spec_env' -> (spec_env', errs)
1130 Failed err -> (spec_env, err:errs)
1132 ----------------------------------------
1133 type SpecM a = UniqSM a
1138 getUniqSM = getUniqueUs
1139 getUniqSupplySM = getUs
1140 setUniqSupplySM = setUs
1144 mapAndCombineSM f [] = returnSM ([], emptyUDs)
1145 mapAndCombineSM f (x:xs) = f x `thenSM` \ (y, uds1) ->
1146 mapAndCombineSM f xs `thenSM` \ (ys, uds2) ->
1147 returnSM (y:ys, uds1 `plusUDs` uds2)
1149 newIdSM old_id new_ty
1150 = getUniqSM `thenSM` \ uniq ->
1152 -- Give the new Id a similar occurrence name to the old one
1153 new_id = mkUserLocal (mkSpecOcc (nameOccName name)) uniq new_ty (getSrcLoc name)
1154 name = idName old_id
1159 = getUniqSM `thenSM` \ uniq ->
1160 returnSM (mkSysTyVar uniq boxedTypeKind)
1164 Old (but interesting) stuff about unboxed bindings
1165 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1167 What should we do when a value is specialised to a *strict* unboxed value?
1169 map_*_* f (x:xs) = let h = f x
1173 Could convert let to case:
1175 map_*_Int# f (x:xs) = case f x of h# ->
1179 This may be undesirable since it forces evaluation here, but the value
1180 may not be used in all branches of the body. In the general case this
1181 transformation is impossible since the mutual recursion in a letrec
1182 cannot be expressed as a case.
1184 There is also a problem with top-level unboxed values, since our
1185 implementation cannot handle unboxed values at the top level.
1187 Solution: Lift the binding of the unboxed value and extract it when it
1190 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1195 Now give it to the simplifier and the _Lifting will be optimised away.
1197 The benfit is that we have given the specialised "unboxed" values a
1198 very simplep lifted semantics and then leave it up to the simplifier to
1199 optimise it --- knowing that the overheads will be removed in nearly
1202 In particular, the value will only be evaluted in the branches of the
1203 program which use it, rather than being forced at the point where the
1204 value is bound. For example:
1206 filtermap_*_* p f (x:xs)
1213 filtermap_*_Int# p f (x:xs)
1214 = let h = case (f x) of h# -> _Lift h#
1217 True -> case h of _Lift h#
1221 The binding for h can still be inlined in the one branch and the
1222 _Lifting eliminated.
1225 Question: When won't the _Lifting be eliminated?
1227 Answer: When they at the top-level (where it is necessary) or when
1228 inlining would duplicate work (or possibly code depending on
1229 options). However, the _Lifting will still be eliminated if the
1230 strictness analyser deems the lifted binding strict.