2 % (c) The GRASP/AQUA Project, Glasgow University, 1993-1998
4 \section[Specialise]{Stamping out overloading, and (optionally) polymorphism}
8 -- The above warning supression flag is a temporary kludge.
9 -- While working on this module you are encouraged to remove it and fix
10 -- any warnings in the module. See
11 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
14 module Specialise ( specProgram ) where
16 #include "HsVersions.h"
18 import DynFlags ( DynFlags, DynFlag(..) )
19 import Id ( Id, idName, idType, mkUserLocal,
20 idInlinePragma, setInlinePragma )
21 import TcType ( Type, mkTyVarTy, tcSplitSigmaTy,
22 tyVarsOfTypes, tyVarsOfTheta, isClassPred,
23 tcCmpType, isUnLiftedType
25 import CoreSubst ( Subst, mkEmptySubst, extendTvSubstList, lookupIdSubst,
26 substBndr, substBndrs, substTy, substInScope,
27 cloneIdBndr, cloneIdBndrs, cloneRecIdBndrs
32 import CoreUtils ( applyTypeToArgs, mkPiTypes )
33 import CoreFVs ( exprFreeVars, exprsFreeVars, idFreeVars )
34 import CoreTidy ( tidyRules )
35 import CoreLint ( showPass, endPass )
36 import Rules ( addIdSpecialisations, mkLocalRule, lookupRule, emptyRuleBase, rulesOfBinds )
37 import PprCore ( pprRules )
38 import UniqSupply ( UniqSupply,
39 UniqSM, initUs_, thenUs, returnUs, getUniqueUs,
43 import MkId ( voidArgId, realWorldPrimId )
45 import Maybes ( catMaybes, maybeToBool )
46 import ErrUtils ( dumpIfSet_dyn )
47 import BasicTypes ( Activation( AlwaysActive ) )
49 import List ( partition )
50 import Util ( zipEqual, zipWithEqual, cmpList, lengthIs,
51 equalLength, lengthAtLeast, notNull )
58 %************************************************************************
60 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
62 %************************************************************************
64 These notes describe how we implement specialisation to eliminate
67 The specialisation pass works on Core
68 syntax, complete with all the explicit dictionary application,
69 abstraction and construction as added by the type checker. The
70 existing type checker remains largely as it is.
72 One important thought: the {\em types} passed to an overloaded
73 function, and the {\em dictionaries} passed are mutually redundant.
74 If the same function is applied to the same type(s) then it is sure to
75 be applied to the same dictionary(s)---or rather to the same {\em
76 values}. (The arguments might look different but they will evaluate
79 Second important thought: we know that we can make progress by
80 treating dictionary arguments as static and worth specialising on. So
81 we can do without binding-time analysis, and instead specialise on
82 dictionary arguments and no others.
91 and suppose f is overloaded.
93 STEP 1: CALL-INSTANCE COLLECTION
95 We traverse <body>, accumulating all applications of f to types and
98 (Might there be partial applications, to just some of its types and
99 dictionaries? In principle yes, but in practice the type checker only
100 builds applications of f to all its types and dictionaries, so partial
101 applications could only arise as a result of transformation, and even
102 then I think it's unlikely. In any case, we simply don't accumulate such
103 partial applications.)
108 So now we have a collection of calls to f:
112 Notice that f may take several type arguments. To avoid ambiguity, we
113 say that f is called at type t1/t2 and t3/t4.
115 We take equivalence classes using equality of the *types* (ignoring
116 the dictionary args, which as mentioned previously are redundant).
118 STEP 3: SPECIALISATION
120 For each equivalence class, choose a representative (f t1 t2 d1 d2),
121 and create a local instance of f, defined thus:
123 f@t1/t2 = <f_rhs> t1 t2 d1 d2
125 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
126 of simplification will now result. However we don't actually *do* that
127 simplification. Rather, we leave it for the simplifier to do. If we
128 *did* do it, though, we'd get more call instances from the specialised
129 RHS. We can work out what they are by instantiating the call-instance
130 set from f's RHS with the types t1, t2.
132 Add this new id to f's IdInfo, to record that f has a specialised version.
134 Before doing any of this, check that f's IdInfo doesn't already
135 tell us about an existing instance of f at the required type/s.
136 (This might happen if specialisation was applied more than once, or
137 it might arise from user SPECIALIZE pragmas.)
141 Wait a minute! What if f is recursive? Then we can't just plug in
142 its right-hand side, can we?
144 But it's ok. The type checker *always* creates non-recursive definitions
145 for overloaded recursive functions. For example:
147 f x = f (x+x) -- Yes I know its silly
151 f a (d::Num a) = let p = +.sel a d
153 letrec fl (y::a) = fl (p y y)
157 We still have recusion for non-overloaded functions which we
158 speciailise, but the recursive call should get specialised to the
159 same recursive version.
165 All this is crystal clear when the function is applied to *constant
166 types*; that is, types which have no type variables inside. But what if
167 it is applied to non-constant types? Suppose we find a call of f at type
168 t1/t2. There are two possibilities:
170 (a) The free type variables of t1, t2 are in scope at the definition point
171 of f. In this case there's no problem, we proceed just as before. A common
172 example is as follows. Here's the Haskell:
177 After typechecking we have
179 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
180 in +.sel a d (f a d y) (f a d y)
182 Notice that the call to f is at type type "a"; a non-constant type.
183 Both calls to f are at the same type, so we can specialise to give:
185 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
186 in +.sel a d (f@a y) (f@a y)
189 (b) The other case is when the type variables in the instance types
190 are *not* in scope at the definition point of f. The example we are
191 working with above is a good case. There are two instances of (+.sel a d),
192 but "a" is not in scope at the definition of +.sel. Can we do anything?
193 Yes, we can "common them up", a sort of limited common sub-expression deal.
196 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
197 f@a (x::a) = +.sel@a x x
198 in +.sel@a (f@a y) (f@a y)
200 This can save work, and can't be spotted by the type checker, because
201 the two instances of +.sel weren't originally at the same type.
205 * There are quite a few variations here. For example, the defn of
206 +.sel could be floated ouside the \y, to attempt to gain laziness.
207 It certainly mustn't be floated outside the \d because the d has to
210 * We don't want to inline f_rhs in this case, because
211 that will duplicate code. Just commoning up the call is the point.
213 * Nothing gets added to +.sel's IdInfo.
215 * Don't bother unless the equivalence class has more than one item!
217 Not clear whether this is all worth it. It is of course OK to
218 simply discard call-instances when passing a big lambda.
220 Polymorphism 2 -- Overloading
222 Consider a function whose most general type is
224 f :: forall a b. Ord a => [a] -> b -> b
226 There is really no point in making a version of g at Int/Int and another
227 at Int/Bool, because it's only instancing the type variable "a" which
228 buys us any efficiency. Since g is completely polymorphic in b there
229 ain't much point in making separate versions of g for the different
232 That suggests that we should identify which of g's type variables
233 are constrained (like "a") and which are unconstrained (like "b").
234 Then when taking equivalence classes in STEP 2, we ignore the type args
235 corresponding to unconstrained type variable. In STEP 3 we make
236 polymorphic versions. Thus:
238 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
247 f a (d::Num a) = let g = ...
249 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
251 Here, g is only called at one type, but the dictionary isn't in scope at the
252 definition point for g. Usually the type checker would build a
253 definition for d1 which enclosed g, but the transformation system
254 might have moved d1's defn inward. Solution: float dictionary bindings
255 outwards along with call instances.
259 f x = let g p q = p==q
265 Before specialisation, leaving out type abstractions we have
267 f df x = let g :: Eq a => a -> a -> Bool
269 h :: Num a => a -> a -> (a, Bool)
270 h dh r s = let deq = eqFromNum dh
271 in (+ dh r s, g deq r s)
275 After specialising h we get a specialised version of h, like this:
277 h' r s = let deq = eqFromNum df
278 in (+ df r s, g deq r s)
280 But we can't naively make an instance for g from this, because deq is not in scope
281 at the defn of g. Instead, we have to float out the (new) defn of deq
282 to widen its scope. Notice that this floating can't be done in advance -- it only
283 shows up when specialisation is done.
285 User SPECIALIZE pragmas
286 ~~~~~~~~~~~~~~~~~~~~~~~
287 Specialisation pragmas can be digested by the type checker, and implemented
288 by adding extra definitions along with that of f, in the same way as before
290 f@t1/t2 = <f_rhs> t1 t2 d1 d2
292 Indeed the pragmas *have* to be dealt with by the type checker, because
293 only it knows how to build the dictionaries d1 and d2! For example
295 g :: Ord a => [a] -> [a]
296 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
298 Here, the specialised version of g is an application of g's rhs to the
299 Ord dictionary for (Tree Int), which only the type checker can conjure
300 up. There might not even *be* one, if (Tree Int) is not an instance of
301 Ord! (All the other specialision has suitable dictionaries to hand
304 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
305 it is buried in a complex (as-yet-un-desugared) binding group.
308 f@t1/t2 = f* t1 t2 d1 d2
310 where f* is the Id f with an IdInfo which says "inline me regardless!".
311 Indeed all the specialisation could be done in this way.
312 That in turn means that the simplifier has to be prepared to inline absolutely
313 any in-scope let-bound thing.
316 Again, the pragma should permit polymorphism in unconstrained variables:
318 h :: Ord a => [a] -> b -> b
319 {-# SPECIALIZE h :: [Int] -> b -> b #-}
321 We *insist* that all overloaded type variables are specialised to ground types,
322 (and hence there can be no context inside a SPECIALIZE pragma).
323 We *permit* unconstrained type variables to be specialised to
325 - or left as a polymorphic type variable
326 but nothing in between. So
328 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
330 is *illegal*. (It can be handled, but it adds complication, and gains the
334 SPECIALISING INSTANCE DECLARATIONS
335 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
338 instance Foo a => Foo [a] where
340 {-# SPECIALIZE instance Foo [Int] #-}
342 The original instance decl creates a dictionary-function
345 dfun.Foo.List :: forall a. Foo a -> Foo [a]
347 The SPECIALIZE pragma just makes a specialised copy, just as for
348 ordinary function definitions:
350 dfun.Foo.List@Int :: Foo [Int]
351 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
353 The information about what instance of the dfun exist gets added to
354 the dfun's IdInfo in the same way as a user-defined function too.
357 Automatic instance decl specialisation?
358 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
359 Can instance decls be specialised automatically? It's tricky.
360 We could collect call-instance information for each dfun, but
361 then when we specialised their bodies we'd get new call-instances
362 for ordinary functions; and when we specialised their bodies, we might get
363 new call-instances of the dfuns, and so on. This all arises because of
364 the unrestricted mutual recursion between instance decls and value decls.
366 Still, there's no actual problem; it just means that we may not do all
367 the specialisation we could theoretically do.
369 Furthermore, instance decls are usually exported and used non-locally,
370 so we'll want to compile enough to get those specialisations done.
372 Lastly, there's no such thing as a local instance decl, so we can
373 survive solely by spitting out *usage* information, and then reading that
374 back in as a pragma when next compiling the file. So for now,
375 we only specialise instance decls in response to pragmas.
378 SPITTING OUT USAGE INFORMATION
379 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
381 To spit out usage information we need to traverse the code collecting
382 call-instance information for all imported (non-prelude?) functions
383 and data types. Then we equivalence-class it and spit it out.
385 This is done at the top-level when all the call instances which escape
386 must be for imported functions and data types.
388 *** Not currently done ***
391 Partial specialisation by pragmas
392 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
393 What about partial specialisation:
395 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
396 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
400 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
402 Seems quite reasonable. Similar things could be done with instance decls:
404 instance (Foo a, Foo b) => Foo (a,b) where
406 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
407 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
409 Ho hum. Things are complex enough without this. I pass.
412 Requirements for the simplifer
413 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
414 The simplifier has to be able to take advantage of the specialisation.
416 * When the simplifier finds an application of a polymorphic f, it looks in
417 f's IdInfo in case there is a suitable instance to call instead. This converts
419 f t1 t2 d1 d2 ===> f_t1_t2
421 Note that the dictionaries get eaten up too!
423 * Dictionary selection operations on constant dictionaries must be
426 +.sel Int d ===> +Int
428 The obvious way to do this is in the same way as other specialised
429 calls: +.sel has inside it some IdInfo which tells that if it's applied
430 to the type Int then it should eat a dictionary and transform to +Int.
432 In short, dictionary selectors need IdInfo inside them for constant
435 * Exactly the same applies if a superclass dictionary is being
438 Eq.sel Int d ===> dEqInt
440 * Something similar applies to dictionary construction too. Suppose
441 dfun.Eq.List is the function taking a dictionary for (Eq a) to
442 one for (Eq [a]). Then we want
444 dfun.Eq.List Int d ===> dEq.List_Int
446 Where does the Eq [Int] dictionary come from? It is built in
447 response to a SPECIALIZE pragma on the Eq [a] instance decl.
449 In short, dfun Ids need IdInfo with a specialisation for each
450 constant instance of their instance declaration.
452 All this uses a single mechanism: the SpecEnv inside an Id
455 What does the specialisation IdInfo look like?
456 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
458 The SpecEnv of an Id maps a list of types (the template) to an expression
462 For example, if f has this SpecInfo:
464 [Int, a] -> \d:Ord Int. f' a
466 it means that we can replace the call
468 f Int t ===> (\d. f' t)
470 This chucks one dictionary away and proceeds with the
471 specialised version of f, namely f'.
474 What can't be done this way?
475 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
476 There is no way, post-typechecker, to get a dictionary for (say)
477 Eq a from a dictionary for Eq [a]. So if we find
481 we can't transform to
486 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
488 Of course, we currently have no way to automatically derive
489 eqList, nor to connect it to the Eq [a] instance decl, but you
490 can imagine that it might somehow be possible. Taking advantage
491 of this is permanently ruled out.
493 Still, this is no great hardship, because we intend to eliminate
494 overloading altogether anyway!
498 A note about non-tyvar dictionaries
499 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
500 Some Ids have types like
502 forall a,b,c. Eq a -> Ord [a] -> tau
504 This seems curious at first, because we usually only have dictionary
505 args whose types are of the form (C a) where a is a type variable.
506 But this doesn't hold for the functions arising from instance decls,
507 which sometimes get arguements with types of form (C (T a)) for some
510 Should we specialise wrt this compound-type dictionary? We used to say
512 "This is a heuristic judgement, as indeed is the fact that we
513 specialise wrt only dictionaries. We choose *not* to specialise
514 wrt compound dictionaries because at the moment the only place
515 they show up is in instance decls, where they are simply plugged
516 into a returned dictionary. So nothing is gained by specialising
519 But it is simpler and more uniform to specialise wrt these dicts too;
520 and in future GHC is likely to support full fledged type signatures
522 f ;: Eq [(a,b)] => ...
525 %************************************************************************
527 \subsubsection{The new specialiser}
529 %************************************************************************
531 Our basic game plan is this. For let(rec) bound function
532 f :: (C a, D c) => (a,b,c,d) -> Bool
534 * Find any specialised calls of f, (f ts ds), where
535 ts are the type arguments t1 .. t4, and
536 ds are the dictionary arguments d1 .. d2.
538 * Add a new definition for f1 (say):
540 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
542 Note that we abstract over the unconstrained type arguments.
546 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
548 to the specialisations of f. This will be used by the
549 simplifier to replace calls
550 (f t1 t2 t3 t4) da db
552 (\d1 d1 -> f1 t2 t4) da db
554 All the stuff about how many dictionaries to discard, and what types
555 to apply the specialised function to, are handled by the fact that the
556 SpecEnv contains a template for the result of the specialisation.
558 We don't build *partial* specialisations for f. For example:
560 f :: Eq a => a -> a -> Bool
561 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
563 Here, little is gained by making a specialised copy of f.
564 There's a distinct danger that the specialised version would
565 first build a dictionary for (Eq b, Eq c), and then select the (==)
566 method from it! Even if it didn't, not a great deal is saved.
568 We do, however, generate polymorphic, but not overloaded, specialisations:
570 f :: Eq a => [a] -> b -> b -> b
571 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
573 Hence, the invariant is this:
575 *** no specialised version is overloaded ***
578 %************************************************************************
580 \subsubsection{The exported function}
582 %************************************************************************
585 specProgram :: DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind]
586 specProgram dflags us binds
588 showPass dflags "Specialise"
590 let binds' = initSM us (go binds `thenSM` \ (binds', uds') ->
591 returnSM (dumpAllDictBinds uds' binds'))
593 endPass dflags "Specialise" Opt_D_dump_spec binds'
595 dumpIfSet_dyn dflags Opt_D_dump_rules "Top-level specialisations"
596 (pprRules (tidyRules emptyTidyEnv (rulesOfBinds binds')))
600 -- We need to start with a Subst that knows all the things
601 -- that are in scope, so that the substitution engine doesn't
602 -- accidentally re-use a unique that's already in use
603 -- Easiest thing is to do it all at once, as if all the top-level
604 -- decls were mutually recursive
605 top_subst = mkEmptySubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
607 go [] = returnSM ([], emptyUDs)
608 go (bind:binds) = go binds `thenSM` \ (binds', uds) ->
609 specBind top_subst bind uds `thenSM` \ (bind', uds') ->
610 returnSM (bind' ++ binds', uds')
613 %************************************************************************
615 \subsubsection{@specExpr@: the main function}
617 %************************************************************************
620 specVar :: Subst -> Id -> CoreExpr
621 specVar subst v = lookupIdSubst subst v
623 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
624 -- We carry a substitution down:
625 -- a) we must clone any binding that might flaot outwards,
626 -- to avoid name clashes
627 -- b) we carry a type substitution to use when analysing
628 -- the RHS of specialised bindings (no type-let!)
630 ---------------- First the easy cases --------------------
631 specExpr subst (Type ty) = returnSM (Type (substTy subst ty), emptyUDs)
632 specExpr subst (Var v) = returnSM (specVar subst v, emptyUDs)
633 specExpr subst (Lit lit) = returnSM (Lit lit, emptyUDs)
634 specExpr subst (Cast e co) =
635 specExpr subst e `thenSM` \ (e', uds) ->
636 returnSM ((Cast e' (substTy subst co)), uds)
637 specExpr subst (Note note body)
638 = specExpr subst body `thenSM` \ (body', uds) ->
639 returnSM (Note (specNote subst note) body', uds)
642 ---------------- Applications might generate a call instance --------------------
643 specExpr subst expr@(App fun arg)
646 go (App fun arg) args = specExpr subst arg `thenSM` \ (arg', uds_arg) ->
647 go fun (arg':args) `thenSM` \ (fun', uds_app) ->
648 returnSM (App fun' arg', uds_arg `plusUDs` uds_app)
650 go (Var f) args = case specVar subst f of
651 Var f' -> returnSM (Var f', mkCallUDs subst f' args)
652 e' -> returnSM (e', emptyUDs) -- I don't expect this!
653 go other args = specExpr subst other
655 ---------------- Lambda/case require dumping of usage details --------------------
656 specExpr subst e@(Lam _ _)
657 = specExpr subst' body `thenSM` \ (body', uds) ->
659 (filtered_uds, body'') = dumpUDs bndrs' uds body'
661 returnSM (mkLams bndrs' body'', filtered_uds)
663 (bndrs, body) = collectBinders e
664 (subst', bndrs') = substBndrs subst bndrs
665 -- More efficient to collect a group of binders together all at once
666 -- and we don't want to split a lambda group with dumped bindings
668 specExpr subst (Case scrut case_bndr ty alts)
669 = specExpr subst scrut `thenSM` \ (scrut', uds_scrut) ->
670 mapAndCombineSM spec_alt alts `thenSM` \ (alts', uds_alts) ->
671 returnSM (Case scrut' case_bndr' (substTy subst ty) alts', uds_scrut `plusUDs` uds_alts)
673 (subst_alt, case_bndr') = substBndr subst case_bndr
674 -- No need to clone case binder; it can't float like a let(rec)
676 spec_alt (con, args, rhs)
677 = specExpr subst_rhs rhs `thenSM` \ (rhs', uds) ->
679 (uds', rhs'') = dumpUDs args uds rhs'
681 returnSM ((con, args', rhs''), uds')
683 (subst_rhs, args') = substBndrs subst_alt args
685 ---------------- Finally, let is the interesting case --------------------
686 specExpr subst (Let bind body)
688 cloneBindSM subst bind `thenSM` \ (rhs_subst, body_subst, bind') ->
690 -- Deal with the body
691 specExpr body_subst body `thenSM` \ (body', body_uds) ->
693 -- Deal with the bindings
694 specBind rhs_subst bind' body_uds `thenSM` \ (binds', uds) ->
697 returnSM (foldr Let body' binds', uds)
699 -- Must apply the type substitution to coerceions
700 specNote subst note = note
703 %************************************************************************
705 \subsubsection{Dealing with a binding}
707 %************************************************************************
710 specBind :: Subst -- Use this for RHSs
712 -> UsageDetails -- Info on how the scope of the binding
713 -> SpecM ([CoreBind], -- New bindings
714 UsageDetails) -- And info to pass upstream
716 specBind rhs_subst bind body_uds
717 = specBindItself rhs_subst bind (calls body_uds) `thenSM` \ (bind', bind_uds) ->
719 bndrs = bindersOf bind
720 all_uds = zapCalls bndrs (body_uds `plusUDs` bind_uds)
721 -- It's important that the `plusUDs` is this way round,
722 -- because body_uds may bind dictionaries that are
723 -- used in the calls passed to specDefn. So the
724 -- dictionary bindings in bind_uds may mention
725 -- dictionaries bound in body_uds.
727 case splitUDs bndrs all_uds of
729 (_, ([],[])) -- This binding doesn't bind anything needed
730 -- in the UDs, so put the binding here
731 -- This is the case for most non-dict bindings, except
732 -- for the few that are mentioned in a dict binding
733 -- that is floating upwards in body_uds
734 -> returnSM ([bind'], all_uds)
736 (float_uds, (dict_binds, calls)) -- This binding is needed in the UDs, so float it out
737 -> returnSM ([], float_uds `plusUDs` mkBigUD bind' dict_binds calls)
740 -- A truly gruesome function
741 mkBigUD bind@(NonRec _ _) dbs calls
742 = -- Common case: non-recursive and no specialisations
743 -- (if there were any specialistions it would have been made recursive)
744 MkUD { dict_binds = listToBag (mkDB bind : dbs),
745 calls = listToCallDetails calls }
747 mkBigUD bind dbs calls
749 MkUD { dict_binds = unitBag (mkDB (Rec (bind_prs bind ++ dbsToPairs dbs))),
751 calls = listToCallDetails calls }
753 bind_prs (NonRec b r) = [(b,r)]
754 bind_prs (Rec prs) = prs
757 dbsToPairs ((bind,_):dbs) = bind_prs bind ++ dbsToPairs dbs
759 -- specBindItself deals with the RHS, specialising it according
760 -- to the calls found in the body (if any)
761 specBindItself rhs_subst (NonRec bndr rhs) call_info
762 = specDefn rhs_subst call_info (bndr,rhs) `thenSM` \ ((bndr',rhs'), spec_defns, spec_uds) ->
764 new_bind | null spec_defns = NonRec bndr' rhs'
765 | otherwise = Rec ((bndr',rhs'):spec_defns)
766 -- bndr' mentions the spec_defns in its SpecEnv
767 -- Not sure why we couln't just put the spec_defns first
769 returnSM (new_bind, spec_uds)
771 specBindItself rhs_subst (Rec pairs) call_info
772 = mapSM (specDefn rhs_subst call_info) pairs `thenSM` \ stuff ->
774 (pairs', spec_defns_s, spec_uds_s) = unzip3 stuff
775 spec_defns = concat spec_defns_s
776 spec_uds = plusUDList spec_uds_s
777 new_bind = Rec (spec_defns ++ pairs')
779 returnSM (new_bind, spec_uds)
782 specDefn :: Subst -- Subst to use for RHS
783 -> CallDetails -- Info on how it is used in its scope
784 -> (Id, CoreExpr) -- The thing being bound and its un-processed RHS
785 -> SpecM ((Id, CoreExpr), -- The thing and its processed RHS
786 -- the Id may now have specialisations attached
787 [(Id,CoreExpr)], -- Extra, specialised bindings
788 UsageDetails -- Stuff to fling upwards from the RHS and its
789 ) -- specialised versions
791 specDefn subst calls (fn, rhs)
792 -- The first case is the interesting one
793 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
794 && rhs_ids `lengthAtLeast` n_dicts -- and enough dict args
795 && notNull calls_for_me -- And there are some calls to specialise
797 -- && not (certainlyWillInline (idUnfolding fn)) -- And it's not small
798 -- See Note [Inline specialisation] for why we do not
799 -- switch off specialisation for inline functions
801 = -- Specialise the body of the function
802 specExpr subst rhs `thenSM` \ (rhs', rhs_uds) ->
804 -- Make a specialised version for each call in calls_for_me
805 mapSM spec_call calls_for_me `thenSM` \ stuff ->
807 (spec_defns, spec_uds, spec_rules) = unzip3 stuff
809 fn' = addIdSpecialisations fn spec_rules
811 returnSM ((fn',rhs'),
813 rhs_uds `plusUDs` plusUDList spec_uds)
815 | otherwise -- No calls or RHS doesn't fit our preconceptions
816 = WARN( notNull calls_for_me, ptext SLIT("Missed specialisation opportunity for") <+> ppr fn )
817 -- Note [Specialisation shape]
818 specExpr subst rhs `thenSM` \ (rhs', rhs_uds) ->
819 returnSM ((fn, rhs'), [], rhs_uds)
823 (tyvars, theta, _) = tcSplitSigmaTy fn_type
824 n_tyvars = length tyvars
825 n_dicts = length theta
826 inline_prag = idInlinePragma fn
828 -- It's important that we "see past" any INLINE pragma
829 -- else we'll fail to specialise an INLINE thing
830 (inline_rhs, rhs_inside) = dropInline rhs
831 (rhs_tyvars, rhs_ids, rhs_body) = collectTyAndValBinders rhs_inside
833 rhs_dicts = take n_dicts rhs_ids
834 rhs_bndrs = rhs_tyvars ++ rhs_dicts
835 body = mkLams (drop n_dicts rhs_ids) rhs_body
836 -- Glue back on the non-dict lambdas
838 calls_for_me = case lookupFM calls fn of
840 Just cs -> fmToList cs
842 ----------------------------------------------------------
843 -- Specialise to one particular call pattern
844 spec_call :: (CallKey, ([DictExpr], VarSet)) -- Call instance
845 -> SpecM ((Id,CoreExpr), -- Specialised definition
846 UsageDetails, -- Usage details from specialised body
847 CoreRule) -- Info for the Id's SpecEnv
848 spec_call (CallKey call_ts, (call_ds, call_fvs))
849 = ASSERT( call_ts `lengthIs` n_tyvars && call_ds `lengthIs` n_dicts )
850 -- Calls are only recorded for properly-saturated applications
852 -- Suppose f's defn is f = /\ a b c d -> \ d1 d2 -> rhs
853 -- Supppose the call is for f [Just t1, Nothing, Just t3, Nothing] [dx1, dx2]
855 -- Construct the new binding
856 -- f1 = SUBST[a->t1,c->t3, d1->d1', d2->d2'] (/\ b d -> rhs)
857 -- PLUS the usage-details
858 -- { d1' = dx1; d2' = dx2 }
859 -- where d1', d2' are cloned versions of d1,d2, with the type substitution applied.
861 -- Note that the substitution is applied to the whole thing.
862 -- This is convenient, but just slightly fragile. Notably:
863 -- * There had better be no name clashes in a/b/c/d
866 -- poly_tyvars = [b,d] in the example above
867 -- spec_tyvars = [a,c]
868 -- ty_args = [t1,b,t3,d]
869 poly_tyvars = [tv | (tv, Nothing) <- rhs_tyvars `zip` call_ts]
870 spec_tyvars = [tv | (tv, Just _) <- rhs_tyvars `zip` call_ts]
871 ty_args = zipWithEqual "spec_call" mk_ty_arg rhs_tyvars call_ts
873 mk_ty_arg rhs_tyvar Nothing = Type (mkTyVarTy rhs_tyvar)
874 mk_ty_arg rhs_tyvar (Just ty) = Type ty
875 rhs_subst = extendTvSubstList subst (spec_tyvars `zip` [ty | Just ty <- call_ts])
877 cloneBinders rhs_subst rhs_dicts `thenSM` \ (rhs_subst', rhs_dicts') ->
879 inst_args = ty_args ++ map Var rhs_dicts'
881 -- Figure out the type of the specialised function
882 body_ty = applyTypeToArgs rhs fn_type inst_args
883 (lam_args, app_args) -- Add a dummy argument if body_ty is unlifted
884 | isUnLiftedType body_ty -- C.f. WwLib.mkWorkerArgs
885 = (poly_tyvars ++ [voidArgId], poly_tyvars ++ [realWorldPrimId])
886 | otherwise = (poly_tyvars, poly_tyvars)
887 spec_id_ty = mkPiTypes lam_args body_ty
889 newIdSM fn spec_id_ty `thenSM` \ spec_f ->
890 specExpr rhs_subst' (mkLams lam_args body) `thenSM` \ (spec_rhs, rhs_uds) ->
892 -- The rule to put in the function's specialisation is:
893 -- forall b,d, d1',d2'. f t1 b t3 d d1' d2' = f1 b d
894 spec_env_rule = mkLocalRule (mkFastString ("SPEC " ++ showSDoc (ppr fn)))
895 AlwaysActive (idName fn)
896 (poly_tyvars ++ rhs_dicts')
898 (mkVarApps (Var spec_f) app_args)
900 -- Add the { d1' = dx1; d2' = dx2 } usage stuff
901 final_uds = foldr addDictBind rhs_uds (my_zipEqual "spec_call" rhs_dicts' call_ds)
903 spec_pr | inline_rhs = (spec_f `setInlinePragma` inline_prag, Note InlineMe spec_rhs)
904 | otherwise = (spec_f, spec_rhs)
906 returnSM (spec_pr, final_uds, spec_env_rule)
909 my_zipEqual doc xs ys
911 | not (equalLength xs ys) = pprPanic "my_zipEqual" (vcat
913 , ppr fn <+> ppr call_ts
914 , ppr (idType fn), ppr theta
915 , ppr n_dicts, ppr rhs_dicts
918 | otherwise = zipEqual doc xs ys
921 Note [Specialisation shape]
922 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
923 We only specialise a function if it has visible top-level lambdas
924 corresponding to its overloading. E.g. if
925 f :: forall a. Eq a => ....
926 then its body must look like
929 Reason: when specialising the body for a call (f ty dexp), we want to
930 substitute dexp for d, and pick up specialised calls in the body of f.
932 This doesn't always work. One example I came across was htis:
933 newtype Gen a = MkGen{ unGen :: Int -> a }
935 choose :: Eq a => a -> Gen a
936 choose n = MkGen (\r -> n)
938 oneof = choose (1::Int)
940 It's a silly exapmle, but we get
941 choose = /\a. g `cast` co
942 where choose doesn't have any dict arguments. Thus far I have not
943 tried to fix this (wait till there's a real example).
946 Note [Inline specialisations]
947 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
948 We transfer to the specialised function any INLINE stuff from the
949 original. This means (a) the Activation in the IdInfo, and (b) any
952 This is a change (Jun06). Previously the idea is that the point of
953 inlining was precisely to specialise the function at its call site,
954 and that's not so important for the specialised copies. But
955 *pragma-directed* specialisation now takes place in the
956 typechecker/desugarer, with manually specified INLINEs. The
957 specialiation here is automatic. It'd be very odd if a function
958 marked INLINE was specialised (because of some local use), and then
959 forever after (including importing modules) the specialised version
960 wasn't INLINEd. After all, the programmer said INLINE!
962 You might wonder why we don't just not specialise INLINE functions.
963 It's because even INLINE functions are sometimes not inlined, when
964 they aren't applied to interesting arguments. But perhaps the type
965 arguments alone are enough to specialise (even though the args are too
966 boring to trigger inlining), and it's certainly better to call the
969 A case in point is dictionary functions, which are current marked
970 INLINE, but which are worth specialising.
973 dropInline :: CoreExpr -> (Bool, CoreExpr)
974 dropInline (Note InlineMe rhs) = (True, rhs)
975 dropInline rhs = (False, rhs)
978 %************************************************************************
980 \subsubsection{UsageDetails and suchlike}
982 %************************************************************************
987 dict_binds :: !(Bag DictBind),
988 -- Floated dictionary bindings
989 -- The order is important;
990 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
991 -- (Remember, Bags preserve order in GHC.)
993 calls :: !CallDetails
996 type DictBind = (CoreBind, VarSet)
997 -- The set is the free vars of the binding
998 -- both tyvars and dicts
1000 type DictExpr = CoreExpr
1002 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM }
1004 type ProtoUsageDetails = ([DictBind],
1005 [(Id, CallKey, ([DictExpr], VarSet))]
1008 ------------------------------------------------------------
1009 type CallDetails = FiniteMap Id CallInfo
1010 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
1011 type CallInfo = FiniteMap CallKey
1012 ([DictExpr], VarSet) -- Dict args and the vars of the whole
1013 -- call (including tyvars)
1014 -- [*not* include the main id itself, of course]
1015 -- The finite maps eliminate duplicates
1016 -- The list of types and dictionaries is guaranteed to
1017 -- match the type of f
1019 -- Type isn't an instance of Ord, so that we can control which
1020 -- instance we use. That's tiresome here. Oh well
1021 instance Eq CallKey where
1022 k1 == k2 = case k1 `compare` k2 of { EQ -> True; other -> False }
1024 instance Ord CallKey where
1025 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
1027 cmp Nothing Nothing = EQ
1028 cmp Nothing (Just t2) = LT
1029 cmp (Just t1) Nothing = GT
1030 cmp (Just t1) (Just t2) = tcCmpType t1 t2
1032 unionCalls :: CallDetails -> CallDetails -> CallDetails
1033 unionCalls c1 c2 = plusFM_C plusFM c1 c2
1035 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> CallDetails
1036 singleCall id tys dicts
1037 = unitFM id (unitFM (CallKey tys) (dicts, call_fvs))
1039 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
1040 tys_fvs = tyVarsOfTypes (catMaybes tys)
1041 -- The type args (tys) are guaranteed to be part of the dictionary
1042 -- types, because they are just the constrained types,
1043 -- and the dictionary is therefore sure to be bound
1044 -- inside the binding for any type variables free in the type;
1045 -- hence it's safe to neglect tyvars free in tys when making
1046 -- the free-var set for this call
1047 -- BUT I don't trust this reasoning; play safe and include tys_fvs
1049 -- We don't include the 'id' itself.
1051 listToCallDetails calls
1052 = foldr (unionCalls . mk_call) emptyFM calls
1054 mk_call (id, tys, dicts_w_fvs) = unitFM id (unitFM tys dicts_w_fvs)
1055 -- NB: the free vars of the call are provided
1057 callDetailsToList calls = [ (id,tys,dicts)
1058 | (id,fm) <- fmToList calls,
1059 (tys, dicts) <- fmToList fm
1062 mkCallUDs subst f args
1064 || not (all isClassPred theta)
1065 -- Only specialise if all overloading is on class params.
1066 -- In ptic, with implicit params, the type args
1067 -- *don't* say what the value of the implicit param is!
1068 || not (spec_tys `lengthIs` n_tyvars)
1069 || not ( dicts `lengthIs` n_dicts)
1070 || maybeToBool (lookupRule (\act -> True) (substInScope subst) emptyRuleBase f args)
1071 -- There's already a rule covering this call. A typical case
1072 -- is where there's an explicit user-provided rule. Then
1073 -- we don't want to create a specialised version
1074 -- of the function that overlaps.
1075 = emptyUDs -- Not overloaded, or no specialisation wanted
1078 = MkUD {dict_binds = emptyBag,
1079 calls = singleCall f spec_tys dicts
1082 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1083 constrained_tyvars = tyVarsOfTheta theta
1084 n_tyvars = length tyvars
1085 n_dicts = length theta
1087 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1088 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1091 | tyvar `elemVarSet` constrained_tyvars = Just ty
1092 | otherwise = Nothing
1094 ------------------------------------------------------------
1095 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1096 plusUDs (MkUD {dict_binds = db1, calls = calls1})
1097 (MkUD {dict_binds = db2, calls = calls2})
1098 = MkUD {dict_binds = d, calls = c}
1100 d = db1 `unionBags` db2
1101 c = calls1 `unionCalls` calls2
1103 plusUDList = foldr plusUDs emptyUDs
1105 -- zapCalls deletes calls to ids from uds
1106 zapCalls ids uds = uds {calls = delListFromFM (calls uds) ids}
1108 mkDB bind = (bind, bind_fvs bind)
1110 bind_fvs (NonRec bndr rhs) = pair_fvs (bndr,rhs)
1111 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1114 rhs_fvs = unionVarSets (map pair_fvs prs)
1116 pair_fvs (bndr, rhs) = exprFreeVars rhs `unionVarSet` idFreeVars bndr
1117 -- Don't forget variables mentioned in the
1118 -- rules of the bndr. C.f. OccAnal.addRuleUsage
1119 -- Also tyvars mentioned in its type; they may not appear in the RHS
1123 addDictBind (dict,rhs) uds = uds { dict_binds = mkDB (NonRec dict rhs) `consBag` dict_binds uds }
1125 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
1126 = foldrBag add binds dbs
1128 add (bind,_) binds = bind : binds
1130 dumpUDs :: [CoreBndr]
1131 -> UsageDetails -> CoreExpr
1132 -> (UsageDetails, CoreExpr)
1133 dumpUDs bndrs uds body
1134 = (free_uds, foldr add_let body dict_binds)
1136 (free_uds, (dict_binds, _)) = splitUDs bndrs uds
1137 add_let (bind,_) body = Let bind body
1139 splitUDs :: [CoreBndr]
1141 -> (UsageDetails, -- These don't mention the binders
1142 ProtoUsageDetails) -- These do
1144 splitUDs bndrs uds@(MkUD {dict_binds = orig_dbs,
1145 calls = orig_calls})
1147 = if isEmptyBag dump_dbs && null dump_calls then
1148 -- Common case: binder doesn't affect floats
1152 -- Binders bind some of the fvs of the floats
1153 (MkUD {dict_binds = free_dbs,
1154 calls = listToCallDetails free_calls},
1155 (bagToList dump_dbs, dump_calls)
1159 bndr_set = mkVarSet bndrs
1161 (free_dbs, dump_dbs, dump_idset)
1162 = foldlBag dump_db (emptyBag, emptyBag, bndr_set) orig_dbs
1163 -- Important that it's foldl not foldr;
1164 -- we're accumulating the set of dumped ids in dump_set
1166 -- Filter out any calls that mention things that are being dumped
1167 orig_call_list = callDetailsToList orig_calls
1168 (dump_calls, free_calls) = partition captured orig_call_list
1169 captured (id,tys,(dicts, fvs)) = fvs `intersectsVarSet` dump_idset
1170 || id `elemVarSet` dump_idset
1172 dump_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1173 | dump_idset `intersectsVarSet` fvs -- Dump it
1174 = (free_dbs, dump_dbs `snocBag` db,
1175 extendVarSetList dump_idset (bindersOf bind))
1177 | otherwise -- Don't dump it
1178 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1182 %************************************************************************
1184 \subsubsection{Boring helper functions}
1186 %************************************************************************
1189 type SpecM a = UniqSM a
1193 getUniqSM = getUniqueUs
1197 mapAndCombineSM f [] = returnSM ([], emptyUDs)
1198 mapAndCombineSM f (x:xs) = f x `thenSM` \ (y, uds1) ->
1199 mapAndCombineSM f xs `thenSM` \ (ys, uds2) ->
1200 returnSM (y:ys, uds1 `plusUDs` uds2)
1202 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1203 -- Clone the binders of the bind; return new bind with the cloned binders
1204 -- Return the substitution to use for RHSs, and the one to use for the body
1205 cloneBindSM subst (NonRec bndr rhs)
1206 = getUs `thenUs` \ us ->
1208 (subst', bndr') = cloneIdBndr subst us bndr
1210 returnUs (subst, subst', NonRec bndr' rhs)
1212 cloneBindSM subst (Rec pairs)
1213 = getUs `thenUs` \ us ->
1215 (subst', bndrs') = cloneRecIdBndrs subst us (map fst pairs)
1217 returnUs (subst', subst', Rec (bndrs' `zip` map snd pairs))
1219 cloneBinders subst bndrs
1220 = getUs `thenUs` \ us ->
1221 returnUs (cloneIdBndrs subst us bndrs)
1223 newIdSM old_id new_ty
1224 = getUniqSM `thenSM` \ uniq ->
1226 -- Give the new Id a similar occurrence name to the old one
1227 name = idName old_id
1228 new_id = mkUserLocal (mkSpecOcc (nameOccName name)) uniq new_ty (getSrcSpan name)
1234 Old (but interesting) stuff about unboxed bindings
1235 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1237 What should we do when a value is specialised to a *strict* unboxed value?
1239 map_*_* f (x:xs) = let h = f x
1243 Could convert let to case:
1245 map_*_Int# f (x:xs) = case f x of h# ->
1249 This may be undesirable since it forces evaluation here, but the value
1250 may not be used in all branches of the body. In the general case this
1251 transformation is impossible since the mutual recursion in a letrec
1252 cannot be expressed as a case.
1254 There is also a problem with top-level unboxed values, since our
1255 implementation cannot handle unboxed values at the top level.
1257 Solution: Lift the binding of the unboxed value and extract it when it
1260 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1265 Now give it to the simplifier and the _Lifting will be optimised away.
1267 The benfit is that we have given the specialised "unboxed" values a
1268 very simplep lifted semantics and then leave it up to the simplifier to
1269 optimise it --- knowing that the overheads will be removed in nearly
1272 In particular, the value will only be evaluted in the branches of the
1273 program which use it, rather than being forced at the point where the
1274 value is bound. For example:
1276 filtermap_*_* p f (x:xs)
1283 filtermap_*_Int# p f (x:xs)
1284 = let h = case (f x) of h# -> _Lift h#
1287 True -> case h of _Lift h#
1291 The binding for h can still be inlined in the one branch and the
1292 _Lifting eliminated.
1295 Question: When won't the _Lifting be eliminated?
1297 Answer: When they at the top-level (where it is necessary) or when
1298 inlining would duplicate work (or possibly code depending on
1299 options). However, the _Lifting will still be eliminated if the
1300 strictness analyser deems the lifted binding strict.