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 ( DynFlags, DynFlag(..) )
12 import Id ( Id, idName, idType, mkUserLocal )
13 import TcType ( Type, mkTyVarTy, tcSplitSigmaTy,
14 tyVarsOfTypes, tyVarsOfTheta, isClassPred,
15 tcCmpType, isUnLiftedType
17 import Subst ( Subst, SubstResult(..), mkSubst, mkSubst, extendTvSubstList,
18 simplBndr, simplBndrs, substTy,
19 substAndCloneId, substAndCloneIds, substAndCloneRecIds,
22 import Var ( zapSpecPragmaId )
26 import CoreUtils ( applyTypeToArgs, mkPiTypes )
27 import CoreFVs ( exprFreeVars, exprsFreeVars )
28 import CoreTidy ( pprTidyIdRules )
29 import CoreLint ( showPass, endPass )
30 import Rules ( addIdSpecialisations, lookupRule )
32 import UniqSupply ( UniqSupply,
33 UniqSM, initUs_, thenUs, returnUs, getUniqueUs,
36 import Name ( nameOccName, mkSpecOcc, getSrcLoc )
37 import MkId ( voidArgId, realWorldPrimId )
39 import Maybes ( catMaybes, maybeToBool )
40 import ErrUtils ( dumpIfSet_dyn )
41 import BasicTypes ( Activation( AlwaysActive ) )
43 import List ( partition )
44 import Util ( zipEqual, zipWithEqual, cmpList, lengthIs,
45 equalLength, lengthAtLeast, notNull )
52 %************************************************************************
54 \subsection[notes-Specialise]{Implementation notes [SLPJ, Aug 18 1993]}
56 %************************************************************************
58 These notes describe how we implement specialisation to eliminate
61 The specialisation pass works on Core
62 syntax, complete with all the explicit dictionary application,
63 abstraction and construction as added by the type checker. The
64 existing type checker remains largely as it is.
66 One important thought: the {\em types} passed to an overloaded
67 function, and the {\em dictionaries} passed are mutually redundant.
68 If the same function is applied to the same type(s) then it is sure to
69 be applied to the same dictionary(s)---or rather to the same {\em
70 values}. (The arguments might look different but they will evaluate
73 Second important thought: we know that we can make progress by
74 treating dictionary arguments as static and worth specialising on. So
75 we can do without binding-time analysis, and instead specialise on
76 dictionary arguments and no others.
85 and suppose f is overloaded.
87 STEP 1: CALL-INSTANCE COLLECTION
89 We traverse <body>, accumulating all applications of f to types and
92 (Might there be partial applications, to just some of its types and
93 dictionaries? In principle yes, but in practice the type checker only
94 builds applications of f to all its types and dictionaries, so partial
95 applications could only arise as a result of transformation, and even
96 then I think it's unlikely. In any case, we simply don't accumulate such
97 partial applications.)
102 So now we have a collection of calls to f:
106 Notice that f may take several type arguments. To avoid ambiguity, we
107 say that f is called at type t1/t2 and t3/t4.
109 We take equivalence classes using equality of the *types* (ignoring
110 the dictionary args, which as mentioned previously are redundant).
112 STEP 3: SPECIALISATION
114 For each equivalence class, choose a representative (f t1 t2 d1 d2),
115 and create a local instance of f, defined thus:
117 f@t1/t2 = <f_rhs> t1 t2 d1 d2
119 f_rhs presumably has some big lambdas and dictionary lambdas, so lots
120 of simplification will now result. However we don't actually *do* that
121 simplification. Rather, we leave it for the simplifier to do. If we
122 *did* do it, though, we'd get more call instances from the specialised
123 RHS. We can work out what they are by instantiating the call-instance
124 set from f's RHS with the types t1, t2.
126 Add this new id to f's IdInfo, to record that f has a specialised version.
128 Before doing any of this, check that f's IdInfo doesn't already
129 tell us about an existing instance of f at the required type/s.
130 (This might happen if specialisation was applied more than once, or
131 it might arise from user SPECIALIZE pragmas.)
135 Wait a minute! What if f is recursive? Then we can't just plug in
136 its right-hand side, can we?
138 But it's ok. The type checker *always* creates non-recursive definitions
139 for overloaded recursive functions. For example:
141 f x = f (x+x) -- Yes I know its silly
145 f a (d::Num a) = let p = +.sel a d
147 letrec fl (y::a) = fl (p y y)
151 We still have recusion for non-overloaded functions which we
152 speciailise, but the recursive call should get specialised to the
153 same recursive version.
159 All this is crystal clear when the function is applied to *constant
160 types*; that is, types which have no type variables inside. But what if
161 it is applied to non-constant types? Suppose we find a call of f at type
162 t1/t2. There are two possibilities:
164 (a) The free type variables of t1, t2 are in scope at the definition point
165 of f. In this case there's no problem, we proceed just as before. A common
166 example is as follows. Here's the Haskell:
171 After typechecking we have
173 g a (d::Num a) (y::a) = let f b (d'::Num b) (x::b) = +.sel b d' x x
174 in +.sel a d (f a d y) (f a d y)
176 Notice that the call to f is at type type "a"; a non-constant type.
177 Both calls to f are at the same type, so we can specialise to give:
179 g a (d::Num a) (y::a) = let f@a (x::a) = +.sel a d x x
180 in +.sel a d (f@a y) (f@a y)
183 (b) The other case is when the type variables in the instance types
184 are *not* in scope at the definition point of f. The example we are
185 working with above is a good case. There are two instances of (+.sel a d),
186 but "a" is not in scope at the definition of +.sel. Can we do anything?
187 Yes, we can "common them up", a sort of limited common sub-expression deal.
190 g a (d::Num a) (y::a) = let +.sel@a = +.sel a d
191 f@a (x::a) = +.sel@a x x
192 in +.sel@a (f@a y) (f@a y)
194 This can save work, and can't be spotted by the type checker, because
195 the two instances of +.sel weren't originally at the same type.
199 * There are quite a few variations here. For example, the defn of
200 +.sel could be floated ouside the \y, to attempt to gain laziness.
201 It certainly mustn't be floated outside the \d because the d has to
204 * We don't want to inline f_rhs in this case, because
205 that will duplicate code. Just commoning up the call is the point.
207 * Nothing gets added to +.sel's IdInfo.
209 * Don't bother unless the equivalence class has more than one item!
211 Not clear whether this is all worth it. It is of course OK to
212 simply discard call-instances when passing a big lambda.
214 Polymorphism 2 -- Overloading
216 Consider a function whose most general type is
218 f :: forall a b. Ord a => [a] -> b -> b
220 There is really no point in making a version of g at Int/Int and another
221 at Int/Bool, because it's only instancing the type variable "a" which
222 buys us any efficiency. Since g is completely polymorphic in b there
223 ain't much point in making separate versions of g for the different
226 That suggests that we should identify which of g's type variables
227 are constrained (like "a") and which are unconstrained (like "b").
228 Then when taking equivalence classes in STEP 2, we ignore the type args
229 corresponding to unconstrained type variable. In STEP 3 we make
230 polymorphic versions. Thus:
232 f@t1/ = /\b -> <f_rhs> t1 b d1 d2
241 f a (d::Num a) = let g = ...
243 ...(let d1::Ord a = Num.Ord.sel a d in g a d1)...
245 Here, g is only called at one type, but the dictionary isn't in scope at the
246 definition point for g. Usually the type checker would build a
247 definition for d1 which enclosed g, but the transformation system
248 might have moved d1's defn inward. Solution: float dictionary bindings
249 outwards along with call instances.
253 f x = let g p q = p==q
259 Before specialisation, leaving out type abstractions we have
261 f df x = let g :: Eq a => a -> a -> Bool
263 h :: Num a => a -> a -> (a, Bool)
264 h dh r s = let deq = eqFromNum dh
265 in (+ dh r s, g deq r s)
269 After specialising h we get a specialised version of h, like this:
271 h' r s = let deq = eqFromNum df
272 in (+ df r s, g deq r s)
274 But we can't naively make an instance for g from this, because deq is not in scope
275 at the defn of g. Instead, we have to float out the (new) defn of deq
276 to widen its scope. Notice that this floating can't be done in advance -- it only
277 shows up when specialisation is done.
279 User SPECIALIZE pragmas
280 ~~~~~~~~~~~~~~~~~~~~~~~
281 Specialisation pragmas can be digested by the type checker, and implemented
282 by adding extra definitions along with that of f, in the same way as before
284 f@t1/t2 = <f_rhs> t1 t2 d1 d2
286 Indeed the pragmas *have* to be dealt with by the type checker, because
287 only it knows how to build the dictionaries d1 and d2! For example
289 g :: Ord a => [a] -> [a]
290 {-# SPECIALIZE f :: [Tree Int] -> [Tree Int] #-}
292 Here, the specialised version of g is an application of g's rhs to the
293 Ord dictionary for (Tree Int), which only the type checker can conjure
294 up. There might not even *be* one, if (Tree Int) is not an instance of
295 Ord! (All the other specialision has suitable dictionaries to hand
298 Problem. The type checker doesn't have to hand a convenient <f_rhs>, because
299 it is buried in a complex (as-yet-un-desugared) binding group.
302 f@t1/t2 = f* t1 t2 d1 d2
304 where f* is the Id f with an IdInfo which says "inline me regardless!".
305 Indeed all the specialisation could be done in this way.
306 That in turn means that the simplifier has to be prepared to inline absolutely
307 any in-scope let-bound thing.
310 Again, the pragma should permit polymorphism in unconstrained variables:
312 h :: Ord a => [a] -> b -> b
313 {-# SPECIALIZE h :: [Int] -> b -> b #-}
315 We *insist* that all overloaded type variables are specialised to ground types,
316 (and hence there can be no context inside a SPECIALIZE pragma).
317 We *permit* unconstrained type variables to be specialised to
319 - or left as a polymorphic type variable
320 but nothing in between. So
322 {-# SPECIALIZE h :: [Int] -> [c] -> [c] #-}
324 is *illegal*. (It can be handled, but it adds complication, and gains the
328 SPECIALISING INSTANCE DECLARATIONS
329 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
332 instance Foo a => Foo [a] where
334 {-# SPECIALIZE instance Foo [Int] #-}
336 The original instance decl creates a dictionary-function
339 dfun.Foo.List :: forall a. Foo a -> Foo [a]
341 The SPECIALIZE pragma just makes a specialised copy, just as for
342 ordinary function definitions:
344 dfun.Foo.List@Int :: Foo [Int]
345 dfun.Foo.List@Int = dfun.Foo.List Int dFooInt
347 The information about what instance of the dfun exist gets added to
348 the dfun's IdInfo in the same way as a user-defined function too.
351 Automatic instance decl specialisation?
352 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
353 Can instance decls be specialised automatically? It's tricky.
354 We could collect call-instance information for each dfun, but
355 then when we specialised their bodies we'd get new call-instances
356 for ordinary functions; and when we specialised their bodies, we might get
357 new call-instances of the dfuns, and so on. This all arises because of
358 the unrestricted mutual recursion between instance decls and value decls.
360 Still, there's no actual problem; it just means that we may not do all
361 the specialisation we could theoretically do.
363 Furthermore, instance decls are usually exported and used non-locally,
364 so we'll want to compile enough to get those specialisations done.
366 Lastly, there's no such thing as a local instance decl, so we can
367 survive solely by spitting out *usage* information, and then reading that
368 back in as a pragma when next compiling the file. So for now,
369 we only specialise instance decls in response to pragmas.
372 SPITTING OUT USAGE INFORMATION
373 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
375 To spit out usage information we need to traverse the code collecting
376 call-instance information for all imported (non-prelude?) functions
377 and data types. Then we equivalence-class it and spit it out.
379 This is done at the top-level when all the call instances which escape
380 must be for imported functions and data types.
382 *** Not currently done ***
385 Partial specialisation by pragmas
386 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
387 What about partial specialisation:
389 k :: (Ord a, Eq b) => [a] -> b -> b -> [a]
390 {-# SPECIALIZE k :: Eq b => [Int] -> b -> b -> [a] #-}
394 {-# SPECIALIZE k :: Eq b => [Int] -> [b] -> [b] -> [a] #-}
396 Seems quite reasonable. Similar things could be done with instance decls:
398 instance (Foo a, Foo b) => Foo (a,b) where
400 {-# SPECIALIZE instance Foo a => Foo (a,Int) #-}
401 {-# SPECIALIZE instance Foo b => Foo (Int,b) #-}
403 Ho hum. Things are complex enough without this. I pass.
406 Requirements for the simplifer
407 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
408 The simplifier has to be able to take advantage of the specialisation.
410 * When the simplifier finds an application of a polymorphic f, it looks in
411 f's IdInfo in case there is a suitable instance to call instead. This converts
413 f t1 t2 d1 d2 ===> f_t1_t2
415 Note that the dictionaries get eaten up too!
417 * Dictionary selection operations on constant dictionaries must be
420 +.sel Int d ===> +Int
422 The obvious way to do this is in the same way as other specialised
423 calls: +.sel has inside it some IdInfo which tells that if it's applied
424 to the type Int then it should eat a dictionary and transform to +Int.
426 In short, dictionary selectors need IdInfo inside them for constant
429 * Exactly the same applies if a superclass dictionary is being
432 Eq.sel Int d ===> dEqInt
434 * Something similar applies to dictionary construction too. Suppose
435 dfun.Eq.List is the function taking a dictionary for (Eq a) to
436 one for (Eq [a]). Then we want
438 dfun.Eq.List Int d ===> dEq.List_Int
440 Where does the Eq [Int] dictionary come from? It is built in
441 response to a SPECIALIZE pragma on the Eq [a] instance decl.
443 In short, dfun Ids need IdInfo with a specialisation for each
444 constant instance of their instance declaration.
446 All this uses a single mechanism: the SpecEnv inside an Id
449 What does the specialisation IdInfo look like?
450 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
452 The SpecEnv of an Id maps a list of types (the template) to an expression
456 For example, if f has this SpecInfo:
458 [Int, a] -> \d:Ord Int. f' a
460 it means that we can replace the call
462 f Int t ===> (\d. f' t)
464 This chucks one dictionary away and proceeds with the
465 specialised version of f, namely f'.
468 What can't be done this way?
469 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
470 There is no way, post-typechecker, to get a dictionary for (say)
471 Eq a from a dictionary for Eq [a]. So if we find
475 we can't transform to
480 eqList :: (a->a->Bool) -> [a] -> [a] -> Bool
482 Of course, we currently have no way to automatically derive
483 eqList, nor to connect it to the Eq [a] instance decl, but you
484 can imagine that it might somehow be possible. Taking advantage
485 of this is permanently ruled out.
487 Still, this is no great hardship, because we intend to eliminate
488 overloading altogether anyway!
492 A note about non-tyvar dictionaries
493 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
494 Some Ids have types like
496 forall a,b,c. Eq a -> Ord [a] -> tau
498 This seems curious at first, because we usually only have dictionary
499 args whose types are of the form (C a) where a is a type variable.
500 But this doesn't hold for the functions arising from instance decls,
501 which sometimes get arguements with types of form (C (T a)) for some
504 Should we specialise wrt this compound-type dictionary? We used to say
506 "This is a heuristic judgement, as indeed is the fact that we
507 specialise wrt only dictionaries. We choose *not* to specialise
508 wrt compound dictionaries because at the moment the only place
509 they show up is in instance decls, where they are simply plugged
510 into a returned dictionary. So nothing is gained by specialising
513 But it is simpler and more uniform to specialise wrt these dicts too;
514 and in future GHC is likely to support full fledged type signatures
516 f ;: Eq [(a,b)] => ...
519 %************************************************************************
521 \subsubsection{The new specialiser}
523 %************************************************************************
525 Our basic game plan is this. For let(rec) bound function
526 f :: (C a, D c) => (a,b,c,d) -> Bool
528 * Find any specialised calls of f, (f ts ds), where
529 ts are the type arguments t1 .. t4, and
530 ds are the dictionary arguments d1 .. d2.
532 * Add a new definition for f1 (say):
534 f1 = /\ b d -> (..body of f..) t1 b t3 d d1 d2
536 Note that we abstract over the unconstrained type arguments.
540 [t1,b,t3,d] |-> \d1 d2 -> f1 b d
542 to the specialisations of f. This will be used by the
543 simplifier to replace calls
544 (f t1 t2 t3 t4) da db
546 (\d1 d1 -> f1 t2 t4) da db
548 All the stuff about how many dictionaries to discard, and what types
549 to apply the specialised function to, are handled by the fact that the
550 SpecEnv contains a template for the result of the specialisation.
552 We don't build *partial* specialisations for f. For example:
554 f :: Eq a => a -> a -> Bool
555 {-# SPECIALISE f :: (Eq b, Eq c) => (b,c) -> (b,c) -> Bool #-}
557 Here, little is gained by making a specialised copy of f.
558 There's a distinct danger that the specialised version would
559 first build a dictionary for (Eq b, Eq c), and then select the (==)
560 method from it! Even if it didn't, not a great deal is saved.
562 We do, however, generate polymorphic, but not overloaded, specialisations:
564 f :: Eq a => [a] -> b -> b -> b
565 {#- SPECIALISE f :: [Int] -> b -> b -> b #-}
567 Hence, the invariant is this:
569 *** no specialised version is overloaded ***
572 %************************************************************************
574 \subsubsection{The exported function}
576 %************************************************************************
579 specProgram :: DynFlags -> UniqSupply -> [CoreBind] -> IO [CoreBind]
580 specProgram dflags us binds
582 showPass dflags "Specialise"
584 let binds' = initSM us (go binds `thenSM` \ (binds', uds') ->
585 returnSM (dumpAllDictBinds uds' binds'))
587 endPass dflags "Specialise" Opt_D_dump_spec binds'
589 dumpIfSet_dyn dflags Opt_D_dump_rules "Top-level specialisations"
590 (vcat (map pprTidyIdRules (concat (map bindersOf binds'))))
594 -- We need to start with a Subst that knows all the things
595 -- that are in scope, so that the substitution engine doesn't
596 -- accidentally re-use a unique that's already in use
597 -- Easiest thing is to do it all at once, as if all the top-level
598 -- decls were mutually recursive
599 top_subst = mkSubst (mkInScopeSet (mkVarSet (bindersOfBinds binds)))
601 go [] = returnSM ([], emptyUDs)
602 go (bind:binds) = go binds `thenSM` \ (binds', uds) ->
603 specBind top_subst bind uds `thenSM` \ (bind', uds') ->
604 returnSM (bind' ++ binds', uds')
607 %************************************************************************
609 \subsubsection{@specExpr@: the main function}
611 %************************************************************************
614 specVar :: Subst -> Id -> CoreExpr
615 specVar subst v = case substId subst v of
619 specExpr :: Subst -> CoreExpr -> SpecM (CoreExpr, UsageDetails)
620 -- We carry a substitution down:
621 -- a) we must clone any binding that might flaot outwards,
622 -- to avoid name clashes
623 -- b) we carry a type substitution to use when analysing
624 -- the RHS of specialised bindings (no type-let!)
626 ---------------- First the easy cases --------------------
627 specExpr subst (Type ty) = returnSM (Type (substTy subst ty), emptyUDs)
628 specExpr subst (Var v) = returnSM (specVar subst v, emptyUDs)
629 specExpr subst (Lit lit) = returnSM (Lit lit, emptyUDs)
631 specExpr subst (Note note body)
632 = specExpr subst body `thenSM` \ (body', uds) ->
633 returnSM (Note (specNote subst note) body', uds)
636 ---------------- Applications might generate a call instance --------------------
637 specExpr subst expr@(App fun arg)
640 go (App fun arg) args = specExpr subst arg `thenSM` \ (arg', uds_arg) ->
641 go fun (arg':args) `thenSM` \ (fun', uds_app) ->
642 returnSM (App fun' arg', uds_arg `plusUDs` uds_app)
644 go (Var f) args = case specVar subst f of
645 Var f' -> returnSM (Var f', mkCallUDs subst f' args)
646 e' -> returnSM (e', emptyUDs) -- I don't expect this!
647 go other args = specExpr subst other
649 ---------------- Lambda/case require dumping of usage details --------------------
650 specExpr subst e@(Lam _ _)
651 = specExpr subst' body `thenSM` \ (body', uds) ->
653 (filtered_uds, body'') = dumpUDs bndrs' uds body'
655 returnSM (mkLams bndrs' body'', filtered_uds)
657 (bndrs, body) = collectBinders e
658 (subst', bndrs') = simplBndrs subst bndrs
659 -- More efficient to collect a group of binders together all at once
660 -- and we don't want to split a lambda group with dumped bindings
662 specExpr subst (Case scrut case_bndr ty alts)
663 = specExpr subst scrut `thenSM` \ (scrut', uds_scrut) ->
664 mapAndCombineSM spec_alt alts `thenSM` \ (alts', uds_alts) ->
665 returnSM (Case scrut' case_bndr' (substTy subst ty) alts', uds_scrut `plusUDs` uds_alts)
667 (subst_alt, case_bndr') = simplBndr subst case_bndr
668 -- No need to clone case binder; it can't float like a let(rec)
670 spec_alt (con, args, rhs)
671 = specExpr subst_rhs rhs `thenSM` \ (rhs', uds) ->
673 (uds', rhs'') = dumpUDs args uds rhs'
675 returnSM ((con, args', rhs''), uds')
677 (subst_rhs, args') = simplBndrs subst_alt args
679 ---------------- Finally, let is the interesting case --------------------
680 specExpr subst (Let bind body)
682 cloneBindSM subst bind `thenSM` \ (rhs_subst, body_subst, bind') ->
684 -- Deal with the body
685 specExpr body_subst body `thenSM` \ (body', body_uds) ->
687 -- Deal with the bindings
688 specBind rhs_subst bind' body_uds `thenSM` \ (binds', uds) ->
691 returnSM (foldr Let body' binds', uds)
693 -- Must apply the type substitution to coerceions
694 specNote subst (Coerce t1 t2) = Coerce (substTy subst t1) (substTy subst t2)
695 specNote subst note = note
698 %************************************************************************
700 \subsubsection{Dealing with a binding}
702 %************************************************************************
705 specBind :: Subst -- Use this for RHSs
707 -> UsageDetails -- Info on how the scope of the binding
708 -> SpecM ([CoreBind], -- New bindings
709 UsageDetails) -- And info to pass upstream
711 specBind rhs_subst bind body_uds
712 = specBindItself rhs_subst bind (calls body_uds) `thenSM` \ (bind', bind_uds) ->
714 bndrs = bindersOf bind
715 all_uds = zapCalls bndrs (body_uds `plusUDs` bind_uds)
716 -- It's important that the `plusUDs` is this way round,
717 -- because body_uds may bind dictionaries that are
718 -- used in the calls passed to specDefn. So the
719 -- dictionary bindings in bind_uds may mention
720 -- dictionaries bound in body_uds.
722 case splitUDs bndrs all_uds of
724 (_, ([],[])) -- This binding doesn't bind anything needed
725 -- in the UDs, so put the binding here
726 -- This is the case for most non-dict bindings, except
727 -- for the few that are mentioned in a dict binding
728 -- that is floating upwards in body_uds
729 -> returnSM ([bind'], all_uds)
731 (float_uds, (dict_binds, calls)) -- This binding is needed in the UDs, so float it out
732 -> returnSM ([], float_uds `plusUDs` mkBigUD bind' dict_binds calls)
735 -- A truly gruesome function
736 mkBigUD bind@(NonRec _ _) dbs calls
737 = -- Common case: non-recursive and no specialisations
738 -- (if there were any specialistions it would have been made recursive)
739 MkUD { dict_binds = listToBag (mkDB bind : dbs),
740 calls = listToCallDetails calls }
742 mkBigUD bind dbs calls
744 MkUD { dict_binds = unitBag (mkDB (Rec (bind_prs bind ++ dbsToPairs dbs))),
746 calls = listToCallDetails calls }
748 bind_prs (NonRec b r) = [(b,r)]
749 bind_prs (Rec prs) = prs
752 dbsToPairs ((bind,_):dbs) = bind_prs bind ++ dbsToPairs dbs
754 -- specBindItself deals with the RHS, specialising it according
755 -- to the calls found in the body (if any)
756 specBindItself rhs_subst (NonRec bndr rhs) call_info
757 = specDefn rhs_subst call_info (bndr,rhs) `thenSM` \ ((bndr',rhs'), spec_defns, spec_uds) ->
759 new_bind | null spec_defns = NonRec bndr' rhs'
760 | otherwise = Rec ((bndr',rhs'):spec_defns)
761 -- bndr' mentions the spec_defns in its SpecEnv
762 -- Not sure why we couln't just put the spec_defns first
764 returnSM (new_bind, spec_uds)
766 specBindItself rhs_subst (Rec pairs) call_info
767 = mapSM (specDefn rhs_subst call_info) pairs `thenSM` \ stuff ->
769 (pairs', spec_defns_s, spec_uds_s) = unzip3 stuff
770 spec_defns = concat spec_defns_s
771 spec_uds = plusUDList spec_uds_s
772 new_bind = Rec (spec_defns ++ pairs')
774 returnSM (new_bind, spec_uds)
777 specDefn :: Subst -- Subst to use for RHS
778 -> CallDetails -- Info on how it is used in its scope
779 -> (Id, CoreExpr) -- The thing being bound and its un-processed RHS
780 -> SpecM ((Id, CoreExpr), -- The thing and its processed RHS
781 -- the Id may now have specialisations attached
782 [(Id,CoreExpr)], -- Extra, specialised bindings
783 UsageDetails -- Stuff to fling upwards from the RHS and its
784 ) -- specialised versions
786 specDefn subst calls (fn, rhs)
787 -- The first case is the interesting one
788 | rhs_tyvars `lengthIs` n_tyvars -- Rhs of fn's defn has right number of big lambdas
789 && rhs_bndrs `lengthAtLeast` n_dicts -- and enough dict args
790 && notNull calls_for_me -- And there are some calls to specialise
792 -- At one time I tried not specialising small functions
793 -- but sometimes there are big functions marked INLINE
794 -- that we'd like to specialise. In particular, dictionary
795 -- functions, which Marcin is keen to inline
796 -- && not (certainlyWillInline fn) -- And it's not small
797 -- If it's small, it's better just to inline
798 -- it than to construct lots of specialisations
799 = -- Specialise the body of the function
800 specExpr subst rhs `thenSM` \ (rhs', rhs_uds) ->
802 -- Make a specialised version for each call in calls_for_me
803 mapSM spec_call calls_for_me `thenSM` \ stuff ->
805 (spec_defns, spec_uds, spec_rules) = unzip3 stuff
807 fn' = addIdSpecialisations zapped_fn spec_rules
809 returnSM ((fn',rhs'),
811 rhs_uds `plusUDs` plusUDList spec_uds)
813 | otherwise -- No calls or RHS doesn't fit our preconceptions
814 = specExpr subst rhs `thenSM` \ (rhs', rhs_uds) ->
815 returnSM ((zapped_fn, rhs'), [], rhs_uds)
818 zapped_fn = zapSpecPragmaId fn
819 -- If the fn is a SpecPragmaId, make it discardable
820 -- It's role as a holder for a call instance is o'er
821 -- But it might be alive for some other reason by now.
824 (tyvars, theta, _) = tcSplitSigmaTy fn_type
825 n_tyvars = length tyvars
826 n_dicts = length theta
828 (rhs_tyvars, rhs_ids, rhs_body)
829 = collectTyAndValBinders (dropInline rhs)
830 -- It's important that we "see past" any INLINE pragma
831 -- else we'll fail to specialise an INLINE thing
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 = Rule (mkFastString ("SPEC " ++ showSDoc (ppr 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 -- NOTE: we don't add back in any INLINE pragma on the RHS, so even if
904 -- the original function said INLINE, the specialised copies won't.
905 -- The idea is that the point of inlining was precisely to specialise
906 -- the function at its call site, and that's not so important for the
907 -- specialised copies. But it still smells like an ad hoc decision.
910 returnSM ((spec_f, spec_rhs),
915 my_zipEqual doc xs ys
916 | not (equalLength xs ys) = pprPanic "my_zipEqual" (ppr xs $$ ppr ys $$ (ppr fn <+> ppr call_ts) $$ ppr rhs)
917 | otherwise = zipEqual doc xs ys
919 dropInline :: CoreExpr -> CoreExpr
920 dropInline (Note InlineMe rhs) = rhs
924 %************************************************************************
926 \subsubsection{UsageDetails and suchlike}
928 %************************************************************************
933 dict_binds :: !(Bag DictBind),
934 -- Floated dictionary bindings
935 -- The order is important;
936 -- in ds1 `union` ds2, bindings in ds2 can depend on those in ds1
937 -- (Remember, Bags preserve order in GHC.)
939 calls :: !CallDetails
942 type DictBind = (CoreBind, VarSet)
943 -- The set is the free vars of the binding
944 -- both tyvars and dicts
946 type DictExpr = CoreExpr
948 emptyUDs = MkUD { dict_binds = emptyBag, calls = emptyFM }
950 type ProtoUsageDetails = ([DictBind],
951 [(Id, CallKey, ([DictExpr], VarSet))]
954 ------------------------------------------------------------
955 type CallDetails = FiniteMap Id CallInfo
956 newtype CallKey = CallKey [Maybe Type] -- Nothing => unconstrained type argument
957 type CallInfo = FiniteMap CallKey
958 ([DictExpr], VarSet) -- Dict args and the vars of the whole
959 -- call (including tyvars)
960 -- [*not* include the main id itself, of course]
961 -- The finite maps eliminate duplicates
962 -- The list of types and dictionaries is guaranteed to
963 -- match the type of f
965 -- Type isn't an instance of Ord, so that we can control which
966 -- instance we use. That's tiresome here. Oh well
967 instance Eq CallKey where
968 k1 == k2 = case k1 `compare` k2 of { EQ -> True; other -> False }
970 instance Ord CallKey where
971 compare (CallKey k1) (CallKey k2) = cmpList cmp k1 k2
973 cmp Nothing Nothing = EQ
974 cmp Nothing (Just t2) = LT
975 cmp (Just t1) Nothing = GT
976 cmp (Just t1) (Just t2) = tcCmpType t1 t2
978 unionCalls :: CallDetails -> CallDetails -> CallDetails
979 unionCalls c1 c2 = plusFM_C plusFM c1 c2
981 singleCall :: Id -> [Maybe Type] -> [DictExpr] -> CallDetails
982 singleCall id tys dicts
983 = unitFM id (unitFM (CallKey tys) (dicts, call_fvs))
985 call_fvs = exprsFreeVars dicts `unionVarSet` tys_fvs
986 tys_fvs = tyVarsOfTypes (catMaybes tys)
987 -- The type args (tys) are guaranteed to be part of the dictionary
988 -- types, because they are just the constrained types,
989 -- and the dictionary is therefore sure to be bound
990 -- inside the binding for any type variables free in the type;
991 -- hence it's safe to neglect tyvars free in tys when making
992 -- the free-var set for this call
993 -- BUT I don't trust this reasoning; play safe and include tys_fvs
995 -- We don't include the 'id' itself.
997 listToCallDetails calls
998 = foldr (unionCalls . mk_call) emptyFM calls
1000 mk_call (id, tys, dicts_w_fvs) = unitFM id (unitFM tys dicts_w_fvs)
1001 -- NB: the free vars of the call are provided
1003 callDetailsToList calls = [ (id,tys,dicts)
1004 | (id,fm) <- fmToList calls,
1005 (tys, dicts) <- fmToList fm
1008 mkCallUDs subst f args
1010 || not (all isClassPred theta)
1011 -- Only specialise if all overloading is on class params.
1012 -- In ptic, with implicit params, the type args
1013 -- *don't* say what the value of the implicit param is!
1014 || not (spec_tys `lengthIs` n_tyvars)
1015 || not ( dicts `lengthIs` n_dicts)
1016 || maybeToBool (lookupRule (\act -> True) (substInScope subst) f args)
1017 -- There's already a rule covering this call. A typical case
1018 -- is where there's an explicit user-provided rule. Then
1019 -- we don't want to create a specialised version
1020 -- of the function that overlaps.
1021 = emptyUDs -- Not overloaded, or no specialisation wanted
1024 = MkUD {dict_binds = emptyBag,
1025 calls = singleCall f spec_tys dicts
1028 (tyvars, theta, _) = tcSplitSigmaTy (idType f)
1029 constrained_tyvars = tyVarsOfTheta theta
1030 n_tyvars = length tyvars
1031 n_dicts = length theta
1033 spec_tys = [mk_spec_ty tv ty | (tv, Type ty) <- tyvars `zip` args]
1034 dicts = [dict_expr | (_, dict_expr) <- theta `zip` (drop n_tyvars args)]
1037 | tyvar `elemVarSet` constrained_tyvars = Just ty
1038 | otherwise = Nothing
1040 ------------------------------------------------------------
1041 plusUDs :: UsageDetails -> UsageDetails -> UsageDetails
1042 plusUDs (MkUD {dict_binds = db1, calls = calls1})
1043 (MkUD {dict_binds = db2, calls = calls2})
1044 = MkUD {dict_binds = d, calls = c}
1046 d = db1 `unionBags` db2
1047 c = calls1 `unionCalls` calls2
1049 plusUDList = foldr plusUDs emptyUDs
1051 -- zapCalls deletes calls to ids from uds
1052 zapCalls ids uds = uds {calls = delListFromFM (calls uds) ids}
1054 mkDB bind = (bind, bind_fvs bind)
1056 bind_fvs (NonRec bndr rhs) = exprFreeVars rhs
1057 bind_fvs (Rec prs) = foldl delVarSet rhs_fvs bndrs
1060 rhs_fvs = unionVarSets [exprFreeVars rhs | (bndr,rhs) <- prs]
1062 addDictBind (dict,rhs) uds = uds { dict_binds = mkDB (NonRec dict rhs) `consBag` dict_binds uds }
1064 dumpAllDictBinds (MkUD {dict_binds = dbs}) binds
1065 = foldrBag add binds dbs
1067 add (bind,_) binds = bind : binds
1069 dumpUDs :: [CoreBndr]
1070 -> UsageDetails -> CoreExpr
1071 -> (UsageDetails, CoreExpr)
1072 dumpUDs bndrs uds body
1073 = (free_uds, foldr add_let body dict_binds)
1075 (free_uds, (dict_binds, _)) = splitUDs bndrs uds
1076 add_let (bind,_) body = Let bind body
1078 splitUDs :: [CoreBndr]
1080 -> (UsageDetails, -- These don't mention the binders
1081 ProtoUsageDetails) -- These do
1083 splitUDs bndrs uds@(MkUD {dict_binds = orig_dbs,
1084 calls = orig_calls})
1086 = if isEmptyBag dump_dbs && null dump_calls then
1087 -- Common case: binder doesn't affect floats
1091 -- Binders bind some of the fvs of the floats
1092 (MkUD {dict_binds = free_dbs,
1093 calls = listToCallDetails free_calls},
1094 (bagToList dump_dbs, dump_calls)
1098 bndr_set = mkVarSet bndrs
1100 (free_dbs, dump_dbs, dump_idset)
1101 = foldlBag dump_db (emptyBag, emptyBag, bndr_set) orig_dbs
1102 -- Important that it's foldl not foldr;
1103 -- we're accumulating the set of dumped ids in dump_set
1105 -- Filter out any calls that mention things that are being dumped
1106 orig_call_list = callDetailsToList orig_calls
1107 (dump_calls, free_calls) = partition captured orig_call_list
1108 captured (id,tys,(dicts, fvs)) = fvs `intersectsVarSet` dump_idset
1109 || id `elemVarSet` dump_idset
1111 dump_db (free_dbs, dump_dbs, dump_idset) db@(bind, fvs)
1112 | dump_idset `intersectsVarSet` fvs -- Dump it
1113 = (free_dbs, dump_dbs `snocBag` db,
1114 extendVarSetList dump_idset (bindersOf bind))
1116 | otherwise -- Don't dump it
1117 = (free_dbs `snocBag` db, dump_dbs, dump_idset)
1121 %************************************************************************
1123 \subsubsection{Boring helper functions}
1125 %************************************************************************
1128 type SpecM a = UniqSM a
1132 getUniqSM = getUniqueUs
1136 mapAndCombineSM f [] = returnSM ([], emptyUDs)
1137 mapAndCombineSM f (x:xs) = f x `thenSM` \ (y, uds1) ->
1138 mapAndCombineSM f xs `thenSM` \ (ys, uds2) ->
1139 returnSM (y:ys, uds1 `plusUDs` uds2)
1141 cloneBindSM :: Subst -> CoreBind -> SpecM (Subst, Subst, CoreBind)
1142 -- Clone the binders of the bind; return new bind with the cloned binders
1143 -- Return the substitution to use for RHSs, and the one to use for the body
1144 cloneBindSM subst (NonRec bndr rhs)
1145 = getUs `thenUs` \ us ->
1147 (subst', bndr') = substAndCloneId subst us bndr
1149 returnUs (subst, subst', NonRec bndr' rhs)
1151 cloneBindSM subst (Rec pairs)
1152 = getUs `thenUs` \ us ->
1154 (subst', bndrs') = substAndCloneRecIds subst us (map fst pairs)
1156 returnUs (subst', subst', Rec (bndrs' `zip` map snd pairs))
1158 cloneBinders subst bndrs
1159 = getUs `thenUs` \ us ->
1160 returnUs (substAndCloneIds subst us bndrs)
1162 newIdSM old_id new_ty
1163 = getUniqSM `thenSM` \ uniq ->
1165 -- Give the new Id a similar occurrence name to the old one
1166 name = idName old_id
1167 new_id = mkUserLocal (mkSpecOcc (nameOccName name)) uniq new_ty (getSrcLoc name)
1173 Old (but interesting) stuff about unboxed bindings
1174 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1176 What should we do when a value is specialised to a *strict* unboxed value?
1178 map_*_* f (x:xs) = let h = f x
1182 Could convert let to case:
1184 map_*_Int# f (x:xs) = case f x of h# ->
1188 This may be undesirable since it forces evaluation here, but the value
1189 may not be used in all branches of the body. In the general case this
1190 transformation is impossible since the mutual recursion in a letrec
1191 cannot be expressed as a case.
1193 There is also a problem with top-level unboxed values, since our
1194 implementation cannot handle unboxed values at the top level.
1196 Solution: Lift the binding of the unboxed value and extract it when it
1199 map_*_Int# f (x:xs) = let h = case (f x) of h# -> _Lift h#
1204 Now give it to the simplifier and the _Lifting will be optimised away.
1206 The benfit is that we have given the specialised "unboxed" values a
1207 very simplep lifted semantics and then leave it up to the simplifier to
1208 optimise it --- knowing that the overheads will be removed in nearly
1211 In particular, the value will only be evaluted in the branches of the
1212 program which use it, rather than being forced at the point where the
1213 value is bound. For example:
1215 filtermap_*_* p f (x:xs)
1222 filtermap_*_Int# p f (x:xs)
1223 = let h = case (f x) of h# -> _Lift h#
1226 True -> case h of _Lift h#
1230 The binding for h can still be inlined in the one branch and the
1231 _Lifting eliminated.
1234 Question: When won't the _Lifting be eliminated?
1236 Answer: When they at the top-level (where it is necessary) or when
1237 inlining would duplicate work (or possibly code depending on
1238 options). However, the _Lifting will still be eliminated if the
1239 strictness analyser deems the lifted binding strict.