2 % (c) The University of Glasgow 2006
3 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
10 tcSimplifyInfer, tcSimplifyInferCheck,
11 tcSimplifyCheck, tcSimplifyRestricted,
12 tcSimplifyRuleLhs, tcSimplifyIPs,
13 tcSimplifySuperClasses,
14 tcSimplifyTop, tcSimplifyInteractive,
15 tcSimplifyBracket, tcSimplifyCheckPat,
17 tcSimplifyDeriv, tcSimplifyDefault,
23 #include "HsVersions.h"
25 import {-# SOURCE #-} TcUnify( unifyType )
29 import TcHsSyn ( hsLPatType )
37 import DsUtils -- Big-tuple functions
65 %************************************************************************
69 %************************************************************************
71 --------------------------------------
72 Notes on functional dependencies (a bug)
73 --------------------------------------
80 instance D a b => C a b -- Undecidable
81 -- (Not sure if it's crucial to this eg)
82 f :: C a b => a -> Bool
85 g :: C a b => a -> Bool
88 Here f typechecks, but g does not!! Reason: before doing improvement,
89 we reduce the (C a b1) constraint from the call of f to (D a b1).
91 Here is a more complicated example:
94 > class Foo a b | a->b
96 > class Bar a b | a->b
100 > instance Bar Obj Obj
102 > instance (Bar a b) => Foo a b
104 > foo:: (Foo a b) => a -> String
107 > runFoo:: (forall a b. (Foo a b) => a -> w) -> w
113 Could not deduce (Bar a b) from the context (Foo a b)
114 arising from use of `foo' at <interactive>:1
116 Add (Bar a b) to the expected type of an expression
117 In the first argument of `runFoo', namely `foo'
118 In the definition of `it': it = runFoo foo
120 Why all of the sudden does GHC need the constraint Bar a b? The
121 function foo didn't ask for that...
124 The trouble is that to type (runFoo foo), GHC has to solve the problem:
126 Given constraint Foo a b
127 Solve constraint Foo a b'
129 Notice that b and b' aren't the same. To solve this, just do
130 improvement and then they are the same. But GHC currently does
135 That is usually fine, but it isn't here, because it sees that Foo a b is
136 not the same as Foo a b', and so instead applies the instance decl for
137 instance Bar a b => Foo a b. And that's where the Bar constraint comes
140 The Right Thing is to improve whenever the constraint set changes at
141 all. Not hard in principle, but it'll take a bit of fiddling to do.
143 Note [Choosing which variables to quantify]
144 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
145 Suppose we are about to do a generalisation step. We have in our hand
148 T the type of the RHS
149 C the constraints from that RHS
151 The game is to figure out
153 Q the set of type variables over which to quantify
154 Ct the constraints we will *not* quantify over
155 Cq the constraints we will quantify over
157 So we're going to infer the type
161 and float the constraints Ct further outwards.
163 Here are the things that *must* be true:
165 (A) Q intersect fv(G) = EMPTY limits how big Q can be
166 (B) Q superset fv(Cq union T) \ oclose(fv(G),C) limits how small Q can be
168 (A) says we can't quantify over a variable that's free in the environment.
169 (B) says we must quantify over all the truly free variables in T, else
170 we won't get a sufficiently general type.
172 We do not *need* to quantify over any variable that is fixed by the
173 free vars of the environment G.
175 BETWEEN THESE TWO BOUNDS, ANY Q WILL DO!
177 Example: class H x y | x->y where ...
179 fv(G) = {a} C = {H a b, H c d}
182 (A) Q intersect {a} is empty
183 (B) Q superset {a,b,c,d} \ oclose({a}, C) = {a,b,c,d} \ {a,b} = {c,d}
185 So Q can be {c,d}, {b,c,d}
187 In particular, it's perfectly OK to quantify over more type variables
188 than strictly necessary; there is no need to quantify over 'b', since
189 it is determined by 'a' which is free in the envt, but it's perfectly
190 OK to do so. However we must not quantify over 'a' itself.
192 Other things being equal, however, we'd like to quantify over as few
193 variables as possible: smaller types, fewer type applications, more
194 constraints can get into Ct instead of Cq. Here's a good way to
197 Q = grow( fv(T), C ) \ oclose( fv(G), C )
199 That is, quantify over all variable that that MIGHT be fixed by the
200 call site (which influences T), but which aren't DEFINITELY fixed by
201 G. This choice definitely quantifies over enough type variables,
202 albeit perhaps too many.
204 Why grow( fv(T), C ) rather than fv(T)? Consider
206 class H x y | x->y where ...
211 If we used fv(T) = {c} we'd get the type
213 forall c. H c d => c -> b
215 And then if the fn was called at several different c's, each of
216 which fixed d differently, we'd get a unification error, because
217 d isn't quantified. Solution: quantify d. So we must quantify
218 everything that might be influenced by c.
220 Why not oclose( fv(T), C )? Because we might not be able to see
221 all the functional dependencies yet:
223 class H x y | x->y where ...
224 instance H x y => Eq (T x y) where ...
229 Now oclose(fv(T),C) = {c}, because the functional dependency isn't
230 apparent yet, and that's wrong. We must really quantify over d too.
232 There really isn't any point in quantifying over any more than
233 grow( fv(T), C ), because the call sites can't possibly influence
234 any other type variables.
238 -------------------------------------
240 -------------------------------------
242 It's very hard to be certain when a type is ambiguous. Consider
246 instance H x y => K (x,y)
248 Is this type ambiguous?
249 forall a b. (K (a,b), Eq b) => a -> a
251 Looks like it! But if we simplify (K (a,b)) we get (H a b) and
252 now we see that a fixes b. So we can't tell about ambiguity for sure
253 without doing a full simplification. And even that isn't possible if
254 the context has some free vars that may get unified. Urgle!
256 Here's another example: is this ambiguous?
257 forall a b. Eq (T b) => a -> a
258 Not if there's an insance decl (with no context)
259 instance Eq (T b) where ...
261 You may say of this example that we should use the instance decl right
262 away, but you can't always do that:
264 class J a b where ...
265 instance J Int b where ...
267 f :: forall a b. J a b => a -> a
269 (Notice: no functional dependency in J's class decl.)
270 Here f's type is perfectly fine, provided f is only called at Int.
271 It's premature to complain when meeting f's signature, or even
272 when inferring a type for f.
276 However, we don't *need* to report ambiguity right away. It'll always
277 show up at the call site.... and eventually at main, which needs special
278 treatment. Nevertheless, reporting ambiguity promptly is an excellent thing.
280 So here's the plan. We WARN about probable ambiguity if
282 fv(Cq) is not a subset of oclose(fv(T) union fv(G), C)
284 (all tested before quantification).
285 That is, all the type variables in Cq must be fixed by the the variables
286 in the environment, or by the variables in the type.
288 Notice that we union before calling oclose. Here's an example:
290 class J a b c | a b -> c
294 forall b c. (J a b c) => b -> b
296 Only if we union {a} from G with {b} from T before using oclose,
297 do we see that c is fixed.
299 It's a bit vague exactly which C we should use for this oclose call. If we
300 don't fix enough variables we might complain when we shouldn't (see
301 the above nasty example). Nothing will be perfect. That's why we can
302 only issue a warning.
305 Can we ever be *certain* about ambiguity? Yes: if there's a constraint
307 c in C such that fv(c) intersect (fv(G) union fv(T)) = EMPTY
309 then c is a "bubble"; there's no way it can ever improve, and it's
310 certainly ambiguous. UNLESS it is a constant (sigh). And what about
315 instance H x y => K (x,y)
317 Is this type ambiguous?
318 forall a b. (K (a,b), Eq b) => a -> a
320 Urk. The (Eq b) looks "definitely ambiguous" but it isn't. What we are after
321 is a "bubble" that's a set of constraints
323 Cq = Ca union Cq' st fv(Ca) intersect (fv(Cq') union fv(T) union fv(G)) = EMPTY
325 Hence another idea. To decide Q start with fv(T) and grow it
326 by transitive closure in Cq (no functional dependencies involved).
327 Now partition Cq using Q, leaving the definitely-ambiguous and probably-ok.
328 The definitely-ambiguous can then float out, and get smashed at top level
329 (which squashes out the constants, like Eq (T a) above)
332 --------------------------------------
333 Notes on principal types
334 --------------------------------------
339 f x = let g y = op (y::Int) in True
341 Here the principal type of f is (forall a. a->a)
342 but we'll produce the non-principal type
343 f :: forall a. C Int => a -> a
346 --------------------------------------
347 The need for forall's in constraints
348 --------------------------------------
350 [Exchange on Haskell Cafe 5/6 Dec 2000]
352 class C t where op :: t -> Bool
353 instance C [t] where op x = True
355 p y = (let f :: c -> Bool; f x = op (y >> return x) in f, y ++ [])
356 q y = (y ++ [], let f :: c -> Bool; f x = op (y >> return x) in f)
358 The definitions of p and q differ only in the order of the components in
359 the pair on their right-hand sides. And yet:
361 ghc and "Typing Haskell in Haskell" reject p, but accept q;
362 Hugs rejects q, but accepts p;
363 hbc rejects both p and q;
364 nhc98 ... (Malcolm, can you fill in the blank for us!).
366 The type signature for f forces context reduction to take place, and
367 the results of this depend on whether or not the type of y is known,
368 which in turn depends on which component of the pair the type checker
371 Solution: if y::m a, float out the constraints
372 Monad m, forall c. C (m c)
373 When m is later unified with [], we can solve both constraints.
376 --------------------------------------
377 Notes on implicit parameters
378 --------------------------------------
380 Note [Inheriting implicit parameters]
381 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
386 where f is *not* a top-level binding.
387 From the RHS of f we'll get the constraint (?y::Int).
388 There are two types we might infer for f:
392 (so we get ?y from the context of f's definition), or
394 f :: (?y::Int) => Int -> Int
396 At first you might think the first was better, becuase then
397 ?y behaves like a free variable of the definition, rather than
398 having to be passed at each call site. But of course, the WHOLE
399 IDEA is that ?y should be passed at each call site (that's what
400 dynamic binding means) so we'd better infer the second.
402 BOTTOM LINE: when *inferring types* you *must* quantify
403 over implicit parameters. See the predicate isFreeWhenInferring.
406 Note [Implicit parameters and ambiguity]
407 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
408 Only a *class* predicate can give rise to ambiguity
409 An *implicit parameter* cannot. For example:
410 foo :: (?x :: [a]) => Int
412 is fine. The call site will suppply a particular 'x'
414 Furthermore, the type variables fixed by an implicit parameter
415 propagate to the others. E.g.
416 foo :: (Show a, ?x::[a]) => Int
418 The type of foo looks ambiguous. But it isn't, because at a call site
420 let ?x = 5::Int in foo
421 and all is well. In effect, implicit parameters are, well, parameters,
422 so we can take their type variables into account as part of the
423 "tau-tvs" stuff. This is done in the function 'FunDeps.grow'.
426 Question 2: type signatures
427 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
428 BUT WATCH OUT: When you supply a type signature, we can't force you
429 to quantify over implicit parameters. For example:
433 This is perfectly reasonable. We do not want to insist on
435 (?x + 1) :: (?x::Int => Int)
437 That would be silly. Here, the definition site *is* the occurrence site,
438 so the above strictures don't apply. Hence the difference between
439 tcSimplifyCheck (which *does* allow implicit paramters to be inherited)
440 and tcSimplifyCheckBind (which does not).
442 What about when you supply a type signature for a binding?
443 Is it legal to give the following explicit, user type
444 signature to f, thus:
449 At first sight this seems reasonable, but it has the nasty property
450 that adding a type signature changes the dynamic semantics.
453 (let f x = (x::Int) + ?y
454 in (f 3, f 3 with ?y=5)) with ?y = 6
460 in (f 3, f 3 with ?y=5)) with ?y = 6
464 Indeed, simply inlining f (at the Haskell source level) would change the
467 Nevertheless, as Launchbury says (email Oct 01) we can't really give the
468 semantics for a Haskell program without knowing its typing, so if you
469 change the typing you may change the semantics.
471 To make things consistent in all cases where we are *checking* against
472 a supplied signature (as opposed to inferring a type), we adopt the
475 a signature does not need to quantify over implicit params.
477 [This represents a (rather marginal) change of policy since GHC 5.02,
478 which *required* an explicit signature to quantify over all implicit
479 params for the reasons mentioned above.]
481 But that raises a new question. Consider
483 Given (signature) ?x::Int
484 Wanted (inferred) ?x::Int, ?y::Bool
486 Clearly we want to discharge the ?x and float the ?y out. But
487 what is the criterion that distinguishes them? Clearly it isn't
488 what free type variables they have. The Right Thing seems to be
489 to float a constraint that
490 neither mentions any of the quantified type variables
491 nor any of the quantified implicit parameters
493 See the predicate isFreeWhenChecking.
496 Question 3: monomorphism
497 ~~~~~~~~~~~~~~~~~~~~~~~~
498 There's a nasty corner case when the monomorphism restriction bites:
502 The argument above suggests that we *must* generalise
503 over the ?y parameter, to get
504 z :: (?y::Int) => Int,
505 but the monomorphism restriction says that we *must not*, giving
507 Why does the momomorphism restriction say this? Because if you have
509 let z = x + ?y in z+z
511 you might not expect the addition to be done twice --- but it will if
512 we follow the argument of Question 2 and generalise over ?y.
515 Question 4: top level
516 ~~~~~~~~~~~~~~~~~~~~~
517 At the top level, monomorhism makes no sense at all.
520 main = let ?x = 5 in print foo
524 woggle :: (?x :: Int) => Int -> Int
527 We definitely don't want (foo :: Int) with a top-level implicit parameter
528 (?x::Int) becuase there is no way to bind it.
533 (A) Always generalise over implicit parameters
534 Bindings that fall under the monomorphism restriction can't
538 * Inlining remains valid
539 * No unexpected loss of sharing
540 * But simple bindings like
542 will be rejected, unless you add an explicit type signature
543 (to avoid the monomorphism restriction)
544 z :: (?y::Int) => Int
546 This seems unacceptable
548 (B) Monomorphism restriction "wins"
549 Bindings that fall under the monomorphism restriction can't
551 Always generalise over implicit parameters *except* for bindings
552 that fall under the monomorphism restriction
555 * Inlining isn't valid in general
556 * No unexpected loss of sharing
557 * Simple bindings like
559 accepted (get value of ?y from binding site)
561 (C) Always generalise over implicit parameters
562 Bindings that fall under the monomorphism restriction can't
563 be generalised, EXCEPT for implicit parameters
565 * Inlining remains valid
566 * Unexpected loss of sharing (from the extra generalisation)
567 * Simple bindings like
569 accepted (get value of ?y from occurrence sites)
574 None of these choices seems very satisfactory. But at least we should
575 decide which we want to do.
577 It's really not clear what is the Right Thing To Do. If you see
581 would you expect the value of ?y to be got from the *occurrence sites*
582 of 'z', or from the valuue of ?y at the *definition* of 'z'? In the
583 case of function definitions, the answer is clearly the former, but
584 less so in the case of non-fucntion definitions. On the other hand,
585 if we say that we get the value of ?y from the definition site of 'z',
586 then inlining 'z' might change the semantics of the program.
588 Choice (C) really says "the monomorphism restriction doesn't apply
589 to implicit parameters". Which is fine, but remember that every
590 innocent binding 'x = ...' that mentions an implicit parameter in
591 the RHS becomes a *function* of that parameter, called at each
592 use of 'x'. Now, the chances are that there are no intervening 'with'
593 clauses that bind ?y, so a decent compiler should common up all
594 those function calls. So I think I strongly favour (C). Indeed,
595 one could make a similar argument for abolishing the monomorphism
596 restriction altogether.
598 BOTTOM LINE: we choose (B) at present. See tcSimplifyRestricted
602 %************************************************************************
604 \subsection{tcSimplifyInfer}
606 %************************************************************************
608 tcSimplify is called when we *inferring* a type. Here's the overall game plan:
610 1. Compute Q = grow( fvs(T), C )
612 2. Partition C based on Q into Ct and Cq. Notice that ambiguous
613 predicates will end up in Ct; we deal with them at the top level
615 3. Try improvement, using functional dependencies
617 4. If Step 3 did any unification, repeat from step 1
618 (Unification can change the result of 'grow'.)
620 Note: we don't reduce dictionaries in step 2. For example, if we have
621 Eq (a,b), we don't simplify to (Eq a, Eq b). So Q won't be different
622 after step 2. However note that we may therefore quantify over more
623 type variables than we absolutely have to.
625 For the guts, we need a loop, that alternates context reduction and
626 improvement with unification. E.g. Suppose we have
628 class C x y | x->y where ...
630 and tcSimplify is called with:
632 Then improvement unifies a with b, giving
635 If we need to unify anything, we rattle round the whole thing all over
642 -> TcTyVarSet -- fv(T); type vars
644 -> TcM ([TcTyVar], -- Tyvars to quantify (zonked and quantified)
645 [Inst], -- Dict Ids that must be bound here (zonked)
646 TcDictBinds) -- Bindings
647 -- Any free (escaping) Insts are tossed into the environment
652 tcSimplifyInfer doc tau_tvs wanted
653 = do { tau_tvs1 <- zonkTcTyVarsAndFV (varSetElems tau_tvs)
654 ; wanted' <- mapM zonkInst wanted -- Zonk before deciding quantified tyvars
655 ; gbl_tvs <- tcGetGlobalTyVars
656 ; let preds1 = fdPredsOfInsts wanted'
657 gbl_tvs1 = oclose preds1 gbl_tvs
658 qtvs = growInstsTyVars wanted' tau_tvs1 `minusVarSet` gbl_tvs1
659 -- See Note [Choosing which variables to quantify]
661 -- To maximise sharing, remove from consideration any
662 -- constraints that don't mention qtvs at all
663 ; let (free, bound) = partition (isFreeWhenInferring qtvs) wanted'
666 -- To make types simple, reduce as much as possible
667 ; traceTc (text "infer" <+> (ppr preds1 $$ ppr (growInstsTyVars wanted' tau_tvs1) $$ ppr gbl_tvs $$
668 ppr gbl_tvs1 $$ ppr free $$ ppr bound))
669 ; (irreds1, binds1) <- tryHardCheckLoop doc bound
671 -- Note [Inference and implication constraints]
672 ; let want_dict d = tyVarsOfInst d `intersectsVarSet` qtvs
673 ; (irreds2, binds2) <- approximateImplications doc want_dict irreds1
675 -- Now work out all over again which type variables to quantify,
676 -- exactly in the same way as before, but starting from irreds2. Why?
677 -- a) By now improvment may have taken place, and we must *not*
678 -- quantify over any variable free in the environment
679 -- tc137 (function h inside g) is an example
681 -- b) Do not quantify over constraints that *now* do not
682 -- mention quantified type variables, because they are
683 -- simply ambiguous (or might be bound further out). Example:
684 -- f :: Eq b => a -> (a, b)
686 -- From the RHS of g we get the MethodInst f77 :: alpha -> (alpha, beta)
687 -- We decide to quantify over 'alpha' alone, but free1 does not include f77
688 -- because f77 mentions 'alpha'. Then reducing leaves only the (ambiguous)
689 -- constraint (Eq beta), which we dump back into the free set
690 -- See test tcfail181
692 -- c) irreds may contain type variables not previously mentioned,
693 -- e.g. instance D a x => Foo [a]
695 -- Then after simplifying we'll get (D a x), and x is fresh
696 -- We must quantify over x else it'll be totally unbound
697 ; tau_tvs2 <- zonkTcTyVarsAndFV (varSetElems tau_tvs1)
698 ; gbl_tvs2 <- zonkTcTyVarsAndFV (varSetElems gbl_tvs1)
699 -- Note that we start from gbl_tvs1
700 -- We use tcGetGlobalTyVars, then oclose wrt preds2, because
701 -- we've already put some of the original preds1 into frees
702 -- E.g. wanteds = C a b (where a->b)
705 -- Then b is fixed by gbl_tvs, so (C a b) will be in free, and
706 -- irreds2 will be empty. But we don't want to generalise over b!
707 ; let preds2 = fdPredsOfInsts irreds2 -- irreds2 is zonked
708 qtvs = growInstsTyVars irreds2 tau_tvs2 `minusVarSet` oclose preds2 gbl_tvs2
709 ---------------------------------------------------
710 -- BUG WARNING: there's a nasty bug lurking here
711 -- fdPredsOfInsts may return preds that mention variables quantified in
712 -- one of the implication constraints in irreds2; and that is clearly wrong:
713 -- we might quantify over too many variables through accidental capture
714 ---------------------------------------------------
715 ; let (free, irreds3) = partition (isFreeWhenInferring qtvs) irreds2
718 -- Turn the quantified meta-type variables into real type variables
719 ; qtvs2 <- zonkQuantifiedTyVars (varSetElems qtvs)
721 -- We can't abstract over any remaining unsolved
722 -- implications so instead just float them outwards. Ugh.
723 ; let (q_dicts0, implics) = partition isAbstractableInst irreds3
724 ; loc <- getInstLoc (ImplicOrigin doc)
725 ; implic_bind <- bindIrreds loc qtvs2 q_dicts0 implics
727 -- Prepare equality instances for quantification
728 ; let (q_eqs0,q_dicts) = partition isEqInst q_dicts0
729 ; q_eqs <- mapM finalizeEqInst q_eqs0
731 ; return (qtvs2, q_eqs ++ q_dicts, binds1 `unionBags` binds2 `unionBags` implic_bind) }
732 -- NB: when we are done, we might have some bindings, but
733 -- the final qtvs might be empty. See Note [NO TYVARS] below.
735 approximateImplications :: SDoc -> (Inst -> Bool) -> [Inst] -> TcM ([Inst], TcDictBinds)
736 -- Note [Inference and implication constraints]
737 -- Given a bunch of Dict and ImplicInsts, try to approximate the implications by
738 -- - fetching any dicts inside them that are free
739 -- - using those dicts as cruder constraints, to solve the implications
740 -- - returning the extra ones too
742 approximateImplications doc want_dict irreds
744 = return (irreds, emptyBag)
746 = do { extra_dicts' <- mapM cloneDict extra_dicts
747 ; tryHardCheckLoop doc (extra_dicts' ++ irreds) }
748 -- By adding extra_dicts', we make them
749 -- available to solve the implication constraints
751 extra_dicts = get_dicts (filter isImplicInst irreds)
753 get_dicts :: [Inst] -> [Inst] -- Returns only Dicts
754 -- Find the wanted constraints in implication constraints that satisfy
755 -- want_dict, and are not bound by forall's in the constraint itself
756 get_dicts ds = concatMap get_dict ds
758 get_dict d@(Dict {}) | want_dict d = [d]
760 get_dict (ImplicInst {tci_tyvars = tvs, tci_wanted = wanteds})
761 = [ d | let tv_set = mkVarSet tvs
762 , d <- get_dicts wanteds
763 , not (tyVarsOfInst d `intersectsVarSet` tv_set)]
764 get_dict i@(EqInst {}) | want_dict i = [i]
766 get_dict other = pprPanic "approximateImplications" (ppr other)
769 Note [Inference and implication constraints]
770 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
771 Suppose we have a wanted implication constraint (perhaps arising from
772 a nested pattern match) like
774 and we are now trying to quantify over 'a' when inferring the type for
775 a function. In principle it's possible that there might be an instance
776 instance (C a, E a) => D [a]
777 so the context (E a) would suffice. The Right Thing is to abstract over
778 the implication constraint, but we don't do that (a) because it'll be
779 surprising to programmers and (b) because we don't have the machinery to deal
780 with 'given' implications.
782 So our best approximation is to make (D [a]) part of the inferred
783 context, so we can use that to discharge the implication. Hence
784 the strange function get_dicts in approximateImplications.
786 The common cases are more clear-cut, when we have things like
788 Here, abstracting over (C b) is not an approximation at all -- but see
789 Note [Freeness and implications].
791 See Trac #1430 and test tc228.
795 -----------------------------------------------------------
796 -- tcSimplifyInferCheck is used when we know the constraints we are to simplify
797 -- against, but we don't know the type variables over which we are going to quantify.
798 -- This happens when we have a type signature for a mutually recursive group
801 -> TcTyVarSet -- fv(T)
804 -> TcM ([TyVar], -- Fully zonked, and quantified
805 TcDictBinds) -- Bindings
807 tcSimplifyInferCheck loc tau_tvs givens wanteds
808 = do { traceTc (text "tcSimplifyInferCheck <-" <+> ppr wanteds)
809 ; (irreds, binds) <- gentleCheckLoop loc givens wanteds
811 -- Figure out which type variables to quantify over
812 -- You might think it should just be the signature tyvars,
813 -- but in bizarre cases you can get extra ones
814 -- f :: forall a. Num a => a -> a
815 -- f x = fst (g (x, head [])) + 1
817 -- Here we infer g :: forall a b. a -> b -> (b,a)
818 -- We don't want g to be monomorphic in b just because
819 -- f isn't quantified over b.
820 ; let all_tvs = varSetElems (tau_tvs `unionVarSet` tyVarsOfInsts givens)
821 ; all_tvs <- zonkTcTyVarsAndFV all_tvs
822 ; gbl_tvs <- tcGetGlobalTyVars
823 ; let qtvs = varSetElems (all_tvs `minusVarSet` gbl_tvs)
824 -- We could close gbl_tvs, but its not necessary for
825 -- soundness, and it'll only affect which tyvars, not which
826 -- dictionaries, we quantify over
828 ; qtvs' <- zonkQuantifiedTyVars qtvs
830 -- Now we are back to normal (c.f. tcSimplCheck)
831 ; implic_bind <- bindIrreds loc qtvs' givens irreds
833 ; traceTc (text "tcSimplifyInferCheck ->" <+> ppr (implic_bind))
834 ; return (qtvs', binds `unionBags` implic_bind) }
837 Note [Squashing methods]
838 ~~~~~~~~~~~~~~~~~~~~~~~~~
839 Be careful if you want to float methods more:
840 truncate :: forall a. RealFrac a => forall b. Integral b => a -> b
841 From an application (truncate f i) we get
844 If we have also have a second occurrence of truncate, we get
847 When simplifying with i,f free, we might still notice that
848 t1=t3; but alas, the binding for t2 (which mentions t1)
849 may continue to float out!
854 class Y a b | a -> b where
857 instance Y [[a]] a where
860 k :: X a -> X a -> X a
862 g :: Num a => [X a] -> [X a]
865 h ys = ys ++ map (k (y [[0]])) xs
867 The excitement comes when simplifying the bindings for h. Initially
868 try to simplify {y @ [[t1]] t2, 0 @ t1}, with initial qtvs = {t2}.
869 From this we get t1~t2, but also various bindings. We can't forget
870 the bindings (because of [LOOP]), but in fact t1 is what g is
873 The net effect of [NO TYVARS]
879 %************************************************************************
881 \subsection{tcSimplifyCheck}
883 %************************************************************************
885 @tcSimplifyCheck@ is used when we know exactly the set of variables
886 we are going to quantify over. For example, a class or instance declaration.
889 -----------------------------------------------------------
890 -- tcSimplifyCheck is used when checking expression type signatures,
891 -- class decls, instance decls etc.
892 tcSimplifyCheck :: InstLoc
893 -> [TcTyVar] -- Quantify over these
896 -> TcM TcDictBinds -- Bindings
897 tcSimplifyCheck loc qtvs givens wanteds
898 = ASSERT( all isTcTyVar qtvs && all isSkolemTyVar qtvs )
899 do { traceTc (text "tcSimplifyCheck")
900 ; (irreds, binds) <- gentleCheckLoop loc givens wanteds
901 ; implic_bind <- bindIrreds loc qtvs givens irreds
902 ; return (binds `unionBags` implic_bind) }
904 -----------------------------------------------------------
905 -- tcSimplifyCheckPat is used for existential pattern match
906 tcSimplifyCheckPat :: InstLoc
907 -> [TcTyVar] -- Quantify over these
910 -> TcM TcDictBinds -- Bindings
911 tcSimplifyCheckPat loc qtvs givens wanteds
912 = ASSERT( all isTcTyVar qtvs && all isSkolemTyVar qtvs )
913 do { traceTc (text "tcSimplifyCheckPat")
914 ; (irreds, binds) <- gentleCheckLoop loc givens wanteds
915 ; implic_bind <- bindIrredsR loc qtvs givens irreds
916 ; return (binds `unionBags` implic_bind) }
918 -----------------------------------------------------------
919 bindIrreds :: InstLoc -> [TcTyVar]
922 bindIrreds loc qtvs givens irreds
923 = bindIrredsR loc qtvs givens irreds
925 bindIrredsR :: InstLoc -> [TcTyVar] -> [Inst] -> [Inst] -> TcM TcDictBinds
926 -- Make a binding that binds 'irreds', by generating an implication
927 -- constraint for them, *and* throwing the constraint into the LIE
928 bindIrredsR loc qtvs givens irreds
932 = do { let givens' = filter isAbstractableInst givens
933 -- The givens can (redundantly) include methods
934 -- We want to retain both EqInsts and Dicts
935 -- There should be no implicadtion constraints
936 -- See Note [Pruning the givens in an implication constraint]
938 -- If there are no 'givens', then it's safe to
939 -- partition the 'wanteds' by their qtvs, thereby trimming irreds
940 -- See Note [Freeness and implications]
941 ; irreds' <- if null givens'
943 { let qtv_set = mkVarSet qtvs
944 (frees, real_irreds) = partition (isFreeWrtTyVars qtv_set) irreds
946 ; return real_irreds }
949 ; (implics, bind) <- makeImplicationBind loc qtvs givens' irreds'
950 -- This call does the real work
951 -- If irreds' is empty, it does something sensible
956 makeImplicationBind :: InstLoc -> [TcTyVar]
958 -> TcM ([Inst], TcDictBinds)
959 -- Make a binding that binds 'irreds', by generating an implication
960 -- constraint for them.
962 -- The binding looks like
963 -- (ir1, .., irn) = f qtvs givens
964 -- where f is (evidence for) the new implication constraint
965 -- f :: forall qtvs. givens => (ir1, .., irn)
966 -- qtvs includes coercion variables
968 -- This binding must line up the 'rhs' in reduceImplication
969 makeImplicationBind loc all_tvs
970 givens -- Guaranteed all Dicts or EqInsts
972 | null irreds -- If there are no irreds, we are done
973 = return ([], emptyBag)
974 | otherwise -- Otherwise we must generate a binding
975 = do { uniq <- newUnique
976 ; span <- getSrcSpanM
977 ; let (eq_givens, dict_givens) = partition isEqInst givens
979 -- extract equality binders
980 eq_cotvs = map eqInstType eq_givens
982 -- make the implication constraint instance
983 name = mkInternalName uniq (mkVarOcc "ic") span
984 implic_inst = ImplicInst { tci_name = name,
985 tci_tyvars = all_tvs,
986 tci_given = eq_givens ++ dict_givens,
987 -- same order as binders
991 -- create binders for the irreducible dictionaries
992 dict_irreds = filter (not . isEqInst) irreds
993 dict_irred_ids = map instToId dict_irreds
994 lpat = mkBigLHsPatTup (map (L span . VarPat) dict_irred_ids)
996 -- create the binding
997 rhs = L span (mkHsWrap co (HsVar (instToId implic_inst)))
998 co = mkWpApps (map instToId dict_givens)
999 <.> mkWpTyApps eq_cotvs
1000 <.> mkWpTyApps (mkTyVarTys all_tvs)
1001 bind | [dict_irred_id] <- dict_irred_ids
1002 = mkVarBind dict_irred_id rhs
1005 PatBind { pat_lhs = lpat
1006 , pat_rhs = unguardedGRHSs rhs
1007 , pat_rhs_ty = hsLPatType lpat
1008 , bind_fvs = placeHolderNames
1011 ; traceTc $ text "makeImplicationBind" <+> ppr implic_inst
1012 ; return ([implic_inst], unitBag bind)
1015 -----------------------------------------------------------
1016 tryHardCheckLoop :: SDoc
1018 -> TcM ([Inst], TcDictBinds)
1020 tryHardCheckLoop doc wanteds
1021 = do { (irreds,binds) <- checkLoop (mkInferRedEnv doc try_me) wanteds
1022 ; return (irreds,binds)
1026 -- Here's the try-hard bit
1028 -----------------------------------------------------------
1029 gentleCheckLoop :: InstLoc
1032 -> TcM ([Inst], TcDictBinds)
1034 gentleCheckLoop inst_loc givens wanteds
1035 = do { (irreds,binds) <- checkLoop env wanteds
1036 ; return (irreds,binds)
1039 env = mkRedEnv (pprInstLoc inst_loc) try_me givens
1041 try_me inst | isMethodOrLit inst = ReduceMe
1043 -- When checking against a given signature
1044 -- we MUST be very gentle: Note [Check gently]
1046 gentleInferLoop :: SDoc -> [Inst]
1047 -> TcM ([Inst], TcDictBinds)
1048 gentleInferLoop doc wanteds
1049 = do { (irreds, binds) <- checkLoop env wanteds
1050 ; return (irreds, binds) }
1052 env = mkInferRedEnv doc try_me
1053 try_me inst | isMethodOrLit inst = ReduceMe
1058 ~~~~~~~~~~~~~~~~~~~~
1059 We have to very careful about not simplifying too vigorously
1064 f :: Show b => T b -> b
1065 f (MkT x) = show [x]
1067 Inside the pattern match, which binds (a:*, x:a), we know that
1069 Hence we have a dictionary for Show [a] available; and indeed we
1070 need it. We are going to build an implication contraint
1071 forall a. (b~[a]) => Show [a]
1072 Later, we will solve this constraint using the knowledge (Show b)
1074 But we MUST NOT reduce (Show [a]) to (Show a), else the whole
1075 thing becomes insoluble. So we simplify gently (get rid of literals
1076 and methods only, plus common up equal things), deferring the real
1077 work until top level, when we solve the implication constraint
1078 with tryHardCheckLooop.
1082 -----------------------------------------------------------
1085 -> TcM ([Inst], TcDictBinds)
1086 -- Precondition: givens are completely rigid
1087 -- Postcondition: returned Insts are zonked
1089 checkLoop env wanteds
1091 where go env wanteds
1092 = do { -- We do need to zonk the givens; cf Note [Zonking RedEnv]
1093 ; env' <- zonkRedEnv env
1094 ; wanteds' <- zonkInsts wanteds
1096 ; (improved, tybinds, binds, irreds)
1097 <- reduceContext env' wanteds'
1098 ; execTcTyVarBinds tybinds
1100 ; if null irreds || not improved then
1101 return (irreds, binds)
1104 -- If improvement did some unification, we go round again.
1105 -- We start again with irreds, not wanteds
1106 -- Using an instance decl might have introduced a fresh type
1107 -- variable which might have been unified, so we'd get an
1108 -- infinite loop if we started again with wanteds!
1110 { (irreds1, binds1) <- go env' irreds
1111 ; return (irreds1, binds `unionBags` binds1) } }
1114 Note [Zonking RedEnv]
1115 ~~~~~~~~~~~~~~~~~~~~~
1116 It might appear as if the givens in RedEnv are always rigid, but that is not
1117 necessarily the case for programs involving higher-rank types that have class
1118 contexts constraining the higher-rank variables. An example from tc237 in the
1121 class Modular s a | s -> a
1123 wim :: forall a w. Integral a
1124 => a -> (forall s. Modular s a => M s w) -> w
1125 wim i k = error "urk"
1127 test5 :: (Modular s a, Integral a) => M s a
1130 test4 = wim 4 test4'
1132 Notice how the variable 'a' of (Modular s a) in the rank-2 type of wim is
1133 quantified further outside. When type checking test4, we have to check
1134 whether the signature of test5 is an instance of
1136 (forall s. Modular s a => M s w)
1138 Consequently, we will get (Modular s t_a), where t_a is a TauTv into the
1141 Given the FD of Modular in this example, class improvement will instantiate
1142 t_a to 'a', where 'a' is the skolem from test5's signatures (due to the
1143 Modular s a predicate in that signature). If we don't zonk (Modular s t_a) in
1144 the givens, we will get into a loop as improveOne uses the unification engine
1145 Unify.tcUnifyTys, which doesn't know about mutable type variables.
1150 class If b t e r | b t e -> r
1153 class Lte a b c | a b -> c where lte :: a -> b -> c
1155 instance (Lte a b l,If l b a c) => Max a b c
1157 Wanted: Max Z (S x) y
1159 Then we'll reduce using the Max instance to:
1160 (Lte Z (S x) l, If l (S x) Z y)
1161 and improve by binding l->T, after which we can do some reduction
1162 on both the Lte and If constraints. What we *can't* do is start again
1163 with (Max Z (S x) y)!
1167 %************************************************************************
1169 tcSimplifySuperClasses
1171 %************************************************************************
1173 Note [SUPERCLASS-LOOP 1]
1174 ~~~~~~~~~~~~~~~~~~~~~~~~
1175 We have to be very, very careful when generating superclasses, lest we
1176 accidentally build a loop. Here's an example:
1180 class S a => C a where { opc :: a -> a }
1181 class S b => D b where { opd :: b -> b }
1183 instance C Int where
1186 instance D Int where
1189 From (instance C Int) we get the constraint set {ds1:S Int, dd:D Int}
1190 Simplifying, we may well get:
1191 $dfCInt = :C ds1 (opd dd)
1194 Notice that we spot that we can extract ds1 from dd.
1196 Alas! Alack! We can do the same for (instance D Int):
1198 $dfDInt = :D ds2 (opc dc)
1202 And now we've defined the superclass in terms of itself.
1203 Two more nasty cases are in
1208 - Satisfy the superclass context *all by itself*
1209 (tcSimplifySuperClasses)
1210 - And do so completely; i.e. no left-over constraints
1211 to mix with the constraints arising from method declarations
1214 Note [Recursive instances and superclases]
1215 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1216 Consider this code, which arises in the context of "Scrap Your
1217 Boilerplate with Class".
1221 instance Sat (ctx Char) => Data ctx Char
1222 instance (Sat (ctx [a]), Data ctx a) => Data ctx [a]
1224 class Data Maybe a => Foo a
1226 instance Foo t => Sat (Maybe t)
1228 instance Data Maybe a => Foo a
1229 instance Foo a => Foo [a]
1232 In the instance for Foo [a], when generating evidence for the superclasses
1233 (ie in tcSimplifySuperClasses) we need a superclass (Data Maybe [a]).
1234 Using the instance for Data, we therefore need
1235 (Sat (Maybe [a], Data Maybe a)
1236 But we are given (Foo a), and hence its superclass (Data Maybe a).
1237 So that leaves (Sat (Maybe [a])). Using the instance for Sat means
1238 we need (Foo [a]). And that is the very dictionary we are bulding
1239 an instance for! So we must put that in the "givens". So in this
1241 Given: Foo a, Foo [a]
1242 Watend: Data Maybe [a]
1244 BUT we must *not not not* put the *superclasses* of (Foo [a]) in
1245 the givens, which is what 'addGiven' would normally do. Why? Because
1246 (Data Maybe [a]) is the superclass, so we'd "satisfy" the wanted
1247 by selecting a superclass from Foo [a], which simply makes a loop.
1249 On the other hand we *must* put the superclasses of (Foo a) in
1250 the givens, as you can see from the derivation described above.
1252 Conclusion: in the very special case of tcSimplifySuperClasses
1253 we have one 'given' (namely the "this" dictionary) whose superclasses
1254 must not be added to 'givens' by addGiven.
1256 There is a complication though. Suppose there are equalities
1257 instance (Eq a, a~b) => Num (a,b)
1258 Then we normalise the 'givens' wrt the equalities, so the original
1259 given "this" dictionary is cast to one of a different type. So it's a
1260 bit trickier than before to identify the "special" dictionary whose
1261 superclasses must not be added. See test
1262 indexed-types/should_run/EqInInstance
1264 We need a persistent property of the dictionary to record this
1265 special-ness. Current I'm using the InstLocOrigin (a bit of a hack,
1266 but cool), which is maintained by dictionary normalisation.
1267 Specifically, the InstLocOrigin is
1269 then the no-superclass thing kicks in. WATCH OUT if you fiddle
1273 tcSimplifySuperClasses
1275 -> Inst -- The dict whose superclasses
1276 -- are being figured out
1280 tcSimplifySuperClasses loc this givens sc_wanteds
1281 = do { traceTc (text "tcSimplifySuperClasses")
1283 -- Note [Recursive instances and superclases]
1284 ; no_sc_loc <- getInstLoc NoScOrigin
1285 ; let no_sc_this = setInstLoc this no_sc_loc
1287 ; let env = RedEnv { red_doc = pprInstLoc loc,
1288 red_try_me = try_me,
1289 red_givens = no_sc_this : givens,
1291 red_improve = False } -- No unification vars
1294 ; (irreds,binds1) <- checkLoop env sc_wanteds
1295 ; let (tidy_env, tidy_irreds) = tidyInsts irreds
1296 ; reportNoInstances tidy_env (Just (loc, givens)) [] tidy_irreds
1299 try_me _ = ReduceMe -- Try hard, so we completely solve the superclass
1300 -- constraints right here. See Note [SUPERCLASS-LOOP 1]
1304 %************************************************************************
1306 \subsection{tcSimplifyRestricted}
1308 %************************************************************************
1310 tcSimplifyRestricted infers which type variables to quantify for a
1311 group of restricted bindings. This isn't trivial.
1314 We want to quantify over a to get id :: forall a. a->a
1317 We do not want to quantify over a, because there's an Eq a
1318 constraint, so we get eq :: a->a->Bool (notice no forall)
1321 RHS has type 'tau', whose free tyvars are tau_tvs
1322 RHS has constraints 'wanteds'
1325 Quantify over (tau_tvs \ ftvs(wanteds))
1326 This is bad. The constraints may contain (Monad (ST s))
1327 where we have instance Monad (ST s) where...
1328 so there's no need to be monomorphic in s!
1330 Also the constraint might be a method constraint,
1331 whose type mentions a perfectly innocent tyvar:
1332 op :: Num a => a -> b -> a
1333 Here, b is unconstrained. A good example would be
1335 We want to infer the polymorphic type
1336 foo :: forall b. b -> b
1339 Plan B (cunning, used for a long time up to and including GHC 6.2)
1340 Step 1: Simplify the constraints as much as possible (to deal
1341 with Plan A's problem). Then set
1342 qtvs = tau_tvs \ ftvs( simplify( wanteds ) )
1344 Step 2: Now simplify again, treating the constraint as 'free' if
1345 it does not mention qtvs, and trying to reduce it otherwise.
1346 The reasons for this is to maximise sharing.
1348 This fails for a very subtle reason. Suppose that in the Step 2
1349 a constraint (Foo (Succ Zero) (Succ Zero) b) gets thrown upstairs as 'free'.
1350 In the Step 1 this constraint might have been simplified, perhaps to
1351 (Foo Zero Zero b), AND THEN THAT MIGHT BE IMPROVED, to bind 'b' to 'T'.
1352 This won't happen in Step 2... but that in turn might prevent some other
1353 constraint (Baz [a] b) being simplified (e.g. via instance Baz [a] T where {..})
1354 and that in turn breaks the invariant that no constraints are quantified over.
1356 Test typecheck/should_compile/tc177 (which failed in GHC 6.2) demonstrates
1361 Step 1: Simplify the constraints as much as possible (to deal
1362 with Plan A's problem). Then set
1363 qtvs = tau_tvs \ ftvs( simplify( wanteds ) )
1364 Return the bindings from Step 1.
1367 A note about Plan C (arising from "bug" reported by George Russel March 2004)
1370 instance (HasBinary ty IO) => HasCodedValue ty
1372 foo :: HasCodedValue a => String -> IO a
1374 doDecodeIO :: HasCodedValue a => () -> () -> IO a
1375 doDecodeIO codedValue view
1376 = let { act = foo "foo" } in act
1378 You might think this should work becuase the call to foo gives rise to a constraint
1379 (HasCodedValue t), which can be satisfied by the type sig for doDecodeIO. But the
1380 restricted binding act = ... calls tcSimplifyRestricted, and PlanC simplifies the
1381 constraint using the (rather bogus) instance declaration, and now we are stuffed.
1383 I claim this is not really a bug -- but it bit Sergey as well as George. So here's
1387 Plan D (a variant of plan B)
1388 Step 1: Simplify the constraints as much as possible (to deal
1389 with Plan A's problem), BUT DO NO IMPROVEMENT. Then set
1390 qtvs = tau_tvs \ ftvs( simplify( wanteds ) )
1392 Step 2: Now simplify again, treating the constraint as 'free' if
1393 it does not mention qtvs, and trying to reduce it otherwise.
1395 The point here is that it's generally OK to have too few qtvs; that is,
1396 to make the thing more monomorphic than it could be. We don't want to
1397 do that in the common cases, but in wierd cases it's ok: the programmer
1398 can always add a signature.
1400 Too few qtvs => too many wanteds, which is what happens if you do less
1405 tcSimplifyRestricted -- Used for restricted binding groups
1406 -- i.e. ones subject to the monomorphism restriction
1409 -> [Name] -- Things bound in this group
1410 -> TcTyVarSet -- Free in the type of the RHSs
1411 -> [Inst] -- Free in the RHSs
1412 -> TcM ([TyVar], -- Tyvars to quantify (zonked and quantified)
1413 TcDictBinds) -- Bindings
1414 -- tcSimpifyRestricted returns no constraints to
1415 -- quantify over; by definition there are none.
1416 -- They are all thrown back in the LIE
1418 tcSimplifyRestricted doc top_lvl bndrs tau_tvs wanteds
1419 -- Zonk everything in sight
1420 = do { traceTc (text "tcSimplifyRestricted")
1421 ; wanteds_z <- zonkInsts wanteds
1423 -- 'ReduceMe': Reduce as far as we can. Don't stop at
1424 -- dicts; the idea is to get rid of as many type
1425 -- variables as possible, and we don't want to stop
1426 -- at (say) Monad (ST s), because that reduces
1427 -- immediately, with no constraint on s.
1429 -- BUT do no improvement! See Plan D above
1430 -- HOWEVER, some unification may take place, if we instantiate
1431 -- a method Inst with an equality constraint
1432 ; let env = mkNoImproveRedEnv doc (\_ -> ReduceMe)
1433 ; (_imp, _tybinds, _binds, constrained_dicts)
1434 <- reduceContext env wanteds_z
1436 -- Next, figure out the tyvars we will quantify over
1437 ; tau_tvs' <- zonkTcTyVarsAndFV (varSetElems tau_tvs)
1438 ; gbl_tvs' <- tcGetGlobalTyVars
1439 ; constrained_dicts' <- zonkInsts constrained_dicts
1441 ; let qtvs1 = tau_tvs' `minusVarSet` oclose (fdPredsOfInsts constrained_dicts) gbl_tvs'
1442 -- As in tcSimplifyInfer
1444 -- Do not quantify over constrained type variables:
1445 -- this is the monomorphism restriction
1446 constrained_tvs' = tyVarsOfInsts constrained_dicts'
1447 qtvs = qtvs1 `minusVarSet` constrained_tvs'
1448 pp_bndrs = pprWithCommas (quotes . ppr) bndrs
1451 ; warn_mono <- doptM Opt_WarnMonomorphism
1452 ; warnTc (warn_mono && (constrained_tvs' `intersectsVarSet` qtvs1))
1453 (vcat[ ptext (sLit "the Monomorphism Restriction applies to the binding")
1454 <> plural bndrs <+> ptext (sLit "for") <+> pp_bndrs,
1455 ptext (sLit "Consider giving a type signature for") <+> pp_bndrs])
1457 ; traceTc (text "tcSimplifyRestricted" <+> vcat [
1458 pprInsts wanteds, pprInsts constrained_dicts',
1460 ppr constrained_tvs', ppr tau_tvs', ppr qtvs ])
1462 -- The first step may have squashed more methods than
1463 -- necessary, so try again, this time more gently, knowing the exact
1464 -- set of type variables to quantify over.
1466 -- We quantify only over constraints that are captured by qtvs;
1467 -- these will just be a subset of non-dicts. This in contrast
1468 -- to normal inference (using isFreeWhenInferring) in which we quantify over
1469 -- all *non-inheritable* constraints too. This implements choice
1470 -- (B) under "implicit parameter and monomorphism" above.
1472 -- Remember that we may need to do *some* simplification, to
1473 -- (for example) squash {Monad (ST s)} into {}. It's not enough
1474 -- just to float all constraints
1476 -- At top level, we *do* squash methods because we want to
1477 -- expose implicit parameters to the test that follows
1478 ; let is_nested_group = isNotTopLevel top_lvl
1479 try_me inst | isFreeWrtTyVars qtvs inst,
1480 (is_nested_group || isDict inst) = Stop
1481 | otherwise = ReduceMe
1482 env = mkNoImproveRedEnv doc try_me
1483 ; (_imp, tybinds, binds, irreds) <- reduceContext env wanteds_z
1484 ; execTcTyVarBinds tybinds
1486 -- See "Notes on implicit parameters, Question 4: top level"
1487 ; ASSERT( all (isFreeWrtTyVars qtvs) irreds ) -- None should be captured
1488 if is_nested_group then
1490 else do { let (bad_ips, non_ips) = partition isIPDict irreds
1491 ; addTopIPErrs bndrs bad_ips
1492 ; extendLIEs non_ips }
1494 ; qtvs' <- zonkQuantifiedTyVars (varSetElems qtvs)
1495 ; return (qtvs', binds) }
1499 %************************************************************************
1503 %************************************************************************
1505 On the LHS of transformation rules we only simplify methods and constants,
1506 getting dictionaries. We want to keep all of them unsimplified, to serve
1507 as the available stuff for the RHS of the rule.
1509 Example. Consider the following left-hand side of a rule
1511 f (x == y) (y > z) = ...
1513 If we typecheck this expression we get constraints
1515 d1 :: Ord a, d2 :: Eq a
1517 We do NOT want to "simplify" to the LHS
1519 forall x::a, y::a, z::a, d1::Ord a.
1520 f ((==) (eqFromOrd d1) x y) ((>) d1 y z) = ...
1524 forall x::a, y::a, z::a, d1::Ord a, d2::Eq a.
1525 f ((==) d2 x y) ((>) d1 y z) = ...
1527 Here is another example:
1529 fromIntegral :: (Integral a, Num b) => a -> b
1530 {-# RULES "foo" fromIntegral = id :: Int -> Int #-}
1532 In the rule, a=b=Int, and Num Int is a superclass of Integral Int. But
1533 we *dont* want to get
1535 forall dIntegralInt.
1536 fromIntegral Int Int dIntegralInt (scsel dIntegralInt) = id Int
1538 because the scsel will mess up RULE matching. Instead we want
1540 forall dIntegralInt, dNumInt.
1541 fromIntegral Int Int dIntegralInt dNumInt = id Int
1545 g (x == y) (y == z) = ..
1547 where the two dictionaries are *identical*, we do NOT WANT
1549 forall x::a, y::a, z::a, d1::Eq a
1550 f ((==) d1 x y) ((>) d1 y z) = ...
1552 because that will only match if the dict args are (visibly) equal.
1553 Instead we want to quantify over the dictionaries separately.
1555 In short, tcSimplifyRuleLhs must *only* squash LitInst and MethInts, leaving
1556 all dicts unchanged, with absolutely no sharing. It's simpler to do this
1557 from scratch, rather than further parameterise simpleReduceLoop etc.
1558 Simpler, maybe, but alas not simple (see Trac #2494)
1560 * Type errors may give rise to an (unsatisfiable) equality constraint
1562 * Applications of a higher-rank function on the LHS may give
1563 rise to an implication constraint, esp if there are unsatisfiable
1564 equality constraints inside.
1567 tcSimplifyRuleLhs :: [Inst] -> TcM ([Inst], TcDictBinds)
1568 tcSimplifyRuleLhs wanteds
1569 = do { wanteds' <- zonkInsts wanteds
1571 -- Simplify equalities
1572 -- It's important to do this: Trac #3346 for example
1573 ; (_, wanteds'', tybinds, binds1) <- tcReduceEqs [] wanteds'
1574 ; execTcTyVarBinds tybinds
1576 -- Simplify other constraints
1577 ; (irreds, binds2) <- go [] emptyBag wanteds''
1579 -- Report anything that is left
1580 ; let (dicts, bad_irreds) = partition isDict irreds
1581 ; traceTc (text "tcSimplifyrulelhs" <+> pprInsts bad_irreds)
1582 ; addNoInstanceErrs (nub bad_irreds)
1583 -- The nub removes duplicates, which has
1584 -- not happened otherwise (see notes above)
1586 ; return (dicts, binds1 `unionBags` binds2) }
1588 go :: [Inst] -> TcDictBinds -> [Inst] -> TcM ([Inst], TcDictBinds)
1590 = return (irreds, binds)
1591 go irreds binds (w:ws)
1593 = go (w:irreds) binds ws
1594 | isImplicInst w -- Have a go at reducing the implication
1595 = do { (binds1, irreds1) <- reduceImplication red_env w
1596 ; let (bad_irreds, ok_irreds) = partition isImplicInst irreds1
1597 ; go (bad_irreds ++ irreds)
1598 (binds `unionBags` binds1)
1601 = do { w' <- zonkInst w -- So that (3::Int) does not generate a call
1602 -- to fromInteger; this looks fragile to me
1603 ; lookup_result <- lookupSimpleInst w'
1604 ; case lookup_result of
1605 NoInstance -> go (w:irreds) binds ws
1606 GenInst ws' rhs -> go irreds binds' (ws' ++ ws)
1608 binds' = addInstToDictBind binds w rhs
1611 -- Sigh: we need to reduce inside implications
1612 red_env = mkInferRedEnv doc try_me
1613 doc = ptext (sLit "Implication constraint in RULE lhs")
1614 try_me inst | isMethodOrLit inst = ReduceMe
1615 | otherwise = Stop -- Be gentle
1618 tcSimplifyBracket is used when simplifying the constraints arising from
1619 a Template Haskell bracket [| ... |]. We want to check that there aren't
1620 any constraints that can't be satisfied (e.g. Show Foo, where Foo has no
1621 Show instance), but we aren't otherwise interested in the results.
1622 Nor do we care about ambiguous dictionaries etc. We will type check
1623 this bracket again at its usage site.
1626 tcSimplifyBracket :: [Inst] -> TcM ()
1627 tcSimplifyBracket wanteds
1628 = do { _ <- tryHardCheckLoop doc wanteds
1631 doc = text "tcSimplifyBracket"
1635 %************************************************************************
1637 \subsection{Filtering at a dynamic binding}
1639 %************************************************************************
1644 we must discharge all the ?x constraints from B. We also do an improvement
1645 step; if we have ?x::t1 and ?x::t2 we must unify t1, t2.
1647 Actually, the constraints from B might improve the types in ?x. For example
1649 f :: (?x::Int) => Char -> Char
1652 then the constraint (?x::Int) arising from the call to f will
1653 force the binding for ?x to be of type Int.
1656 tcSimplifyIPs :: [Inst] -- The implicit parameters bound here
1659 -- We need a loop so that we do improvement, and then
1660 -- (next time round) generate a binding to connect the two
1662 -- Here the two ?x's have different types, and improvement
1663 -- makes them the same.
1665 tcSimplifyIPs given_ips wanteds
1666 = do { wanteds' <- zonkInsts wanteds
1667 ; given_ips' <- zonkInsts given_ips
1668 -- Unusually for checking, we *must* zonk the given_ips
1670 ; let env = mkRedEnv doc try_me given_ips'
1671 ; (improved, tybinds, binds, irreds) <- reduceContext env wanteds'
1672 ; execTcTyVarBinds tybinds
1674 ; if null irreds || not improved then
1675 ASSERT( all is_free irreds )
1676 do { extendLIEs irreds
1679 -- If improvement did some unification, we go round again.
1680 -- We start again with irreds, not wanteds
1681 -- Using an instance decl might have introduced a fresh type
1682 -- variable which might have been unified, so we'd get an
1683 -- infinite loop if we started again with wanteds!
1685 { binds1 <- tcSimplifyIPs given_ips' irreds
1686 ; return $ binds `unionBags` binds1
1689 doc = text "tcSimplifyIPs" <+> ppr given_ips
1690 ip_set = mkNameSet (ipNamesOfInsts given_ips)
1691 is_free inst = isFreeWrtIPs ip_set inst
1693 -- Simplify any methods that mention the implicit parameter
1694 try_me inst | is_free inst = Stop
1695 | otherwise = ReduceMe
1699 %************************************************************************
1701 \subsection[binds-for-local-funs]{@bindInstsOfLocalFuns@}
1703 %************************************************************************
1705 When doing a binding group, we may have @Insts@ of local functions.
1706 For example, we might have...
1708 let f x = x + 1 -- orig local function (overloaded)
1709 f.1 = f Int -- two instances of f
1714 The point is: we must drop the bindings for @f.1@ and @f.2@ here,
1715 where @f@ is in scope; those @Insts@ must certainly not be passed
1716 upwards towards the top-level. If the @Insts@ were binding-ified up
1717 there, they would have unresolvable references to @f@.
1719 We pass in an @init_lie@ of @Insts@ and a list of locally-bound @Ids@.
1720 For each method @Inst@ in the @init_lie@ that mentions one of the
1721 @Ids@, we create a binding. We return the remaining @Insts@ (in an
1722 @LIE@), as well as the @HsBinds@ generated.
1725 bindInstsOfLocalFuns :: [Inst] -> [TcId] -> TcM TcDictBinds
1726 -- Simlifies only MethodInsts, and generate only bindings of form
1728 -- We're careful not to even generate bindings of the form
1730 -- You'd think that'd be fine, but it interacts with what is
1731 -- arguably a bug in Match.tidyEqnInfo (see notes there)
1733 bindInstsOfLocalFuns wanteds local_ids
1734 | null overloaded_ids = do
1737 return emptyLHsBinds
1740 = do { (irreds, binds) <- gentleInferLoop doc for_me
1741 ; extendLIEs not_for_me
1745 doc = text "bindInsts" <+> ppr local_ids
1746 overloaded_ids = filter is_overloaded local_ids
1747 is_overloaded id = isOverloadedTy (idType id)
1748 (for_me, not_for_me) = partition (isMethodFor overloaded_set) wanteds
1750 overloaded_set = mkVarSet overloaded_ids -- There can occasionally be a lot of them
1751 -- so it's worth building a set, so that
1752 -- lookup (in isMethodFor) is faster
1756 %************************************************************************
1758 \subsection{Data types for the reduction mechanism}
1760 %************************************************************************
1762 The main control over context reduction is here
1766 = RedEnv { red_doc :: SDoc -- The context
1767 , red_try_me :: Inst -> WhatToDo
1768 , red_improve :: Bool -- True <=> do improvement
1769 , red_givens :: [Inst] -- All guaranteed rigid
1770 -- Always dicts & equalities
1771 -- but see Note [Rigidity]
1773 , red_stack :: (Int, [Inst]) -- Recursion stack (for err msg)
1774 -- See Note [RedStack]
1778 -- The red_givens are rigid so far as cmpInst is concerned.
1779 -- There is one case where they are not totally rigid, namely in tcSimplifyIPs
1780 -- let ?x = e in ...
1781 -- Here, the given is (?x::a), where 'a' is not necy a rigid type
1782 -- But that doesn't affect the comparison, which is based only on mame.
1785 -- The red_stack pair (n,insts) pair is just used for error reporting.
1786 -- 'n' is always the depth of the stack.
1787 -- The 'insts' is the stack of Insts being reduced: to produce X
1788 -- I had to produce Y, to produce Y I had to produce Z, and so on.
1791 mkRedEnv :: SDoc -> (Inst -> WhatToDo) -> [Inst] -> RedEnv
1792 mkRedEnv doc try_me givens
1793 = RedEnv { red_doc = doc, red_try_me = try_me,
1794 red_givens = givens,
1796 red_improve = True }
1798 mkInferRedEnv :: SDoc -> (Inst -> WhatToDo) -> RedEnv
1800 mkInferRedEnv doc try_me
1801 = RedEnv { red_doc = doc, red_try_me = try_me,
1804 red_improve = True }
1806 mkNoImproveRedEnv :: SDoc -> (Inst -> WhatToDo) -> RedEnv
1807 -- Do not do improvement; no givens
1808 mkNoImproveRedEnv doc try_me
1809 = RedEnv { red_doc = doc, red_try_me = try_me,
1812 red_improve = True }
1815 = ReduceMe -- Try to reduce this
1816 -- If there's no instance, add the inst to the
1817 -- irreductible ones, but don't produce an error
1818 -- message of any kind.
1819 -- It might be quite legitimate such as (Eq a)!
1821 | Stop -- Return as irreducible unless it can
1822 -- be reduced to a constant in one step
1823 -- Do not add superclasses; see
1825 data WantSCs = NoSCs | AddSCs -- Tells whether we should add the superclasses
1826 -- of a predicate when adding it to the avails
1827 -- The reason for this flag is entirely the super-class loop problem
1828 -- Note [SUPER-CLASS LOOP 1]
1830 zonkRedEnv :: RedEnv -> TcM RedEnv
1832 = do { givens' <- mapM zonkInst (red_givens env)
1833 ; return $ env {red_givens = givens'}
1838 %************************************************************************
1840 \subsection[reduce]{@reduce@}
1842 %************************************************************************
1844 Note [Ancestor Equalities]
1845 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1846 During context reduction, we add to the wanted equalities also those
1847 equalities that (transitively) occur in superclass contexts of wanted
1848 class constraints. Consider the following code
1850 class a ~ Int => C a
1853 If (C a) is wanted, we want to add (a ~ Int), which will be discharged by
1854 substituting Int for a. Hence, we ultimately want (C Int), which we
1855 discharge with the explicit instance.
1858 reduceContext :: RedEnv
1860 -> TcM (ImprovementDone,
1861 TcTyVarBinds, -- Type variable bindings
1862 TcDictBinds, -- Dictionary bindings
1863 [Inst]) -- Irreducible
1865 reduceContext env wanteds0
1866 = do { traceTc (text "reduceContext" <+> (vcat [
1867 text "----------------------",
1869 text "given" <+> ppr (red_givens env),
1870 text "wanted" <+> ppr wanteds0,
1871 text "----------------------"
1874 -- We want to add as wanted equalities those that (transitively)
1875 -- occur in superclass contexts of wanted class constraints.
1876 -- See Note [Ancestor Equalities]
1877 ; ancestor_eqs <- ancestorEqualities wanteds0
1878 ; traceTc $ text "reduceContext: ancestor eqs" <+> ppr ancestor_eqs
1880 -- Normalise and solve all equality constraints as far as possible
1881 -- and normalise all dictionary constraints wrt to the reduced
1882 -- equalities. The returned wanted constraints include the
1883 -- irreducible wanted equalities.
1884 ; let wanteds = wanteds0 ++ ancestor_eqs
1885 givens = red_givens env
1889 normalise_binds) <- tcReduceEqs givens wanteds
1890 ; traceTc $ text "reduceContext: tcReduceEqs result" <+> vcat
1891 [ppr givens', ppr wanteds', ppr tybinds,
1892 ppr normalise_binds]
1894 -- Build the Avail mapping from "given_dicts"
1895 ; (init_state, _) <- getConstraints $ do
1896 { init_state <- foldlM addGiven emptyAvails givens'
1900 -- Solve the *wanted* *dictionary* constraints (not implications)
1901 -- This may expose some further equational constraints in the course
1902 -- of improvement due to functional dependencies if any of the
1903 -- involved unifications gets deferred.
1904 ; let (wanted_implics, wanted_dicts) = partition isImplicInst wanteds'
1905 ; (avails, extra_eqs) <- getConstraints (reduceList env wanted_dicts init_state)
1906 -- The getConstraints is reqd because reduceList does improvement
1907 -- (via extendAvails) which may in turn do unification
1910 dict_irreds) <- extractResults avails wanted_dicts
1911 ; traceTc $ text "reduceContext: extractResults" <+> vcat
1912 [ppr avails, ppr wanted_dicts, ppr dict_binds]
1914 -- Solve the wanted *implications*. In doing so, we can provide
1915 -- as "given" all the dicts that were originally given,
1916 -- *or* for which we now have bindings,
1917 -- *or* which are now irreds
1918 -- NB: Equality irreds need to be converted, as the recursive
1919 -- invocation of the solver will still treat them as wanteds
1921 ; let implic_env = env { red_givens
1922 = givens ++ bound_dicts ++
1923 map wantedToLocalEqInst dict_irreds }
1924 ; (implic_binds_s, implic_irreds_s)
1925 <- mapAndUnzipM (reduceImplication implic_env) wanted_implics
1926 ; let implic_binds = unionManyBags implic_binds_s
1927 implic_irreds = concat implic_irreds_s
1929 -- Collect all irreducible instances, and determine whether we should
1930 -- go round again. We do so in either of two cases:
1931 -- (1) If dictionary reduction or equality solving led to
1932 -- improvement (i.e., bindings for type variables).
1933 -- (2) If we reduced dictionaries (i.e., got dictionary bindings),
1934 -- they may have exposed further opportunities to normalise
1935 -- family applications. See Note [Dictionary Improvement]
1937 -- NB: We do *not* go around for new extra_eqs. Morally, we should,
1938 -- but we can't without risking non-termination (see #2688). By
1939 -- not going around, we miss some legal programs mixing FDs and
1940 -- TFs, but we never claimed to support such programs in the
1941 -- current implementation anyway.
1943 ; let all_irreds = dict_irreds ++ implic_irreds ++ extra_eqs
1944 avails_improved = availsImproved avails
1945 eq_improved = anyBag (not . isCoVarBind) tybinds
1946 improvedFlexible = avails_improved || eq_improved
1947 reduced_dicts = not (isEmptyBag dict_binds)
1948 improved = improvedFlexible || reduced_dicts
1950 improvedHint = (if avails_improved then " [AVAILS]" else "") ++
1951 (if eq_improved then " [EQ]" else "")
1953 ; traceTc (text "reduceContext end" <+> (vcat [
1954 text "----------------------",
1956 text "given" <+> ppr givens,
1957 text "wanted" <+> ppr wanteds0,
1959 text "tybinds" <+> ppr tybinds,
1960 text "avails" <+> pprAvails avails,
1961 text "improved =" <+> ppr improved <+> text improvedHint,
1962 text "(all) irreds = " <+> ppr all_irreds,
1963 text "dict-binds = " <+> ppr dict_binds,
1964 text "implic-binds = " <+> ppr implic_binds,
1965 text "----------------------"
1970 normalise_binds `unionBags` dict_binds
1971 `unionBags` implic_binds,
1975 isCoVarBind (TcTyVarBind tv _) = isCoVar tv
1977 tcImproveOne :: Avails -> Inst -> TcM ImprovementDone
1978 tcImproveOne avails inst
1979 | not (isDict inst) = return False
1981 = do { inst_envs <- tcGetInstEnvs
1982 ; let eqns = improveOne (classInstances inst_envs)
1983 (dictPred inst, pprInstArising inst)
1984 [ (dictPred p, pprInstArising p)
1985 | p <- availsInsts avails, isDict p ]
1986 -- Avails has all the superclasses etc (good)
1987 -- It also has all the intermediates of the deduction (good)
1988 -- It does not have duplicates (good)
1989 -- NB that (?x::t1) and (?x::t2) will be held separately in
1990 -- avails so that improve will see them separate
1991 ; traceTc (text "improveOne" <+> ppr inst)
1994 unifyEqns :: [(Equation, (PredType, SDoc), (PredType, SDoc))]
1995 -> TcM ImprovementDone
1996 unifyEqns [] = return False
1998 = do { traceTc (ptext (sLit "Improve:") <+> vcat (map pprEquationDoc eqns))
1999 ; improved <- mapM unify eqns
2000 ; return $ or improved
2003 unify ((qtvs, pairs), what1, what2)
2004 = addErrCtxtM (mkEqnMsg what1 what2) $
2005 do { let freeTyVars = unionVarSets (map tvs_pr pairs)
2007 ; (_, _, tenv) <- tcInstTyVars (varSetElems qtvs)
2008 ; mapM_ (unif_pr tenv) pairs
2009 ; anyM isFilledMetaTyVar $ varSetElems freeTyVars
2012 unif_pr tenv (ty1, ty2) = unifyType (substTy tenv ty1) (substTy tenv ty2)
2014 tvs_pr (ty1, ty2) = tyVarsOfType ty1 `unionVarSet` tyVarsOfType ty2
2016 pprEquationDoc :: (Equation, (PredType, SDoc), (PredType, SDoc)) -> SDoc
2017 pprEquationDoc (eqn, (p1, _), (p2, _))
2018 = vcat [pprEquation eqn, nest 2 (ppr p1), nest 2 (ppr p2)]
2020 mkEqnMsg :: (TcPredType, SDoc) -> (TcPredType, SDoc) -> TidyEnv
2021 -> TcM (TidyEnv, SDoc)
2022 mkEqnMsg (pred1,from1) (pred2,from2) tidy_env
2023 = do { pred1' <- zonkTcPredType pred1
2024 ; pred2' <- zonkTcPredType pred2
2025 ; let { pred1'' = tidyPred tidy_env pred1'
2026 ; pred2'' = tidyPred tidy_env pred2' }
2027 ; let msg = vcat [ptext (sLit "When using functional dependencies to combine"),
2028 nest 2 (sep [ppr pred1'' <> comma, nest 2 from1]),
2029 nest 2 (sep [ppr pred2'' <> comma, nest 2 from2])]
2030 ; return (tidy_env, msg) }
2033 Note [Dictionary Improvement]
2034 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2035 In reduceContext, we first reduce equalities and then class constraints.
2036 However, the letter may expose further opportunities for the former. Hence,
2037 we need to go around again if dictionary reduction produced any dictionary
2038 bindings. The following example demonstrated the point:
2040 data EX _x _y (p :: * -> *)
2045 class Base (Def p) => Prop p where
2049 instance Prop () where
2052 instance (Base (Def (p ANY))) => Base (EX _x _y p)
2053 instance (Prop (p ANY)) => Prop (EX _x _y p) where
2054 type Def (EX _x _y p) = EX _x _y p
2057 instance Prop (FOO x) where
2058 type Def (FOO x) = ()
2061 instance Prop BAR where
2062 type Def BAR = EX () () FOO
2064 During checking the last instance declaration, we need to check the superclass
2065 cosntraint Base (Def BAR), which family normalisation reduced to
2066 Base (EX () () FOO). Chasing the instance for Base (EX _x _y p), gives us
2067 Base (Def (FOO ANY)), which again requires family normalisation of Def to
2068 Base () before we can finish.
2071 The main context-reduction function is @reduce@. Here's its game plan.
2074 reduceList :: RedEnv -> [Inst] -> Avails -> TcM Avails
2075 reduceList env@(RedEnv {red_stack = (n,stk)}) wanteds state
2076 = do { traceTc (text "reduceList " <+> (ppr wanteds $$ ppr state))
2078 ; when (debugIsOn && (n > 8)) $ do
2079 debugDumpTcRn (hang (ptext (sLit "Interesting! Context reduction stack depth") <+> int n)
2080 2 (ifPprDebug (nest 2 (pprStack stk))))
2081 ; if n >= ctxtStkDepth dopts then
2082 failWithTc (reduceDepthErr n stk)
2086 go [] state = return state
2087 go (w:ws) state = do { state' <- reduce (env {red_stack = (n+1, w:stk)}) w state
2090 -- Base case: we're done!
2091 reduce :: RedEnv -> Inst -> Avails -> TcM Avails
2092 reduce env wanted avails
2094 -- We don't reduce equalities here (and they must not end up as irreds
2099 -- It's the same as an existing inst, or a superclass thereof
2100 | Just _ <- findAvail avails wanted
2101 = do { traceTc (text "reduce: found " <+> ppr wanted)
2106 = do { traceTc (text "reduce" <+> ppr wanted $$ ppr avails)
2107 ; case red_try_me env wanted of {
2108 Stop -> try_simple (addIrred NoSCs);
2109 -- See Note [No superclasses for Stop]
2111 ReduceMe -> do -- It should be reduced
2112 { (avails, lookup_result) <- reduceInst env avails wanted
2113 ; case lookup_result of
2114 NoInstance -> addIrred AddSCs avails wanted
2115 -- Add it and its superclasses
2117 GenInst [] rhs -> addWanted AddSCs avails wanted rhs []
2119 GenInst wanteds' rhs
2120 -> do { avails1 <- addIrred NoSCs avails wanted
2121 ; avails2 <- reduceList env wanteds' avails1
2122 ; addWanted AddSCs avails2 wanted rhs wanteds' } }
2123 -- Temporarily do addIrred *before* the reduceList,
2124 -- which has the effect of adding the thing we are trying
2125 -- to prove to the database before trying to prove the things it
2126 -- needs. See note [RECURSIVE DICTIONARIES]
2127 -- NB: we must not do an addWanted before, because that adds the
2128 -- superclasses too, and that can lead to a spurious loop; see
2129 -- the examples in [SUPERCLASS-LOOP]
2130 -- So we do an addIrred before, and then overwrite it afterwards with addWanted
2133 -- First, see if the inst can be reduced to a constant in one step
2134 -- Works well for literals (1::Int) and constant dictionaries (d::Num Int)
2135 -- Don't bother for implication constraints, which take real work
2136 try_simple do_this_otherwise
2137 = do { res <- lookupSimpleInst wanted
2139 GenInst [] rhs -> addWanted AddSCs avails wanted rhs []
2140 _ -> do_this_otherwise avails wanted }
2144 Note [RECURSIVE DICTIONARIES]
2145 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2147 data D r = ZeroD | SuccD (r (D r));
2149 instance (Eq (r (D r))) => Eq (D r) where
2150 ZeroD == ZeroD = True
2151 (SuccD a) == (SuccD b) = a == b
2154 equalDC :: D [] -> D [] -> Bool;
2157 We need to prove (Eq (D [])). Here's how we go:
2161 by instance decl, holds if
2165 by instance decl of Eq, holds if
2167 where d2 = dfEqList d3
2170 But now we can "tie the knot" to give
2176 and it'll even run! The trick is to put the thing we are trying to prove
2177 (in this case Eq (D []) into the database before trying to prove its
2178 contributing clauses.
2180 Note [SUPERCLASS-LOOP 2]
2181 ~~~~~~~~~~~~~~~~~~~~~~~~
2182 We need to be careful when adding "the constaint we are trying to prove".
2183 Suppose we are *given* d1:Ord a, and want to deduce (d2:C [a]) where
2185 class Ord a => C a where
2186 instance Ord [a] => C [a] where ...
2188 Then we'll use the instance decl to deduce C [a] from Ord [a], and then add the
2189 superclasses of C [a] to avails. But we must not overwrite the binding
2190 for Ord [a] (which is obtained from Ord a) with a superclass selection or we'll just
2193 Here's another variant, immortalised in tcrun020
2194 class Monad m => C1 m
2195 class C1 m => C2 m x
2196 instance C2 Maybe Bool
2197 For the instance decl we need to build (C1 Maybe), and it's no good if
2198 we run around and add (C2 Maybe Bool) and its superclasses to the avails
2199 before we search for C1 Maybe.
2201 Here's another example
2202 class Eq b => Foo a b
2203 instance Eq a => Foo [a] a
2207 we'll first deduce that it holds (via the instance decl). We must not
2208 then overwrite the Eq t constraint with a superclass selection!
2210 At first I had a gross hack, whereby I simply did not add superclass constraints
2211 in addWanted, though I did for addGiven and addIrred. This was sub-optimal,
2212 becuase it lost legitimate superclass sharing, and it still didn't do the job:
2213 I found a very obscure program (now tcrun021) in which improvement meant the
2214 simplifier got two bites a the cherry... so something seemed to be an Stop
2215 first time, but reducible next time.
2217 Now we implement the Right Solution, which is to check for loops directly
2218 when adding superclasses. It's a bit like the occurs check in unification.
2222 %************************************************************************
2224 Reducing a single constraint
2226 %************************************************************************
2229 ---------------------------------------------
2230 reduceInst :: RedEnv -> Avails -> Inst -> TcM (Avails, LookupInstResult)
2231 reduceInst _ avails other_inst
2232 = do { result <- lookupSimpleInst other_inst
2233 ; return (avails, result) }
2236 Note [Equational Constraints in Implication Constraints]
2237 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2239 An implication constraint is of the form
2241 where Given and Wanted may contain both equational and dictionary
2242 constraints. The delay and reduction of these two kinds of constraints
2245 -) In the generated code, wanted Dictionary constraints are wrapped up in an
2246 implication constraint that is created at the code site where the wanted
2247 dictionaries can be reduced via a let-binding. This let-bound implication
2248 constraint is deconstructed at the use-site of the wanted dictionaries.
2250 -) While the reduction of equational constraints is also delayed, the delay
2251 is not manifest in the generated code. The required evidence is generated
2252 in the code directly at the use-site. There is no let-binding and deconstruction
2253 necessary. The main disadvantage is that we cannot exploit sharing as the
2254 same evidence may be generated at multiple use-sites. However, this disadvantage
2255 is limited because it only concerns coercions which are erased.
2257 The different treatment is motivated by the different in representation. Dictionary
2258 constraints require manifest runtime dictionaries, while equations require coercions
2262 ---------------------------------------------
2263 reduceImplication :: RedEnv
2265 -> TcM (TcDictBinds, [Inst])
2268 Suppose we are simplifying the constraint
2269 forall bs. extras => wanted
2270 in the context of an overall simplification problem with givens 'givens'.
2273 * The 'givens' need not mention any of the quantified type variables
2274 e.g. forall {}. Eq a => Eq [a]
2275 forall {}. C Int => D (Tree Int)
2277 This happens when you have something like
2279 T1 :: Eq a => a -> T a
2282 f x = ...(case x of { T1 v -> v==v })...
2285 -- ToDo: should we instantiate tvs? I think it's not necessary
2287 -- Note on coercion variables:
2289 -- The extra given coercion variables are bound at two different
2292 -- -) in the creation context of the implication constraint
2293 -- the solved equational constraints use these binders
2295 -- -) at the solving site of the implication constraint
2296 -- the solved dictionaries use these binders;
2297 -- these binders are generated by reduceImplication
2299 -- Note [Binders for equalities]
2300 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2301 -- To reuse the binders of local/given equalities in the binders of
2302 -- implication constraints, it is crucial that these given equalities
2303 -- always have the form
2305 -- where cotv is a simple coercion type variable (and not a more
2306 -- complex coercion term). We require that the extra_givens always
2307 -- have this form and exploit the special form when generating binders.
2308 reduceImplication env
2309 orig_implic@(ImplicInst { tci_name = name, tci_loc = inst_loc,
2311 tci_given = extra_givens, tci_wanted = wanteds
2313 = do { -- Solve the sub-problem
2314 ; let try_me _ = ReduceMe -- Note [Freeness and implications]
2315 env' = env { red_givens = extra_givens ++ red_givens env
2316 , red_doc = sep [ptext (sLit "reduceImplication for")
2318 nest 2 (parens $ ptext (sLit "within")
2320 , red_try_me = try_me }
2322 ; traceTc (text "reduceImplication" <+> vcat
2323 [ ppr (red_givens env), ppr extra_givens,
2325 ; (irreds, binds) <- checkLoop env' wanteds
2327 ; traceTc (text "reduceImplication result" <+> vcat
2328 [ppr irreds, ppr binds])
2330 ; -- extract superclass binds
2331 -- (sc_binds,_) <- extractResults avails []
2332 -- ; traceTc (text "reduceImplication sc_binds" <+> vcat
2333 -- [ppr sc_binds, ppr avails])
2336 -- SLPJ Sept 07: what if improvement happened inside the checkLoop?
2337 -- Then we must iterate the outer loop too!
2339 ; didntSolveWantedEqs <- allM wantedEqInstIsUnsolved wanteds
2340 -- we solve wanted eqs by side effect!
2342 -- Progress is no longer measered by the number of bindings
2343 -- If there are any irreds, but no bindings and no solved
2344 -- equalities, we back off and do nothing
2345 ; let backOff = isEmptyLHsBinds binds && -- no new bindings
2346 (not $ null irreds) && -- but still some irreds
2347 didntSolveWantedEqs -- no instantiated cotv
2349 ; if backOff then -- No progress
2350 return (emptyBag, [orig_implic])
2352 { (simpler_implic_insts, bind)
2353 <- makeImplicationBind inst_loc tvs extra_givens irreds
2354 -- This binding is useless if the recursive simplification
2355 -- made no progress; but currently we don't try to optimise that
2356 -- case. After all, we only try hard to reduce at top level, or
2357 -- when inferring types.
2359 ; let -- extract Id binders for dicts and CoTyVar binders for eqs;
2360 -- see Note [Binders for equalities]
2361 (extra_eq_givens, extra_dict_givens) = partition isEqInst
2363 eq_cotvs = map instToVar extra_eq_givens
2364 dict_ids = map instToId extra_dict_givens
2367 <.> mkWpTyLams eq_cotvs
2368 <.> mkWpLams dict_ids
2369 <.> WpLet (binds `unionBags` bind)
2370 rhs = mkLHsWrap co payload
2371 loc = instLocSpan inst_loc
2372 -- wanted equalities are solved by updating their
2373 -- cotv; we don't generate bindings for them
2374 dict_bndrs = map (L loc . HsVar . instToId)
2375 . filter (not . isEqInst)
2377 payload = mkBigLHsTup dict_bndrs
2379 ; traceTc (vcat [text "reduceImplication" <+> ppr name,
2380 ppr simpler_implic_insts,
2381 text "->" <+> ppr rhs])
2382 ; return (unitBag (L loc (VarBind { var_id= instToId orig_implic
2384 , var_inline = notNull dict_ids }
2385 -- See Note [Always inline implication constraints]
2387 simpler_implic_insts)
2390 reduceImplication _ i = pprPanic "reduceImplication" (ppr i)
2393 Note [Always inline implication constraints]
2394 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2395 Suppose an implication constraint floats out of an INLINE function.
2396 Then although the implication has a single call site, it won't be
2397 inlined. And that is bad because it means that even if there is really
2398 *no* overloading (type signatures specify the exact types) there will
2399 still be dictionary passing in the resulting code. To avert this,
2400 we mark the implication constraints themselves as INLINE, at least when
2401 there is no loss of sharing as a result.
2403 Note [Freeness and implications]
2404 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2405 It's hard to say when an implication constraint can be floated out. Consider
2406 forall {} Eq a => Foo [a]
2407 The (Foo [a]) doesn't mention any of the quantified variables, but it
2408 still might be partially satisfied by the (Eq a).
2410 There is a useful special case when it *is* easy to partition the
2411 constraints, namely when there are no 'givens'. Consider
2412 forall {a}. () => Bar b
2413 There are no 'givens', and so there is no reason to capture (Bar b).
2414 We can let it float out. But if there is even one constraint we
2415 must be much more careful:
2416 forall {a}. C a b => Bar (m b)
2417 because (C a b) might have a superclass (D b), from which we might
2418 deduce (Bar [b]) when m later gets instantiated to []. Ha!
2420 Here is an even more exotic example
2422 Now consider the constraint
2423 forall b. D Int b => C Int
2424 We can satisfy the (C Int) from the superclass of D, so we don't want
2425 to float the (C Int) out, even though it mentions no type variable in
2428 One more example: the constraint
2430 instance (C a, E c) => E (a,c)
2432 constraint: forall b. D Int b => E (Int,c)
2434 You might think that the (D Int b) can't possibly contribute
2435 to solving (E (Int,c)), since the latter mentions 'c'. But
2436 in fact it can, because solving the (E (Int,c)) constraint needs
2439 and the (C Int) can be satisfied from the superclass of (D Int b).
2440 So we must still not float (E (Int,c)) out.
2442 To think about: special cases for unary type classes?
2444 Note [Pruning the givens in an implication constraint]
2445 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2446 Suppose we are about to form the implication constraint
2447 forall tvs. Eq a => Ord b
2448 The (Eq a) cannot contribute to the (Ord b), because it has no access to
2449 the type variable 'b'. So we could filter out the (Eq a) from the givens.
2450 But BE CAREFUL of the examples above in [Freeness and implications].
2452 Doing so would be a bit tidier, but all the implication constraints get
2453 simplified away by the optimiser, so it's no great win. So I don't take
2454 advantage of that at the moment.
2456 If you do, BE CAREFUL of wobbly type variables.
2459 %************************************************************************
2461 Avails and AvailHow: the pool of evidence
2463 %************************************************************************
2467 data Avails = Avails !ImprovementDone !AvailEnv
2469 type ImprovementDone = Bool -- True <=> some unification has happened
2470 -- so some Irreds might now be reducible
2471 -- keys that are now
2473 type AvailEnv = FiniteMap Inst AvailHow
2475 = IsIrred -- Used for irreducible dictionaries,
2476 -- which are going to be lambda bound
2478 | Given Inst -- Used for dictionaries for which we have a binding
2479 -- e.g. those "given" in a signature
2481 | Rhs -- Used when there is a RHS
2482 (LHsExpr TcId) -- The RHS
2483 [Inst] -- Insts free in the RHS; we need these too
2485 instance Outputable Avails where
2488 pprAvails :: Avails -> SDoc
2489 pprAvails (Avails imp avails)
2490 = vcat [ ptext (sLit "Avails") <> (if imp then ptext (sLit "[improved]") else empty)
2492 vcat [ sep [ppr inst, nest 2 (equals <+> ppr avail)]
2493 | (inst,avail) <- Map.toList avails ]]
2495 instance Outputable AvailHow where
2498 -------------------------
2499 pprAvail :: AvailHow -> SDoc
2500 pprAvail IsIrred = text "Irred"
2501 pprAvail (Given x) = text "Given" <+> ppr x
2502 pprAvail (Rhs rhs bs) = sep [text "Rhs" <+> ppr bs,
2505 -------------------------
2506 extendAvailEnv :: AvailEnv -> Inst -> AvailHow -> AvailEnv
2507 extendAvailEnv env inst avail = Map.insert inst avail env
2509 findAvailEnv :: AvailEnv -> Inst -> Maybe AvailHow
2510 findAvailEnv env wanted = Map.lookup wanted env
2511 -- NB 1: the Ord instance of Inst compares by the class/type info
2512 -- *not* by unique. So
2513 -- d1::C Int == d2::C Int
2515 emptyAvails :: Avails
2516 emptyAvails = Avails False emptyFM
2518 findAvail :: Avails -> Inst -> Maybe AvailHow
2519 findAvail (Avails _ avails) wanted = findAvailEnv avails wanted
2521 elemAvails :: Inst -> Avails -> Bool
2522 elemAvails wanted (Avails _ avails) = wanted `elemFM` avails
2524 extendAvails :: Avails -> Inst -> AvailHow -> TcM Avails
2526 extendAvails avails@(Avails imp env) inst avail
2527 = do { imp1 <- tcImproveOne avails inst -- Do any improvement
2528 ; return (Avails (imp || imp1) (extendAvailEnv env inst avail)) }
2530 availsInsts :: Avails -> [Inst]
2531 availsInsts (Avails _ avails) = Map.keys avails
2533 availsImproved :: Avails -> ImprovementDone
2534 availsImproved (Avails imp _) = imp
2537 Extracting the bindings from a bunch of Avails.
2538 The bindings do *not* come back sorted in dependency order.
2539 We assume that they'll be wrapped in a big Rec, so that the
2540 dependency analyser can sort them out later
2543 type DoneEnv = FiniteMap Inst [Id]
2544 -- Tracks which things we have evidence for
2546 extractResults :: Avails
2548 -> TcM (TcDictBinds, -- Bindings
2549 [Inst], -- The insts bound by the bindings
2550 [Inst]) -- Irreducible ones
2551 -- Note [Reducing implication constraints]
2553 extractResults (Avails _ avails) wanteds
2554 = go emptyBag [] [] emptyFM wanteds
2556 go :: TcDictBinds -- Bindings for dicts
2557 -> [Inst] -- Bound by the bindings
2559 -> DoneEnv -- Has an entry for each inst in the above three sets
2561 -> TcM (TcDictBinds, [Inst], [Inst])
2562 go binds bound_dicts irreds _ []
2563 = return (binds, bound_dicts, irreds)
2565 go binds bound_dicts irreds done (w:ws)
2567 = go binds bound_dicts (w:irreds) done' ws
2569 | Just done_ids@(done_id : rest_done_ids) <- Map.lookup w done
2570 = if w_id `elem` done_ids then
2571 go binds bound_dicts irreds done ws
2573 go (add_bind (nlHsVar done_id)) bound_dicts irreds
2574 (Map.insert w (done_id : w_id : rest_done_ids) done) ws
2576 | otherwise -- Not yet done
2577 = case findAvailEnv avails w of
2578 Nothing -> pprTrace "Urk: extractResults" (ppr w) $
2579 go binds bound_dicts irreds done ws
2581 Just IsIrred -> go binds bound_dicts (w:irreds) done' ws
2583 Just (Rhs rhs ws') -> go (add_bind rhs) (w:bound_dicts) irreds done' (ws' ++ ws)
2585 Just (Given g) -> go binds' bound_dicts irreds (Map.insert w [g_id] done) ws
2588 binds' | w_id == g_id = binds
2589 | otherwise = add_bind (nlHsVar g_id)
2592 done' = Map.insert w [w_id] done
2593 add_bind rhs = addInstToDictBind binds w rhs
2597 Note [No superclasses for Stop]
2598 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2599 When we decide not to reduce an Inst -- the 'WhatToDo' --- we still
2600 add it to avails, so that any other equal Insts will be commoned up
2601 right here. However, we do *not* add superclasses. If we have
2604 but a is not bound here, then we *don't* want to derive dn from df
2605 here lest we lose sharing.
2608 addWanted :: WantSCs -> Avails -> Inst -> LHsExpr TcId -> [Inst] -> TcM Avails
2609 addWanted want_scs avails wanted rhs_expr wanteds
2610 = addAvailAndSCs want_scs avails wanted avail
2612 avail = Rhs rhs_expr wanteds
2614 addGiven :: Avails -> Inst -> TcM Avails
2615 addGiven avails given
2616 = addAvailAndSCs want_scs avails given (Given given)
2618 want_scs = case instLocOrigin (instLoc given) of
2621 -- Conditionally add superclasses for 'given'
2622 -- See Note [Recursive instances and superclases]
2624 -- No ASSERT( not (given `elemAvails` avails) ) because in an
2625 -- instance decl for Ord t we can add both Ord t and Eq t as
2626 -- 'givens', so the assert isn't true
2630 addIrred :: WantSCs -> Avails -> Inst -> TcM Avails
2631 addIrred want_scs avails irred = ASSERT2( not (irred `elemAvails` avails), ppr irred $$ ppr avails )
2632 addAvailAndSCs want_scs avails irred IsIrred
2634 addAvailAndSCs :: WantSCs -> Avails -> Inst -> AvailHow -> TcM Avails
2635 addAvailAndSCs want_scs avails inst avail
2636 | not (isClassDict inst) = extendAvails avails inst avail
2637 | NoSCs <- want_scs = extendAvails avails inst avail
2638 | otherwise = do { traceTc (text "addAvailAndSCs" <+> vcat [ppr inst, ppr deps])
2639 ; avails' <- extendAvails avails inst avail
2640 ; addSCs is_loop avails' inst }
2642 is_loop pred = any (`tcEqType` mkPredTy pred) dep_tys
2643 -- Note: this compares by *type*, not by Unique
2644 deps = findAllDeps (unitVarSet (instToVar inst)) avail
2645 dep_tys = map idType (varSetElems deps)
2647 findAllDeps :: IdSet -> AvailHow -> IdSet
2648 -- Find all the Insts that this one depends on
2649 -- See Note [SUPERCLASS-LOOP 2]
2650 -- Watch out, though. Since the avails may contain loops
2651 -- (see Note [RECURSIVE DICTIONARIES]), so we need to track the ones we've seen so far
2652 findAllDeps so_far (Rhs _ kids) = foldl find_all so_far kids
2653 findAllDeps so_far _ = so_far
2655 find_all :: IdSet -> Inst -> IdSet
2657 | isEqInst kid = so_far
2658 | kid_id `elemVarSet` so_far = so_far
2659 | Just avail <- findAvail avails kid = findAllDeps so_far' avail
2660 | otherwise = so_far'
2662 so_far' = extendVarSet so_far kid_id -- Add the new kid to so_far
2663 kid_id = instToId kid
2665 addSCs :: (TcPredType -> Bool) -> Avails -> Inst -> TcM Avails
2666 -- Add all the superclasses of the Inst to Avails
2667 -- The first param says "don't do this because the original thing
2668 -- depends on this one, so you'd build a loop"
2669 -- Invariant: the Inst is already in Avails.
2671 addSCs is_loop avails dict
2672 = ASSERT( isDict dict )
2673 do { sc_dicts <- newCtGivens (instLoc dict) sc_theta'
2674 ; foldlM add_sc avails (zipEqual "add_scs" sc_dicts sc_sels) }
2676 (clas, tys) = getDictClassTys dict
2677 (tyvars, sc_theta, sc_sels, _) = classBigSig clas
2678 sc_theta' = filter (not . isEqPred) $
2679 substTheta (zipTopTvSubst tyvars tys) sc_theta
2681 add_sc avails (sc_dict, sc_sel)
2682 | is_loop (dictPred sc_dict) = return avails -- See Note [SUPERCLASS-LOOP 2]
2683 | is_given sc_dict = return avails
2684 | otherwise = do { avails' <- extendAvails avails sc_dict (Rhs sc_sel_rhs [dict])
2685 ; addSCs is_loop avails' sc_dict }
2687 sc_sel_rhs = L (instSpan dict) (HsWrap co_fn (HsVar sc_sel))
2688 co_fn = WpApp (instToVar dict) <.> mkWpTyApps tys
2690 is_given :: Inst -> Bool
2691 is_given sc_dict = case findAvail avails sc_dict of
2692 Just (Given _) -> True -- Given is cheaper than superclass selection
2695 -- From the a set of insts obtain all equalities that (transitively) occur in
2696 -- superclass contexts of class constraints (aka the ancestor equalities).
2698 ancestorEqualities :: [Inst] -> TcM [Inst]
2700 = mapM mkWantedEqInst -- turn only equality predicates..
2701 . filter isEqPred -- ..into wanted equality insts
2703 . addAEsToBag emptyBag -- collect the superclass constraints..
2704 . map dictPred -- ..of all predicates in a bag
2705 . filter isClassDict
2707 addAEsToBag :: Bag PredType -> [PredType] -> Bag PredType
2708 addAEsToBag bag [] = bag
2709 addAEsToBag bag (pred:preds)
2710 | pred `elemBag` bag = addAEsToBag bag preds
2711 | isEqPred pred = addAEsToBag bagWithPred preds
2712 | isClassPred pred = addAEsToBag bagWithPred predsWithSCs
2713 | otherwise = addAEsToBag bag preds
2715 bagWithPred = bag `snocBag` pred
2716 predsWithSCs = preds ++ substTheta (zipTopTvSubst tyvars tys) sc_theta
2718 (tyvars, sc_theta, _, _) = classBigSig clas
2719 (clas, tys) = getClassPredTys pred
2723 %************************************************************************
2725 \section{tcSimplifyTop: defaulting}
2727 %************************************************************************
2730 @tcSimplifyTop@ is called once per module to simplify all the constant
2731 and ambiguous Insts.
2733 We need to be careful of one case. Suppose we have
2735 instance Num a => Num (Foo a b) where ...
2737 and @tcSimplifyTop@ is given a constraint (Num (Foo x y)). Then it'll simplify
2738 to (Num x), and default x to Int. But what about y??
2740 It's OK: the final zonking stage should zap y to (), which is fine.
2744 tcSimplifyTop, tcSimplifyInteractive :: [Inst] -> TcM TcDictBinds
2745 tcSimplifyTop wanteds
2746 = tc_simplify_top doc False wanteds
2748 doc = text "tcSimplifyTop"
2750 tcSimplifyInteractive wanteds
2751 = tc_simplify_top doc True wanteds
2753 doc = text "tcSimplifyInteractive"
2755 -- The TcLclEnv should be valid here, solely to improve
2756 -- error message generation for the monomorphism restriction
2757 tc_simplify_top :: SDoc -> Bool -> [Inst] -> TcM (Bag (LHsBind TcId))
2758 tc_simplify_top doc interactive wanteds
2759 = do { dflags <- getDOpts
2760 ; wanteds <- zonkInsts wanteds
2761 ; mapM_ zonkTopTyVar (varSetElems (tyVarsOfInsts wanteds))
2763 ; traceTc (text "tc_simplify_top 0: " <+> ppr wanteds)
2764 ; (irreds1, binds1) <- tryHardCheckLoop doc1 wanteds
2765 -- ; (irreds1, binds1) <- gentleInferLoop doc1 wanteds
2766 ; traceTc (text "tc_simplify_top 1: " <+> ppr irreds1)
2767 ; (irreds2, binds2) <- approximateImplications doc2 (\_ -> True) irreds1
2768 ; traceTc (text "tc_simplify_top 2: " <+> ppr irreds2)
2770 -- Use the defaulting rules to do extra unification
2771 -- NB: irreds2 are already zonked
2772 ; (irreds3, binds3) <- disambiguate doc3 interactive dflags irreds2
2774 -- Deal with implicit parameters
2775 ; let (bad_ips, non_ips) = partition isIPDict irreds3
2776 (ambigs, others) = partition isTyVarDict non_ips
2778 ; topIPErrs bad_ips -- Can arise from f :: Int -> Int
2780 ; addNoInstanceErrs others
2781 ; addTopAmbigErrs ambigs
2783 ; return (binds1 `unionBags` binds2 `unionBags` binds3) }
2785 doc1 = doc <+> ptext (sLit "(first round)")
2786 doc2 = doc <+> ptext (sLit "(approximate)")
2787 doc3 = doc <+> ptext (sLit "(disambiguate)")
2790 If a dictionary constrains a type variable which is
2791 * not mentioned in the environment
2792 * and not mentioned in the type of the expression
2793 then it is ambiguous. No further information will arise to instantiate
2794 the type variable; nor will it be generalised and turned into an extra
2795 parameter to a function.
2797 It is an error for this to occur, except that Haskell provided for
2798 certain rules to be applied in the special case of numeric types.
2800 * at least one of its classes is a numeric class, and
2801 * all of its classes are numeric or standard
2802 then the type variable can be defaulted to the first type in the
2803 default-type list which is an instance of all the offending classes.
2805 So here is the function which does the work. It takes the ambiguous
2806 dictionaries and either resolves them (producing bindings) or
2807 complains. It works by splitting the dictionary list by type
2808 variable, and using @disambigOne@ to do the real business.
2810 @disambigOne@ assumes that its arguments dictionaries constrain all
2811 the same type variable.
2813 ADR Comment 20/6/94: I've changed the @CReturnable@ case to default to
2814 @()@ instead of @Int@. I reckon this is the Right Thing to do since
2815 the most common use of defaulting is code like:
2817 _ccall_ foo `seqPrimIO` bar
2819 Since we're not using the result of @foo@, the result if (presumably)
2823 disambiguate :: SDoc -> Bool -> DynFlags -> [Inst] -> TcM ([Inst], TcDictBinds)
2824 -- Just does unification to fix the default types
2825 -- The Insts are assumed to be pre-zonked
2826 disambiguate doc interactive dflags insts
2828 = return (insts, emptyBag)
2830 | null defaultable_groups
2831 = do { traceTc (text "disambigutate, no defaultable groups" <+> vcat [ppr unaries, ppr insts, ppr bad_tvs, ppr defaultable_groups])
2832 ; return (insts, emptyBag) }
2835 = do { -- Figure out what default types to use
2836 default_tys <- getDefaultTys extended_defaulting ovl_strings
2838 ; traceTc (text "disambiguate1" <+> vcat [ppr insts, ppr unaries, ppr bad_tvs, ppr defaultable_groups])
2839 ; mapM_ (disambigGroup default_tys) defaultable_groups
2841 -- disambigGroup does unification, hence try again
2842 ; tryHardCheckLoop doc insts }
2845 extended_defaulting = interactive || dopt Opt_ExtendedDefaultRules dflags
2846 -- See also Trac #1974
2847 ovl_strings = dopt Opt_OverloadedStrings dflags
2849 unaries :: [(Inst, Class, TcTyVar)] -- (C tv) constraints
2850 bad_tvs :: TcTyVarSet -- Tyvars mentioned by *other* constraints
2851 (unaries, bad_tvs_s) = partitionWith find_unary insts
2852 bad_tvs = unionVarSets bad_tvs_s
2854 -- Finds unary type-class constraints
2855 find_unary d@(Dict {tci_pred = ClassP cls [ty]})
2856 | Just tv <- tcGetTyVar_maybe ty = Left (d,cls,tv)
2857 find_unary inst = Right (tyVarsOfInst inst)
2859 -- Group by type variable
2860 defaultable_groups :: [[(Inst,Class,TcTyVar)]]
2861 defaultable_groups = filter defaultable_group (equivClasses cmp_tv unaries)
2862 cmp_tv (_,_,tv1) (_,_,tv2) = tv1 `compare` tv2
2864 defaultable_group :: [(Inst,Class,TcTyVar)] -> Bool
2865 defaultable_group ds@((_,_,tv):_)
2866 = isTyConableTyVar tv -- Note [Avoiding spurious errors]
2867 && not (tv `elemVarSet` bad_tvs)
2868 && defaultable_classes [c | (_,c,_) <- ds]
2869 defaultable_group [] = panic "defaultable_group"
2871 defaultable_classes clss
2872 | extended_defaulting = any isInteractiveClass clss
2873 | otherwise = all is_std_class clss && (any is_num_class clss)
2875 -- In interactive mode, or with -XExtendedDefaultRules,
2876 -- we default Show a to Show () to avoid graututious errors on "show []"
2877 isInteractiveClass cls
2878 = is_num_class cls || (classKey cls `elem` [showClassKey, eqClassKey, ordClassKey])
2880 is_num_class cls = isNumericClass cls || (ovl_strings && (cls `hasKey` isStringClassKey))
2881 -- is_num_class adds IsString to the standard numeric classes,
2882 -- when -foverloaded-strings is enabled
2884 is_std_class cls = isStandardClass cls || (ovl_strings && (cls `hasKey` isStringClassKey))
2885 -- Similarly is_std_class
2887 -----------------------
2888 disambigGroup :: [Type] -- The default types
2889 -> [(Inst,Class,TcTyVar)] -- All standard classes of form (C a)
2890 -> TcM () -- Just does unification, to fix the default types
2892 disambigGroup default_tys dicts
2893 = do { mb_chosen_ty <- try_default default_tys
2894 ; case mb_chosen_ty of
2895 Nothing -> return ()
2896 Just chosen_ty -> do { _ <- unifyType chosen_ty (mkTyVarTy tyvar)
2897 ; warnDefault dicts chosen_ty } }
2899 (_,_,tyvar) = ASSERT(not (null dicts)) head dicts -- Should be non-empty
2900 classes = [c | (_,c,_) <- dicts]
2902 try_default [] = return Nothing
2903 try_default (default_ty : default_tys)
2904 = tryTcLIE_ (try_default default_tys) $
2905 do { tcSimplifyDefault [mkClassPred clas [default_ty] | clas <- classes]
2906 -- This may fail; then the tryTcLIE_ kicks in
2907 -- Failure here is caused by there being no type in the
2908 -- default list which can satisfy all the ambiguous classes.
2909 -- For example, if Real a is reqd, but the only type in the
2910 -- default list is Int.
2912 ; return (Just default_ty) -- TOMDO: do something with the coercion
2916 -----------------------
2917 getDefaultTys :: Bool -> Bool -> TcM [Type]
2918 getDefaultTys extended_deflts ovl_strings
2919 = do { mb_defaults <- getDeclaredDefaultTys
2920 ; case mb_defaults of {
2921 Just tys -> return tys ; -- User-supplied defaults
2924 -- No use-supplied default
2925 -- Use [Integer, Double], plus modifications
2926 { integer_ty <- tcMetaTy integerTyConName
2927 ; checkWiredInTyCon doubleTyCon
2928 ; string_ty <- tcMetaTy stringTyConName
2929 ; return (opt_deflt extended_deflts unitTy
2930 -- Note [Default unitTy]
2932 [integer_ty,doubleTy]
2934 opt_deflt ovl_strings string_ty) } } }
2936 opt_deflt True ty = [ty]
2937 opt_deflt False _ = []
2940 Note [Default unitTy]
2941 ~~~~~~~~~~~~~~~~~~~~~
2942 In interative mode (or with -XExtendedDefaultRules) we add () as the first type we
2943 try when defaulting. This has very little real impact, except in the following case.
2945 Text.Printf.printf "hello"
2946 This has type (forall a. IO a); it prints "hello", and returns 'undefined'. We don't
2947 want the GHCi repl loop to try to print that 'undefined'. The neatest thing is to
2948 default the 'a' to (), rather than to Integer (which is what would otherwise happen;
2949 and then GHCi doesn't attempt to print the (). So in interactive mode, we add
2950 () to the list of defaulting types. See Trac #1200.
2952 Note [Avoiding spurious errors]
2953 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2954 When doing the unification for defaulting, we check for skolem
2955 type variables, and simply don't default them. For example:
2956 f = (*) -- Monomorphic
2957 g :: Num a => a -> a
2959 Here, we get a complaint when checking the type signature for g,
2960 that g isn't polymorphic enough; but then we get another one when
2961 dealing with the (Num a) context arising from f's definition;
2962 we try to unify a with Int (to default it), but find that it's
2963 already been unified with the rigid variable from g's type sig
2966 %************************************************************************
2968 \subsection[simple]{@Simple@ versions}
2970 %************************************************************************
2972 Much simpler versions when there are no bindings to make!
2974 @tcSimplifyThetas@ simplifies class-type constraints formed by
2975 @deriving@ declarations and when specialising instances. We are
2976 only interested in the simplified bunch of class/type constraints.
2978 It simplifies to constraints of the form (C a b c) where
2979 a,b,c are type variables. This is required for the context of
2980 instance declarations.
2983 tcSimplifyDeriv :: InstOrigin
2985 -> ThetaType -- Wanted
2986 -> TcM ThetaType -- Needed
2987 -- Given instance (wanted) => C inst_ty
2988 -- Simplify 'wanted' as much as possible
2990 tcSimplifyDeriv orig tyvars theta
2991 = do { (tvs, _, tenv) <- tcInstTyVars tyvars
2992 -- The main loop may do unification, and that may crash if
2993 -- it doesn't see a TcTyVar, so we have to instantiate. Sigh
2994 -- ToDo: what if two of them do get unified?
2995 ; wanteds <- newCtGivensO orig (substTheta tenv theta)
2996 ; (irreds, _) <- tryHardCheckLoop doc wanteds
2998 ; let (tv_dicts, others) = partition ok irreds
2999 (tidy_env, tidy_insts) = tidyInsts others
3000 ; reportNoInstances tidy_env Nothing [alt_fix] tidy_insts
3001 -- See Note [Exotic derived instance contexts] in TcMType
3003 ; let rev_env = zipTopTvSubst tvs (mkTyVarTys tyvars)
3004 simpl_theta = substTheta rev_env (map dictPred tv_dicts)
3005 -- This reverse-mapping is a pain, but the result
3006 -- should mention the original TyVars not TcTyVars
3008 ; return simpl_theta }
3010 doc = ptext (sLit "deriving classes for a data type")
3012 ok dict | isDict dict = validDerivPred (dictPred dict)
3014 alt_fix = vcat [ptext (sLit "use a standalone 'deriving instance' declaration instead,"),
3015 ptext (sLit "so you can specify the instance context yourself")]
3019 @tcSimplifyDefault@ just checks class-type constraints, essentially;
3020 used with \tr{default} declarations. We are only interested in
3021 whether it worked or not.
3024 tcSimplifyDefault :: ThetaType -- Wanted; has no type variables in it
3027 tcSimplifyDefault theta = do
3028 wanteds <- newCtGivensO DefaultOrigin theta
3029 (irreds, _) <- tryHardCheckLoop doc wanteds
3030 addNoInstanceErrs irreds
3034 traceTc (ptext (sLit "tcSimplifyDefault failing")) >> failM
3036 doc = ptext (sLit "default declaration")
3041 %************************************************************************
3043 \section{Errors and contexts}
3045 %************************************************************************
3047 ToDo: for these error messages, should we note the location as coming
3048 from the insts, or just whatever seems to be around in the monad just
3052 groupErrs :: ([Inst] -> TcM ()) -- Deal with one group
3053 -> [Inst] -- The offending Insts
3055 -- Group together insts with the same origin
3056 -- We want to report them together in error messages
3060 groupErrs report_err (inst:insts)
3061 = do { do_one (inst:friends)
3062 ; groupErrs report_err others }
3064 -- (It may seem a bit crude to compare the error messages,
3065 -- but it makes sure that we combine just what the user sees,
3066 -- and it avoids need equality on InstLocs.)
3067 (friends, others) = partition is_friend insts
3068 loc_msg = showSDoc (pprInstLoc (instLoc inst))
3069 is_friend friend = showSDoc (pprInstLoc (instLoc friend)) == loc_msg
3070 do_one insts = setInstCtxt (instLoc (head insts)) (report_err insts)
3071 -- Add location and context information derived from the Insts
3073 -- Add the "arising from..." part to a message about bunch of dicts
3074 addInstLoc :: [Inst] -> Message -> Message
3075 addInstLoc insts msg = msg $$ nest 2 (pprInstArising (head insts))
3077 addTopIPErrs :: [Name] -> [Inst] -> TcM ()
3080 addTopIPErrs bndrs ips
3081 = do { dflags <- getDOpts
3082 ; addErrTcM (tidy_env, mk_msg dflags tidy_ips) }
3084 (tidy_env, tidy_ips) = tidyInsts ips
3086 = vcat [sep [ptext (sLit "Implicit parameters escape from"),
3087 nest 2 (ptext (sLit "the monomorphic top-level binding")
3088 <> plural bndrs <+> ptext (sLit "of")
3089 <+> pprBinders bndrs <> colon)],
3090 nest 2 (vcat (map ppr_ip ips)),
3091 monomorphism_fix dflags]
3092 ppr_ip ip = pprPred (dictPred ip) <+> pprInstArising ip
3094 topIPErrs :: [Inst] -> TcM ()
3096 = groupErrs report tidy_dicts
3098 (tidy_env, tidy_dicts) = tidyInsts dicts
3099 report dicts = addErrTcM (tidy_env, mk_msg dicts)
3100 mk_msg dicts = addInstLoc dicts (ptext (sLit "Unbound implicit parameter") <>
3101 plural tidy_dicts <+> pprDictsTheta tidy_dicts)
3103 addNoInstanceErrs :: [Inst] -- Wanted (can include implications)
3105 addNoInstanceErrs insts
3106 = do { let (tidy_env, tidy_insts) = tidyInsts insts
3107 ; reportNoInstances tidy_env Nothing [] tidy_insts }
3111 -> Maybe (InstLoc, [Inst]) -- Context
3112 -- Nothing => top level
3113 -- Just (d,g) => d describes the construct
3115 -> [SDoc] -- Alternative fix for no-such-instance
3116 -> [Inst] -- What is wanted (can include implications)
3119 reportNoInstances tidy_env mb_what alt_fix insts
3120 = groupErrs (report_no_instances tidy_env mb_what alt_fix) insts
3122 report_no_instances :: TidyEnv -> Maybe (InstLoc, [Inst]) -> [SDoc] -> [Inst] -> TcM ()
3123 report_no_instances tidy_env mb_what alt_fixes insts
3124 = do { inst_envs <- tcGetInstEnvs
3125 ; let (implics, insts1) = partition isImplicInst insts
3126 (insts2, overlaps) = partitionWith (check_overlap inst_envs) insts1
3127 (eqInsts, insts3) = partition isEqInst insts2
3128 ; traceTc (text "reportNoInstances" <+> vcat
3129 [ppr insts, ppr implics, ppr insts1, ppr insts2])
3130 ; mapM_ complain_implic implics
3131 ; mapM_ (\doc -> addErrTcM (tidy_env, doc)) overlaps
3132 ; groupErrs complain_no_inst insts3
3133 ; mapM_ (addErrTcM . mk_eq_err) eqInsts
3136 complain_no_inst insts = addErrTcM (tidy_env, mk_no_inst_err insts)
3138 complain_implic inst -- Recurse!
3139 = reportNoInstances tidy_env
3140 (Just (tci_loc inst, tci_given inst))
3141 alt_fixes (tci_wanted inst)
3143 check_overlap :: (InstEnv,InstEnv) -> Inst -> Either Inst SDoc
3144 -- Right msg => overlap message
3145 -- Left inst => no instance
3146 check_overlap inst_envs wanted
3147 | not (isClassDict wanted) = Left wanted
3149 = case lookupInstEnv inst_envs clas tys of
3150 ([], _) -> Left wanted -- No match
3151 -- The case of exactly one match and no unifiers means a
3152 -- successful lookup. That can't happen here, because dicts
3153 -- only end up here if they didn't match in Inst.lookupInst
3155 | debugIsOn -> pprPanic "reportNoInstance" (ppr wanted)
3156 res -> Right (mk_overlap_msg wanted res)
3158 (clas,tys) = getDictClassTys wanted
3160 mk_overlap_msg dict (matches, unifiers)
3161 = ASSERT( not (null matches) )
3162 vcat [ addInstLoc [dict] ((ptext (sLit "Overlapping instances for")
3163 <+> pprPred (dictPred dict))),
3164 sep [ptext (sLit "Matching instances") <> colon,
3165 nest 2 (vcat [pprInstances ispecs, pprInstances unifiers])],
3166 if not (isSingleton matches)
3167 then -- Two or more matches
3169 else -- One match, plus some unifiers
3170 ASSERT( not (null unifiers) )
3171 parens (vcat [ptext (sLit "The choice depends on the instantiation of") <+>
3172 quotes (pprWithCommas ppr (varSetElems (tyVarsOfInst dict))),
3173 ptext (sLit "To pick the first instance above, use -XIncoherentInstances"),
3174 ptext (sLit "when compiling the other instance declarations")])]
3176 ispecs = [ispec | (ispec, _) <- matches]
3178 mk_eq_err :: Inst -> (TidyEnv, SDoc)
3179 mk_eq_err inst = misMatchMsg tidy_env (eqInstTys inst)
3181 mk_no_inst_err insts
3182 | null insts = empty
3184 | Just (loc, givens) <- mb_what, -- Nested (type signatures, instance decls)
3185 not (isEmptyVarSet (tyVarsOfInsts insts))
3186 = vcat [ addInstLoc insts $
3187 sep [ ptext (sLit "Could not deduce") <+> pprDictsTheta insts
3188 , nest 2 $ ptext (sLit "from the context") <+> pprDictsTheta givens]
3189 , show_fixes (fix1 loc : fixes2 ++ alt_fixes) ]
3191 | otherwise -- Top level
3192 = vcat [ addInstLoc insts $
3193 ptext (sLit "No instance") <> plural insts
3194 <+> ptext (sLit "for") <+> pprDictsTheta insts
3195 , show_fixes (fixes2 ++ alt_fixes) ]
3198 fix1 loc = sep [ ptext (sLit "add") <+> pprDictsTheta insts
3199 <+> ptext (sLit "to the context of"),
3200 nest 2 (ppr (instLocOrigin loc)) ]
3201 -- I'm not sure it helps to add the location
3202 -- nest 2 (ptext (sLit "at") <+> ppr (instLocSpan loc)) ]
3204 fixes2 | null instance_dicts = []
3205 | otherwise = [sep [ptext (sLit "add an instance declaration for"),
3206 pprDictsTheta instance_dicts]]
3207 instance_dicts = [d | d <- insts, isClassDict d, not (isTyVarDict d)]
3208 -- Insts for which it is worth suggesting an adding an instance declaration
3209 -- Exclude implicit parameters, and tyvar dicts
3211 show_fixes :: [SDoc] -> SDoc
3212 show_fixes [] = empty
3213 show_fixes (f:fs) = sep [ptext (sLit "Possible fix:"),
3214 nest 2 (vcat (f : map (ptext (sLit "or") <+>) fs))]
3216 addTopAmbigErrs :: [Inst] -> TcRn ()
3217 addTopAmbigErrs dicts
3218 -- Divide into groups that share a common set of ambiguous tyvars
3219 = ifErrsM (return ()) $ -- Only report ambiguity if no other errors happened
3220 -- See Note [Avoiding spurious errors]
3221 mapM_ report (equivClasses cmp [(d, tvs_of d) | d <- tidy_dicts])
3223 (tidy_env, tidy_dicts) = tidyInsts dicts
3225 tvs_of :: Inst -> [TcTyVar]
3226 tvs_of d = varSetElems (tyVarsOfInst d)
3227 cmp (_,tvs1) (_,tvs2) = tvs1 `compare` tvs2
3229 report :: [(Inst,[TcTyVar])] -> TcM ()
3230 report pairs@((inst,tvs) : _) = do -- The pairs share a common set of ambiguous tyvars
3231 (tidy_env, mono_msg) <- mkMonomorphismMsg tidy_env tvs
3232 setSrcSpan (instSpan inst) $
3233 -- the location of the first one will do for the err message
3234 addErrTcM (tidy_env, msg $$ mono_msg)
3236 dicts = map fst pairs
3237 msg = sep [text "Ambiguous type variable" <> plural tvs <+>
3238 pprQuotedList tvs <+> in_msg,
3239 nest 2 (pprDictsInFull dicts)]
3240 in_msg = text "in the constraint" <> plural dicts <> colon
3241 report [] = panic "addTopAmbigErrs"
3244 mkMonomorphismMsg :: TidyEnv -> [TcTyVar] -> TcM (TidyEnv, Message)
3245 -- There's an error with these Insts; if they have free type variables
3246 -- it's probably caused by the monomorphism restriction.
3247 -- Try to identify the offending variable
3248 -- ASSUMPTION: the Insts are fully zonked
3249 mkMonomorphismMsg tidy_env inst_tvs
3250 = do { dflags <- getDOpts
3251 ; (tidy_env, docs) <- findGlobals (mkVarSet inst_tvs) tidy_env
3252 ; return (tidy_env, mk_msg dflags docs) }
3254 mk_msg _ _ | any isRuntimeUnk inst_tvs
3255 = vcat [ptext (sLit "Cannot resolve unknown runtime types:") <+>
3256 (pprWithCommas ppr inst_tvs),
3257 ptext (sLit "Use :print or :force to determine these types")]
3258 mk_msg _ [] = ptext (sLit "Probable fix: add a type signature that fixes these type variable(s)")
3259 -- This happens in things like
3260 -- f x = show (read "foo")
3261 -- where monomorphism doesn't play any role
3263 = vcat [ptext (sLit "Possible cause: the monomorphism restriction applied to the following:"),
3265 monomorphism_fix dflags]
3267 monomorphism_fix :: DynFlags -> SDoc
3268 monomorphism_fix dflags
3269 = ptext (sLit "Probable fix:") <+> vcat
3270 [ptext (sLit "give these definition(s) an explicit type signature"),
3271 if dopt Opt_MonomorphismRestriction dflags
3272 then ptext (sLit "or use -XNoMonomorphismRestriction")
3273 else empty] -- Only suggest adding "-XNoMonomorphismRestriction"
3274 -- if it is not already set!
3276 warnDefault :: [(Inst, Class, Var)] -> Type -> TcM ()
3277 warnDefault ups default_ty = do
3278 warn_flag <- doptM Opt_WarnTypeDefaults
3279 setInstCtxt (instLoc (head (dicts))) (warnTc warn_flag warn_msg)
3281 dicts = [d | (d,_,_) <- ups]
3284 (_, tidy_dicts) = tidyInsts dicts
3285 warn_msg = vcat [ptext (sLit "Defaulting the following constraint(s) to type") <+>
3286 quotes (ppr default_ty),
3287 pprDictsInFull tidy_dicts]
3289 reduceDepthErr :: Int -> [Inst] -> SDoc
3290 reduceDepthErr n stack
3291 = vcat [ptext (sLit "Context reduction stack overflow; size =") <+> int n,
3292 ptext (sLit "Use -fcontext-stack=N to increase stack size to N"),
3293 nest 4 (pprStack stack)]
3295 pprStack :: [Inst] -> SDoc
3296 pprStack stack = vcat (map pprInstInFull stack)