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
66 %************************************************************************
70 %************************************************************************
72 --------------------------------------
73 Notes on functional dependencies (a bug)
74 --------------------------------------
81 instance D a b => C a b -- Undecidable
82 -- (Not sure if it's crucial to this eg)
83 f :: C a b => a -> Bool
86 g :: C a b => a -> Bool
89 Here f typechecks, but g does not!! Reason: before doing improvement,
90 we reduce the (C a b1) constraint from the call of f to (D a b1).
92 Here is a more complicated example:
95 > class Foo a b | a->b
97 > class Bar a b | a->b
101 > instance Bar Obj Obj
103 > instance (Bar a b) => Foo a b
105 > foo:: (Foo a b) => a -> String
108 > runFoo:: (forall a b. (Foo a b) => a -> w) -> w
114 Could not deduce (Bar a b) from the context (Foo a b)
115 arising from use of `foo' at <interactive>:1
117 Add (Bar a b) to the expected type of an expression
118 In the first argument of `runFoo', namely `foo'
119 In the definition of `it': it = runFoo foo
121 Why all of the sudden does GHC need the constraint Bar a b? The
122 function foo didn't ask for that...
125 The trouble is that to type (runFoo foo), GHC has to solve the problem:
127 Given constraint Foo a b
128 Solve constraint Foo a b'
130 Notice that b and b' aren't the same. To solve this, just do
131 improvement and then they are the same. But GHC currently does
136 That is usually fine, but it isn't here, because it sees that Foo a b is
137 not the same as Foo a b', and so instead applies the instance decl for
138 instance Bar a b => Foo a b. And that's where the Bar constraint comes
141 The Right Thing is to improve whenever the constraint set changes at
142 all. Not hard in principle, but it'll take a bit of fiddling to do.
144 Note [Choosing which variables to quantify]
145 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
146 Suppose we are about to do a generalisation step. We have in our hand
149 T the type of the RHS
150 C the constraints from that RHS
152 The game is to figure out
154 Q the set of type variables over which to quantify
155 Ct the constraints we will *not* quantify over
156 Cq the constraints we will quantify over
158 So we're going to infer the type
162 and float the constraints Ct further outwards.
164 Here are the things that *must* be true:
166 (A) Q intersect fv(G) = EMPTY limits how big Q can be
167 (B) Q superset fv(Cq union T) \ oclose(fv(G),C) limits how small Q can be
169 (A) says we can't quantify over a variable that's free in the environment.
170 (B) says we must quantify over all the truly free variables in T, else
171 we won't get a sufficiently general type.
173 We do not *need* to quantify over any variable that is fixed by the
174 free vars of the environment G.
176 BETWEEN THESE TWO BOUNDS, ANY Q WILL DO!
178 Example: class H x y | x->y where ...
180 fv(G) = {a} C = {H a b, H c d}
183 (A) Q intersect {a} is empty
184 (B) Q superset {a,b,c,d} \ oclose({a}, C) = {a,b,c,d} \ {a,b} = {c,d}
186 So Q can be {c,d}, {b,c,d}
188 In particular, it's perfectly OK to quantify over more type variables
189 than strictly necessary; there is no need to quantify over 'b', since
190 it is determined by 'a' which is free in the envt, but it's perfectly
191 OK to do so. However we must not quantify over 'a' itself.
193 Other things being equal, however, we'd like to quantify over as few
194 variables as possible: smaller types, fewer type applications, more
195 constraints can get into Ct instead of Cq. Here's a good way to
198 Q = grow( fv(T), C ) \ oclose( fv(G), C )
200 That is, quantify over all variable that that MIGHT be fixed by the
201 call site (which influences T), but which aren't DEFINITELY fixed by
202 G. This choice definitely quantifies over enough type variables,
203 albeit perhaps too many.
205 Why grow( fv(T), C ) rather than fv(T)? Consider
207 class H x y | x->y where ...
212 If we used fv(T) = {c} we'd get the type
214 forall c. H c d => c -> b
216 And then if the fn was called at several different c's, each of
217 which fixed d differently, we'd get a unification error, because
218 d isn't quantified. Solution: quantify d. So we must quantify
219 everything that might be influenced by c.
221 Why not oclose( fv(T), C )? Because we might not be able to see
222 all the functional dependencies yet:
224 class H x y | x->y where ...
225 instance H x y => Eq (T x y) where ...
230 Now oclose(fv(T),C) = {c}, because the functional dependency isn't
231 apparent yet, and that's wrong. We must really quantify over d too.
233 There really isn't any point in quantifying over any more than
234 grow( fv(T), C ), because the call sites can't possibly influence
235 any other type variables.
239 -------------------------------------
241 -------------------------------------
243 It's very hard to be certain when a type is ambiguous. Consider
247 instance H x y => K (x,y)
249 Is this type ambiguous?
250 forall a b. (K (a,b), Eq b) => a -> a
252 Looks like it! But if we simplify (K (a,b)) we get (H a b) and
253 now we see that a fixes b. So we can't tell about ambiguity for sure
254 without doing a full simplification. And even that isn't possible if
255 the context has some free vars that may get unified. Urgle!
257 Here's another example: is this ambiguous?
258 forall a b. Eq (T b) => a -> a
259 Not if there's an insance decl (with no context)
260 instance Eq (T b) where ...
262 You may say of this example that we should use the instance decl right
263 away, but you can't always do that:
265 class J a b where ...
266 instance J Int b where ...
268 f :: forall a b. J a b => a -> a
270 (Notice: no functional dependency in J's class decl.)
271 Here f's type is perfectly fine, provided f is only called at Int.
272 It's premature to complain when meeting f's signature, or even
273 when inferring a type for f.
277 However, we don't *need* to report ambiguity right away. It'll always
278 show up at the call site.... and eventually at main, which needs special
279 treatment. Nevertheless, reporting ambiguity promptly is an excellent thing.
281 So here's the plan. We WARN about probable ambiguity if
283 fv(Cq) is not a subset of oclose(fv(T) union fv(G), C)
285 (all tested before quantification).
286 That is, all the type variables in Cq must be fixed by the the variables
287 in the environment, or by the variables in the type.
289 Notice that we union before calling oclose. Here's an example:
291 class J a b c | a b -> c
295 forall b c. (J a b c) => b -> b
297 Only if we union {a} from G with {b} from T before using oclose,
298 do we see that c is fixed.
300 It's a bit vague exactly which C we should use for this oclose call. If we
301 don't fix enough variables we might complain when we shouldn't (see
302 the above nasty example). Nothing will be perfect. That's why we can
303 only issue a warning.
306 Can we ever be *certain* about ambiguity? Yes: if there's a constraint
308 c in C such that fv(c) intersect (fv(G) union fv(T)) = EMPTY
310 then c is a "bubble"; there's no way it can ever improve, and it's
311 certainly ambiguous. UNLESS it is a constant (sigh). And what about
316 instance H x y => K (x,y)
318 Is this type ambiguous?
319 forall a b. (K (a,b), Eq b) => a -> a
321 Urk. The (Eq b) looks "definitely ambiguous" but it isn't. What we are after
322 is a "bubble" that's a set of constraints
324 Cq = Ca union Cq' st fv(Ca) intersect (fv(Cq') union fv(T) union fv(G)) = EMPTY
326 Hence another idea. To decide Q start with fv(T) and grow it
327 by transitive closure in Cq (no functional dependencies involved).
328 Now partition Cq using Q, leaving the definitely-ambiguous and probably-ok.
329 The definitely-ambiguous can then float out, and get smashed at top level
330 (which squashes out the constants, like Eq (T a) above)
333 --------------------------------------
334 Notes on principal types
335 --------------------------------------
340 f x = let g y = op (y::Int) in True
342 Here the principal type of f is (forall a. a->a)
343 but we'll produce the non-principal type
344 f :: forall a. C Int => a -> a
347 --------------------------------------
348 The need for forall's in constraints
349 --------------------------------------
351 [Exchange on Haskell Cafe 5/6 Dec 2000]
353 class C t where op :: t -> Bool
354 instance C [t] where op x = True
356 p y = (let f :: c -> Bool; f x = op (y >> return x) in f, y ++ [])
357 q y = (y ++ [], let f :: c -> Bool; f x = op (y >> return x) in f)
359 The definitions of p and q differ only in the order of the components in
360 the pair on their right-hand sides. And yet:
362 ghc and "Typing Haskell in Haskell" reject p, but accept q;
363 Hugs rejects q, but accepts p;
364 hbc rejects both p and q;
365 nhc98 ... (Malcolm, can you fill in the blank for us!).
367 The type signature for f forces context reduction to take place, and
368 the results of this depend on whether or not the type of y is known,
369 which in turn depends on which component of the pair the type checker
372 Solution: if y::m a, float out the constraints
373 Monad m, forall c. C (m c)
374 When m is later unified with [], we can solve both constraints.
377 --------------------------------------
378 Notes on implicit parameters
379 --------------------------------------
381 Note [Inheriting implicit parameters]
382 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
387 where f is *not* a top-level binding.
388 From the RHS of f we'll get the constraint (?y::Int).
389 There are two types we might infer for f:
393 (so we get ?y from the context of f's definition), or
395 f :: (?y::Int) => Int -> Int
397 At first you might think the first was better, becuase then
398 ?y behaves like a free variable of the definition, rather than
399 having to be passed at each call site. But of course, the WHOLE
400 IDEA is that ?y should be passed at each call site (that's what
401 dynamic binding means) so we'd better infer the second.
403 BOTTOM LINE: when *inferring types* you *must* quantify
404 over implicit parameters. See the predicate isFreeWhenInferring.
407 Note [Implicit parameters and ambiguity]
408 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
409 Only a *class* predicate can give rise to ambiguity
410 An *implicit parameter* cannot. For example:
411 foo :: (?x :: [a]) => Int
413 is fine. The call site will suppply a particular 'x'
415 Furthermore, the type variables fixed by an implicit parameter
416 propagate to the others. E.g.
417 foo :: (Show a, ?x::[a]) => Int
419 The type of foo looks ambiguous. But it isn't, because at a call site
421 let ?x = 5::Int in foo
422 and all is well. In effect, implicit parameters are, well, parameters,
423 so we can take their type variables into account as part of the
424 "tau-tvs" stuff. This is done in the function 'FunDeps.grow'.
427 Question 2: type signatures
428 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
429 BUT WATCH OUT: When you supply a type signature, we can't force you
430 to quantify over implicit parameters. For example:
434 This is perfectly reasonable. We do not want to insist on
436 (?x + 1) :: (?x::Int => Int)
438 That would be silly. Here, the definition site *is* the occurrence site,
439 so the above strictures don't apply. Hence the difference between
440 tcSimplifyCheck (which *does* allow implicit paramters to be inherited)
441 and tcSimplifyCheckBind (which does not).
443 What about when you supply a type signature for a binding?
444 Is it legal to give the following explicit, user type
445 signature to f, thus:
450 At first sight this seems reasonable, but it has the nasty property
451 that adding a type signature changes the dynamic semantics.
454 (let f x = (x::Int) + ?y
455 in (f 3, f 3 with ?y=5)) with ?y = 6
461 in (f 3, f 3 with ?y=5)) with ?y = 6
465 Indeed, simply inlining f (at the Haskell source level) would change the
468 Nevertheless, as Launchbury says (email Oct 01) we can't really give the
469 semantics for a Haskell program without knowing its typing, so if you
470 change the typing you may change the semantics.
472 To make things consistent in all cases where we are *checking* against
473 a supplied signature (as opposed to inferring a type), we adopt the
476 a signature does not need to quantify over implicit params.
478 [This represents a (rather marginal) change of policy since GHC 5.02,
479 which *required* an explicit signature to quantify over all implicit
480 params for the reasons mentioned above.]
482 But that raises a new question. Consider
484 Given (signature) ?x::Int
485 Wanted (inferred) ?x::Int, ?y::Bool
487 Clearly we want to discharge the ?x and float the ?y out. But
488 what is the criterion that distinguishes them? Clearly it isn't
489 what free type variables they have. The Right Thing seems to be
490 to float a constraint that
491 neither mentions any of the quantified type variables
492 nor any of the quantified implicit parameters
494 See the predicate isFreeWhenChecking.
497 Question 3: monomorphism
498 ~~~~~~~~~~~~~~~~~~~~~~~~
499 There's a nasty corner case when the monomorphism restriction bites:
503 The argument above suggests that we *must* generalise
504 over the ?y parameter, to get
505 z :: (?y::Int) => Int,
506 but the monomorphism restriction says that we *must not*, giving
508 Why does the momomorphism restriction say this? Because if you have
510 let z = x + ?y in z+z
512 you might not expect the addition to be done twice --- but it will if
513 we follow the argument of Question 2 and generalise over ?y.
516 Question 4: top level
517 ~~~~~~~~~~~~~~~~~~~~~
518 At the top level, monomorhism makes no sense at all.
521 main = let ?x = 5 in print foo
525 woggle :: (?x :: Int) => Int -> Int
528 We definitely don't want (foo :: Int) with a top-level implicit parameter
529 (?x::Int) becuase there is no way to bind it.
534 (A) Always generalise over implicit parameters
535 Bindings that fall under the monomorphism restriction can't
539 * Inlining remains valid
540 * No unexpected loss of sharing
541 * But simple bindings like
543 will be rejected, unless you add an explicit type signature
544 (to avoid the monomorphism restriction)
545 z :: (?y::Int) => Int
547 This seems unacceptable
549 (B) Monomorphism restriction "wins"
550 Bindings that fall under the monomorphism restriction can't
552 Always generalise over implicit parameters *except* for bindings
553 that fall under the monomorphism restriction
556 * Inlining isn't valid in general
557 * No unexpected loss of sharing
558 * Simple bindings like
560 accepted (get value of ?y from binding site)
562 (C) Always generalise over implicit parameters
563 Bindings that fall under the monomorphism restriction can't
564 be generalised, EXCEPT for implicit parameters
566 * Inlining remains valid
567 * Unexpected loss of sharing (from the extra generalisation)
568 * Simple bindings like
570 accepted (get value of ?y from occurrence sites)
575 None of these choices seems very satisfactory. But at least we should
576 decide which we want to do.
578 It's really not clear what is the Right Thing To Do. If you see
582 would you expect the value of ?y to be got from the *occurrence sites*
583 of 'z', or from the valuue of ?y at the *definition* of 'z'? In the
584 case of function definitions, the answer is clearly the former, but
585 less so in the case of non-fucntion definitions. On the other hand,
586 if we say that we get the value of ?y from the definition site of 'z',
587 then inlining 'z' might change the semantics of the program.
589 Choice (C) really says "the monomorphism restriction doesn't apply
590 to implicit parameters". Which is fine, but remember that every
591 innocent binding 'x = ...' that mentions an implicit parameter in
592 the RHS becomes a *function* of that parameter, called at each
593 use of 'x'. Now, the chances are that there are no intervening 'with'
594 clauses that bind ?y, so a decent compiler should common up all
595 those function calls. So I think I strongly favour (C). Indeed,
596 one could make a similar argument for abolishing the monomorphism
597 restriction altogether.
599 BOTTOM LINE: we choose (B) at present. See tcSimplifyRestricted
603 %************************************************************************
605 \subsection{tcSimplifyInfer}
607 %************************************************************************
609 tcSimplify is called when we *inferring* a type. Here's the overall game plan:
611 1. Compute Q = grow( fvs(T), C )
613 2. Partition C based on Q into Ct and Cq. Notice that ambiguous
614 predicates will end up in Ct; we deal with them at the top level
616 3. Try improvement, using functional dependencies
618 4. If Step 3 did any unification, repeat from step 1
619 (Unification can change the result of 'grow'.)
621 Note: we don't reduce dictionaries in step 2. For example, if we have
622 Eq (a,b), we don't simplify to (Eq a, Eq b). So Q won't be different
623 after step 2. However note that we may therefore quantify over more
624 type variables than we absolutely have to.
626 For the guts, we need a loop, that alternates context reduction and
627 improvement with unification. E.g. Suppose we have
629 class C x y | x->y where ...
631 and tcSimplify is called with:
633 Then improvement unifies a with b, giving
636 If we need to unify anything, we rattle round the whole thing all over
643 -> TcTyVarSet -- fv(T); type vars
645 -> TcM ([TcTyVar], -- Tyvars to quantify (zonked and quantified)
646 [Inst], -- Dict Ids that must be bound here (zonked)
647 TcDictBinds) -- Bindings
648 -- Any free (escaping) Insts are tossed into the environment
653 tcSimplifyInfer doc tau_tvs wanted
654 = do { tau_tvs1 <- zonkTcTyVarsAndFV (varSetElems tau_tvs)
655 ; wanted' <- mapM zonkInst wanted -- Zonk before deciding quantified tyvars
656 ; gbl_tvs <- tcGetGlobalTyVars
657 ; let preds1 = fdPredsOfInsts wanted'
658 gbl_tvs1 = oclose preds1 gbl_tvs
659 qtvs = grow preds1 tau_tvs1 `minusVarSet` gbl_tvs1
660 -- See Note [Choosing which variables to quantify]
662 -- To maximise sharing, remove from consideration any
663 -- constraints that don't mention qtvs at all
664 ; let (free, bound) = partition (isFreeWhenInferring qtvs) wanted'
667 -- To make types simple, reduce as much as possible
668 ; traceTc (text "infer" <+> (ppr preds1 $$ ppr (grow preds1 tau_tvs1) $$ ppr gbl_tvs $$
669 ppr gbl_tvs1 $$ ppr free $$ ppr bound))
670 ; (irreds1, binds1) <- tryHardCheckLoop doc bound
672 -- Note [Inference and implication constraints]
673 ; let want_dict d = tyVarsOfInst d `intersectsVarSet` qtvs
674 ; (irreds2, binds2) <- approximateImplications doc want_dict irreds1
676 -- Now work out all over again which type variables to quantify,
677 -- exactly in the same way as before, but starting from irreds2. Why?
678 -- a) By now improvment may have taken place, and we must *not*
679 -- quantify over any variable free in the environment
680 -- tc137 (function h inside g) is an example
682 -- b) Do not quantify over constraints that *now* do not
683 -- mention quantified type variables, because they are
684 -- simply ambiguous (or might be bound further out). Example:
685 -- f :: Eq b => a -> (a, b)
687 -- From the RHS of g we get the MethodInst f77 :: alpha -> (alpha, beta)
688 -- We decide to quantify over 'alpha' alone, but free1 does not include f77
689 -- because f77 mentions 'alpha'. Then reducing leaves only the (ambiguous)
690 -- constraint (Eq beta), which we dump back into the free set
691 -- See test tcfail181
693 -- c) irreds may contain type variables not previously mentioned,
694 -- e.g. instance D a x => Foo [a]
696 -- Then after simplifying we'll get (D a x), and x is fresh
697 -- We must quantify over x else it'll be totally unbound
698 ; tau_tvs2 <- zonkTcTyVarsAndFV (varSetElems tau_tvs1)
699 ; gbl_tvs2 <- zonkTcTyVarsAndFV (varSetElems gbl_tvs1)
700 -- Note that we start from gbl_tvs1
701 -- We use tcGetGlobalTyVars, then oclose wrt preds2, because
702 -- we've already put some of the original preds1 into frees
703 -- E.g. wanteds = C a b (where a->b)
706 -- Then b is fixed by gbl_tvs, so (C a b) will be in free, and
707 -- irreds2 will be empty. But we don't want to generalise over b!
708 ; let preds2 = fdPredsOfInsts irreds2 -- irreds2 is zonked
709 qtvs = grow preds2 tau_tvs2 `minusVarSet` oclose preds2 gbl_tvs2
710 ; let (free, irreds3) = partition (isFreeWhenInferring qtvs) irreds2
713 -- Turn the quantified meta-type variables into real type variables
714 ; qtvs2 <- zonkQuantifiedTyVars (varSetElems qtvs)
716 -- We can't abstract over any remaining unsolved
717 -- implications so instead just float them outwards. Ugh.
718 ; let (q_dicts0, implics) = partition isAbstractableInst irreds3
719 ; loc <- getInstLoc (ImplicOrigin doc)
720 ; implic_bind <- bindIrreds loc qtvs2 q_dicts0 implics
722 -- Prepare equality instances for quantification
723 ; let (q_eqs0,q_dicts) = partition isEqInst q_dicts0
724 ; q_eqs <- mapM finalizeEqInst q_eqs0
726 ; return (qtvs2, q_eqs ++ q_dicts, binds1 `unionBags` binds2 `unionBags` implic_bind) }
727 -- NB: when we are done, we might have some bindings, but
728 -- the final qtvs might be empty. See Note [NO TYVARS] below.
730 approximateImplications :: SDoc -> (Inst -> Bool) -> [Inst] -> TcM ([Inst], TcDictBinds)
731 -- Note [Inference and implication constraints]
732 -- Given a bunch of Dict and ImplicInsts, try to approximate the implications by
733 -- - fetching any dicts inside them that are free
734 -- - using those dicts as cruder constraints, to solve the implications
735 -- - returning the extra ones too
737 approximateImplications doc want_dict irreds
739 = return (irreds, emptyBag)
741 = do { extra_dicts' <- mapM cloneDict extra_dicts
742 ; tryHardCheckLoop doc (extra_dicts' ++ irreds) }
743 -- By adding extra_dicts', we make them
744 -- available to solve the implication constraints
746 extra_dicts = get_dicts (filter isImplicInst irreds)
748 get_dicts :: [Inst] -> [Inst] -- Returns only Dicts
749 -- Find the wanted constraints in implication constraints that satisfy
750 -- want_dict, and are not bound by forall's in the constraint itself
751 get_dicts ds = concatMap get_dict ds
753 get_dict d@(Dict {}) | want_dict d = [d]
755 get_dict (ImplicInst {tci_tyvars = tvs, tci_wanted = wanteds})
756 = [ d | let tv_set = mkVarSet tvs
757 , d <- get_dicts wanteds
758 , not (tyVarsOfInst d `intersectsVarSet` tv_set)]
759 get_dict i@(EqInst {}) | want_dict i = [i]
761 get_dict other = pprPanic "approximateImplications" (ppr other)
764 Note [Inference and implication constraints]
765 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
766 Suppose we have a wanted implication constraint (perhaps arising from
767 a nested pattern match) like
769 and we are now trying to quantify over 'a' when inferring the type for
770 a function. In principle it's possible that there might be an instance
771 instance (C a, E a) => D [a]
772 so the context (E a) would suffice. The Right Thing is to abstract over
773 the implication constraint, but we don't do that (a) because it'll be
774 surprising to programmers and (b) because we don't have the machinery to deal
775 with 'given' implications.
777 So our best approximation is to make (D [a]) part of the inferred
778 context, so we can use that to discharge the implication. Hence
779 the strange function get_dicts in approximateImplications.
781 The common cases are more clear-cut, when we have things like
783 Here, abstracting over (C b) is not an approximation at all -- but see
784 Note [Freeness and implications].
786 See Trac #1430 and test tc228.
790 -----------------------------------------------------------
791 -- tcSimplifyInferCheck is used when we know the constraints we are to simplify
792 -- against, but we don't know the type variables over which we are going to quantify.
793 -- This happens when we have a type signature for a mutually recursive group
796 -> TcTyVarSet -- fv(T)
799 -> TcM ([TyVar], -- Fully zonked, and quantified
800 TcDictBinds) -- Bindings
802 tcSimplifyInferCheck loc tau_tvs givens wanteds
803 = do { traceTc (text "tcSimplifyInferCheck <-" <+> ppr wanteds)
804 ; (irreds, binds) <- gentleCheckLoop loc givens wanteds
806 -- Figure out which type variables to quantify over
807 -- You might think it should just be the signature tyvars,
808 -- but in bizarre cases you can get extra ones
809 -- f :: forall a. Num a => a -> a
810 -- f x = fst (g (x, head [])) + 1
812 -- Here we infer g :: forall a b. a -> b -> (b,a)
813 -- We don't want g to be monomorphic in b just because
814 -- f isn't quantified over b.
815 ; let all_tvs = varSetElems (tau_tvs `unionVarSet` tyVarsOfInsts givens)
816 ; all_tvs <- zonkTcTyVarsAndFV all_tvs
817 ; gbl_tvs <- tcGetGlobalTyVars
818 ; let qtvs = varSetElems (all_tvs `minusVarSet` gbl_tvs)
819 -- We could close gbl_tvs, but its not necessary for
820 -- soundness, and it'll only affect which tyvars, not which
821 -- dictionaries, we quantify over
823 ; qtvs' <- zonkQuantifiedTyVars qtvs
825 -- Now we are back to normal (c.f. tcSimplCheck)
826 ; implic_bind <- bindIrreds loc qtvs' givens irreds
828 ; traceTc (text "tcSimplifyInferCheck ->" <+> ppr (implic_bind))
829 ; return (qtvs', binds `unionBags` implic_bind) }
832 Note [Squashing methods]
833 ~~~~~~~~~~~~~~~~~~~~~~~~~
834 Be careful if you want to float methods more:
835 truncate :: forall a. RealFrac a => forall b. Integral b => a -> b
836 From an application (truncate f i) we get
839 If we have also have a second occurrence of truncate, we get
842 When simplifying with i,f free, we might still notice that
843 t1=t3; but alas, the binding for t2 (which mentions t1)
844 may continue to float out!
849 class Y a b | a -> b where
852 instance Y [[a]] a where
855 k :: X a -> X a -> X a
857 g :: Num a => [X a] -> [X a]
860 h ys = ys ++ map (k (y [[0]])) xs
862 The excitement comes when simplifying the bindings for h. Initially
863 try to simplify {y @ [[t1]] t2, 0 @ t1}, with initial qtvs = {t2}.
864 From this we get t1:=:t2, but also various bindings. We can't forget
865 the bindings (because of [LOOP]), but in fact t1 is what g is
868 The net effect of [NO TYVARS]
871 isFreeWhenInferring :: TyVarSet -> Inst -> Bool
872 isFreeWhenInferring qtvs inst
873 = isFreeWrtTyVars qtvs inst -- Constrains no quantified vars
874 && isInheritableInst inst -- and no implicit parameter involved
875 -- see Note [Inheriting implicit parameters]
877 {- No longer used (with implication constraints)
878 isFreeWhenChecking :: TyVarSet -- Quantified tyvars
879 -> NameSet -- Quantified implicit parameters
881 isFreeWhenChecking qtvs ips inst
882 = isFreeWrtTyVars qtvs inst
883 && isFreeWrtIPs ips inst
886 isFreeWrtTyVars :: VarSet -> Inst -> Bool
887 isFreeWrtTyVars qtvs inst = tyVarsOfInst inst `disjointVarSet` qtvs
888 isFreeWrtIPs :: NameSet -> Inst -> Bool
889 isFreeWrtIPs ips inst = not (any (`elemNameSet` ips) (ipNamesOfInst inst))
893 %************************************************************************
895 \subsection{tcSimplifyCheck}
897 %************************************************************************
899 @tcSimplifyCheck@ is used when we know exactly the set of variables
900 we are going to quantify over. For example, a class or instance declaration.
903 -----------------------------------------------------------
904 -- tcSimplifyCheck is used when checking expression type signatures,
905 -- class decls, instance decls etc.
906 tcSimplifyCheck :: InstLoc
907 -> [TcTyVar] -- Quantify over these
910 -> TcM TcDictBinds -- Bindings
911 tcSimplifyCheck loc qtvs givens wanteds
912 = ASSERT( all isTcTyVar qtvs && all isSkolemTyVar qtvs )
913 do { traceTc (text "tcSimplifyCheck")
914 ; (irreds, binds) <- gentleCheckLoop loc givens wanteds
915 ; implic_bind <- bindIrreds loc qtvs givens irreds
916 ; return (binds `unionBags` implic_bind) }
918 -----------------------------------------------------------
919 -- tcSimplifyCheckPat is used for existential pattern match
920 tcSimplifyCheckPat :: InstLoc
921 -> [TcTyVar] -- Quantify over these
924 -> TcM TcDictBinds -- Bindings
925 tcSimplifyCheckPat loc qtvs givens wanteds
926 = ASSERT( all isTcTyVar qtvs && all isSkolemTyVar qtvs )
927 do { traceTc (text "tcSimplifyCheckPat")
928 ; (irreds, binds) <- gentleCheckLoop loc givens wanteds
929 ; implic_bind <- bindIrredsR loc qtvs givens irreds
930 ; return (binds `unionBags` implic_bind) }
932 -----------------------------------------------------------
933 bindIrreds :: InstLoc -> [TcTyVar]
936 bindIrreds loc qtvs givens irreds
937 = bindIrredsR loc qtvs givens irreds
939 bindIrredsR :: InstLoc -> [TcTyVar] -> [Inst] -> [Inst] -> TcM TcDictBinds
940 -- Make a binding that binds 'irreds', by generating an implication
941 -- constraint for them, *and* throwing the constraint into the LIE
942 bindIrredsR loc qtvs givens irreds
946 = do { let givens' = filter isAbstractableInst givens
947 -- The givens can (redundantly) include methods
948 -- We want to retain both EqInsts and Dicts
949 -- There should be no implicadtion constraints
950 -- See Note [Pruning the givens in an implication constraint]
952 -- If there are no 'givens', then it's safe to
953 -- partition the 'wanteds' by their qtvs, thereby trimming irreds
954 -- See Note [Freeness and implications]
955 ; irreds' <- if null givens'
957 { let qtv_set = mkVarSet qtvs
958 (frees, real_irreds) = partition (isFreeWrtTyVars qtv_set) irreds
960 ; return real_irreds }
963 ; (implics, bind) <- makeImplicationBind loc qtvs givens' irreds'
964 -- This call does the real work
965 -- If irreds' is empty, it does something sensible
970 makeImplicationBind :: InstLoc -> [TcTyVar]
972 -> TcM ([Inst], TcDictBinds)
973 -- Make a binding that binds 'irreds', by generating an implication
974 -- constraint for them.
976 -- The binding looks like
977 -- (ir1, .., irn) = f qtvs givens
978 -- where f is (evidence for) the new implication constraint
979 -- f :: forall qtvs. givens => (ir1, .., irn)
980 -- qtvs includes coercion variables.
982 -- This binding must line up the 'rhs' in reduceImplication
983 makeImplicationBind loc all_tvs
984 givens -- Guaranteed all Dicts or EqInsts
986 | null irreds -- If there are no irreds, we are done
987 = return ([], emptyBag)
988 | otherwise -- Otherwise we must generate a binding
989 = do { uniq <- newUnique
990 ; span <- getSrcSpanM
991 ; let (eq_givens, dict_givens) = partition isEqInst givens
993 -- extract equality binders
994 eq_cotvs = map eqInstType eq_givens
996 -- make the implication constraint instance
997 name = mkInternalName uniq (mkVarOcc "ic") span
998 implic_inst = ImplicInst { tci_name = name,
999 tci_tyvars = all_tvs,
1000 tci_given = (eq_givens ++ dict_givens),
1001 -- same order as binders
1002 tci_wanted = irreds,
1005 -- create binders for the irreducible dictionaries
1006 dict_irreds = filter (not . isEqInst) irreds
1007 dict_irred_ids = map instToId dict_irreds
1008 lpat = mkBigLHsPatTup (map (L span . VarPat) dict_irred_ids)
1010 -- create the binding
1011 rhs = L span (mkHsWrap co (HsVar (instToId implic_inst)))
1012 co = mkWpApps (map instToId dict_givens)
1013 <.> mkWpTyApps eq_cotvs
1014 <.> mkWpTyApps (mkTyVarTys all_tvs)
1015 bind | [dict_irred_id] <- dict_irred_ids
1016 = VarBind dict_irred_id rhs
1018 = PatBind { pat_lhs = lpat
1019 , pat_rhs = unguardedGRHSs rhs
1020 , pat_rhs_ty = hsLPatType lpat
1021 , bind_fvs = placeHolderNames
1024 ; traceTc $ text "makeImplicationBind" <+> ppr implic_inst
1025 ; return ([implic_inst], unitBag (L span bind))
1028 -----------------------------------------------------------
1029 tryHardCheckLoop :: SDoc
1031 -> TcM ([Inst], TcDictBinds)
1033 tryHardCheckLoop doc wanteds
1034 = do { (irreds,binds) <- checkLoop (mkInferRedEnv doc try_me) wanteds
1035 ; return (irreds,binds)
1039 -- Here's the try-hard bit
1041 -----------------------------------------------------------
1042 gentleCheckLoop :: InstLoc
1045 -> TcM ([Inst], TcDictBinds)
1047 gentleCheckLoop inst_loc givens wanteds
1048 = do { (irreds,binds) <- checkLoop env wanteds
1049 ; return (irreds,binds)
1052 env = mkRedEnv (pprInstLoc inst_loc) try_me givens
1054 try_me inst | isMethodOrLit inst = ReduceMe
1056 -- When checking against a given signature
1057 -- we MUST be very gentle: Note [Check gently]
1059 gentleInferLoop :: SDoc -> [Inst]
1060 -> TcM ([Inst], TcDictBinds)
1061 gentleInferLoop doc wanteds
1062 = do { (irreds, binds) <- checkLoop env wanteds
1063 ; return (irreds, binds) }
1065 env = mkInferRedEnv doc try_me
1066 try_me inst | isMethodOrLit inst = ReduceMe
1071 ~~~~~~~~~~~~~~~~~~~~
1072 We have to very careful about not simplifying too vigorously
1077 f :: Show b => T b -> b
1078 f (MkT x) = show [x]
1080 Inside the pattern match, which binds (a:*, x:a), we know that
1082 Hence we have a dictionary for Show [a] available; and indeed we
1083 need it. We are going to build an implication contraint
1084 forall a. (b~[a]) => Show [a]
1085 Later, we will solve this constraint using the knowledge (Show b)
1087 But we MUST NOT reduce (Show [a]) to (Show a), else the whole
1088 thing becomes insoluble. So we simplify gently (get rid of literals
1089 and methods only, plus common up equal things), deferring the real
1090 work until top level, when we solve the implication constraint
1091 with tryHardCheckLooop.
1095 -----------------------------------------------------------
1098 -> TcM ([Inst], TcDictBinds)
1099 -- Precondition: givens are completely rigid
1100 -- Postcondition: returned Insts are zonked
1102 checkLoop env wanteds
1104 where go env wanteds
1105 = do { -- We do need to zonk the givens; cf Note [Zonking RedEnv]
1106 ; env' <- zonkRedEnv env
1107 ; wanteds' <- zonkInsts wanteds
1109 ; (improved, binds, irreds) <- reduceContext env' wanteds'
1111 ; if null irreds || not improved then
1112 return (irreds, binds)
1115 -- If improvement did some unification, we go round again.
1116 -- We start again with irreds, not wanteds
1117 -- Using an instance decl might have introduced a fresh type
1118 -- variable which might have been unified, so we'd get an
1119 -- infinite loop if we started again with wanteds!
1121 { (irreds1, binds1) <- go env' irreds
1122 ; return (irreds1, binds `unionBags` binds1) } }
1125 Note [Zonking RedEnv]
1126 ~~~~~~~~~~~~~~~~~~~~~
1127 It might appear as if the givens in RedEnv are always rigid, but that is not
1128 necessarily the case for programs involving higher-rank types that have class
1129 contexts constraining the higher-rank variables. An example from tc237 in the
1132 class Modular s a | s -> a
1134 wim :: forall a w. Integral a
1135 => a -> (forall s. Modular s a => M s w) -> w
1136 wim i k = error "urk"
1138 test5 :: (Modular s a, Integral a) => M s a
1141 test4 = wim 4 test4'
1143 Notice how the variable 'a' of (Modular s a) in the rank-2 type of wim is
1144 quantified further outside. When type checking test4, we have to check
1145 whether the signature of test5 is an instance of
1147 (forall s. Modular s a => M s w)
1149 Consequently, we will get (Modular s t_a), where t_a is a TauTv into the
1152 Given the FD of Modular in this example, class improvement will instantiate
1153 t_a to 'a', where 'a' is the skolem from test5's signatures (due to the
1154 Modular s a predicate in that signature). If we don't zonk (Modular s t_a) in
1155 the givens, we will get into a loop as improveOne uses the unification engine
1156 Unify.tcUnifyTys, which doesn't know about mutable type variables.
1161 class If b t e r | b t e -> r
1164 class Lte a b c | a b -> c where lte :: a -> b -> c
1166 instance (Lte a b l,If l b a c) => Max a b c
1168 Wanted: Max Z (S x) y
1170 Then we'll reduce using the Max instance to:
1171 (Lte Z (S x) l, If l (S x) Z y)
1172 and improve by binding l->T, after which we can do some reduction
1173 on both the Lte and If constraints. What we *can't* do is start again
1174 with (Max Z (S x) y)!
1178 %************************************************************************
1180 tcSimplifySuperClasses
1182 %************************************************************************
1184 Note [SUPERCLASS-LOOP 1]
1185 ~~~~~~~~~~~~~~~~~~~~~~~~
1186 We have to be very, very careful when generating superclasses, lest we
1187 accidentally build a loop. Here's an example:
1191 class S a => C a where { opc :: a -> a }
1192 class S b => D b where { opd :: b -> b }
1194 instance C Int where
1197 instance D Int where
1200 From (instance C Int) we get the constraint set {ds1:S Int, dd:D Int}
1201 Simplifying, we may well get:
1202 $dfCInt = :C ds1 (opd dd)
1205 Notice that we spot that we can extract ds1 from dd.
1207 Alas! Alack! We can do the same for (instance D Int):
1209 $dfDInt = :D ds2 (opc dc)
1213 And now we've defined the superclass in terms of itself.
1214 Two more nasty cases are in
1219 - Satisfy the superclass context *all by itself*
1220 (tcSimplifySuperClasses)
1221 - And do so completely; i.e. no left-over constraints
1222 to mix with the constraints arising from method declarations
1225 Note [Recursive instances and superclases]
1226 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1227 Consider this code, which arises in the context of "Scrap Your
1228 Boilerplate with Class".
1232 instance Sat (ctx Char) => Data ctx Char
1233 instance (Sat (ctx [a]), Data ctx a) => Data ctx [a]
1235 class Data Maybe a => Foo a
1237 instance Foo t => Sat (Maybe t)
1239 instance Data Maybe a => Foo a
1240 instance Foo a => Foo [a]
1243 In the instance for Foo [a], when generating evidence for the superclasses
1244 (ie in tcSimplifySuperClasses) we need a superclass (Data Maybe [a]).
1245 Using the instance for Data, we therefore need
1246 (Sat (Maybe [a], Data Maybe a)
1247 But we are given (Foo a), and hence its superclass (Data Maybe a).
1248 So that leaves (Sat (Maybe [a])). Using the instance for Sat means
1249 we need (Foo [a]). And that is the very dictionary we are bulding
1250 an instance for! So we must put that in the "givens". So in this
1252 Given: Foo a, Foo [a]
1253 Watend: Data Maybe [a]
1255 BUT we must *not not not* put the *superclasses* of (Foo [a]) in
1256 the givens, which is what 'addGiven' would normally do. Why? Because
1257 (Data Maybe [a]) is the superclass, so we'd "satisfy" the wanted
1258 by selecting a superclass from Foo [a], which simply makes a loop.
1260 On the other hand we *must* put the superclasses of (Foo a) in
1261 the givens, as you can see from the derivation described above.
1263 Conclusion: in the very special case of tcSimplifySuperClasses
1264 we have one 'given' (namely the "this" dictionary) whose superclasses
1265 must not be added to 'givens' by addGiven. That is the *whole* reason
1266 for the red_given_scs field in RedEnv, and the function argument to
1270 tcSimplifySuperClasses
1272 -> Inst -- The dict whose superclasses
1273 -- are being figured out
1277 tcSimplifySuperClasses loc this givens sc_wanteds
1278 = do { traceTc (text "tcSimplifySuperClasses")
1279 ; (irreds,binds1) <- checkLoop env sc_wanteds
1280 ; let (tidy_env, tidy_irreds) = tidyInsts irreds
1281 ; reportNoInstances tidy_env (Just (loc, givens)) tidy_irreds
1284 env = RedEnv { red_doc = pprInstLoc loc,
1285 red_try_me = try_me,
1286 red_givens = this:givens,
1287 red_given_scs = add_scs,
1289 red_improve = False } -- No unification vars
1290 add_scs g | g==this = NoSCs
1291 | otherwise = AddSCs
1293 try_me _ = ReduceMe -- Try hard, so we completely solve the superclass
1294 -- constraints right here. See Note [SUPERCLASS-LOOP 1]
1298 %************************************************************************
1300 \subsection{tcSimplifyRestricted}
1302 %************************************************************************
1304 tcSimplifyRestricted infers which type variables to quantify for a
1305 group of restricted bindings. This isn't trivial.
1308 We want to quantify over a to get id :: forall a. a->a
1311 We do not want to quantify over a, because there's an Eq a
1312 constraint, so we get eq :: a->a->Bool (notice no forall)
1315 RHS has type 'tau', whose free tyvars are tau_tvs
1316 RHS has constraints 'wanteds'
1319 Quantify over (tau_tvs \ ftvs(wanteds))
1320 This is bad. The constraints may contain (Monad (ST s))
1321 where we have instance Monad (ST s) where...
1322 so there's no need to be monomorphic in s!
1324 Also the constraint might be a method constraint,
1325 whose type mentions a perfectly innocent tyvar:
1326 op :: Num a => a -> b -> a
1327 Here, b is unconstrained. A good example would be
1329 We want to infer the polymorphic type
1330 foo :: forall b. b -> b
1333 Plan B (cunning, used for a long time up to and including GHC 6.2)
1334 Step 1: Simplify the constraints as much as possible (to deal
1335 with Plan A's problem). Then set
1336 qtvs = tau_tvs \ ftvs( simplify( wanteds ) )
1338 Step 2: Now simplify again, treating the constraint as 'free' if
1339 it does not mention qtvs, and trying to reduce it otherwise.
1340 The reasons for this is to maximise sharing.
1342 This fails for a very subtle reason. Suppose that in the Step 2
1343 a constraint (Foo (Succ Zero) (Succ Zero) b) gets thrown upstairs as 'free'.
1344 In the Step 1 this constraint might have been simplified, perhaps to
1345 (Foo Zero Zero b), AND THEN THAT MIGHT BE IMPROVED, to bind 'b' to 'T'.
1346 This won't happen in Step 2... but that in turn might prevent some other
1347 constraint (Baz [a] b) being simplified (e.g. via instance Baz [a] T where {..})
1348 and that in turn breaks the invariant that no constraints are quantified over.
1350 Test typecheck/should_compile/tc177 (which failed in GHC 6.2) demonstrates
1355 Step 1: Simplify the constraints as much as possible (to deal
1356 with Plan A's problem). Then set
1357 qtvs = tau_tvs \ ftvs( simplify( wanteds ) )
1358 Return the bindings from Step 1.
1361 A note about Plan C (arising from "bug" reported by George Russel March 2004)
1364 instance (HasBinary ty IO) => HasCodedValue ty
1366 foo :: HasCodedValue a => String -> IO a
1368 doDecodeIO :: HasCodedValue a => () -> () -> IO a
1369 doDecodeIO codedValue view
1370 = let { act = foo "foo" } in act
1372 You might think this should work becuase the call to foo gives rise to a constraint
1373 (HasCodedValue t), which can be satisfied by the type sig for doDecodeIO. But the
1374 restricted binding act = ... calls tcSimplifyRestricted, and PlanC simplifies the
1375 constraint using the (rather bogus) instance declaration, and now we are stuffed.
1377 I claim this is not really a bug -- but it bit Sergey as well as George. So here's
1381 Plan D (a variant of plan B)
1382 Step 1: Simplify the constraints as much as possible (to deal
1383 with Plan A's problem), BUT DO NO IMPROVEMENT. Then set
1384 qtvs = tau_tvs \ ftvs( simplify( wanteds ) )
1386 Step 2: Now simplify again, treating the constraint as 'free' if
1387 it does not mention qtvs, and trying to reduce it otherwise.
1389 The point here is that it's generally OK to have too few qtvs; that is,
1390 to make the thing more monomorphic than it could be. We don't want to
1391 do that in the common cases, but in wierd cases it's ok: the programmer
1392 can always add a signature.
1394 Too few qtvs => too many wanteds, which is what happens if you do less
1399 tcSimplifyRestricted -- Used for restricted binding groups
1400 -- i.e. ones subject to the monomorphism restriction
1403 -> [Name] -- Things bound in this group
1404 -> TcTyVarSet -- Free in the type of the RHSs
1405 -> [Inst] -- Free in the RHSs
1406 -> TcM ([TyVar], -- Tyvars to quantify (zonked and quantified)
1407 TcDictBinds) -- Bindings
1408 -- tcSimpifyRestricted returns no constraints to
1409 -- quantify over; by definition there are none.
1410 -- They are all thrown back in the LIE
1412 tcSimplifyRestricted doc top_lvl bndrs tau_tvs wanteds
1413 -- Zonk everything in sight
1414 = do { traceTc (text "tcSimplifyRestricted")
1415 ; wanteds_z <- zonkInsts wanteds
1417 -- 'ReduceMe': Reduce as far as we can. Don't stop at
1418 -- dicts; the idea is to get rid of as many type
1419 -- variables as possible, and we don't want to stop
1420 -- at (say) Monad (ST s), because that reduces
1421 -- immediately, with no constraint on s.
1423 -- BUT do no improvement! See Plan D above
1424 -- HOWEVER, some unification may take place, if we instantiate
1425 -- a method Inst with an equality constraint
1426 ; let env = mkNoImproveRedEnv doc (\_ -> ReduceMe)
1427 ; (_imp, _binds, constrained_dicts) <- reduceContext env wanteds_z
1429 -- Next, figure out the tyvars we will quantify over
1430 ; tau_tvs' <- zonkTcTyVarsAndFV (varSetElems tau_tvs)
1431 ; gbl_tvs' <- tcGetGlobalTyVars
1432 ; constrained_dicts' <- zonkInsts constrained_dicts
1434 ; let qtvs1 = tau_tvs' `minusVarSet` oclose (fdPredsOfInsts constrained_dicts) gbl_tvs'
1435 -- As in tcSimplifyInfer
1437 -- Do not quantify over constrained type variables:
1438 -- this is the monomorphism restriction
1439 constrained_tvs' = tyVarsOfInsts constrained_dicts'
1440 qtvs = qtvs1 `minusVarSet` constrained_tvs'
1441 pp_bndrs = pprWithCommas (quotes . ppr) bndrs
1444 ; warn_mono <- doptM Opt_WarnMonomorphism
1445 ; warnTc (warn_mono && (constrained_tvs' `intersectsVarSet` qtvs1))
1446 (vcat[ ptext (sLit "the Monomorphism Restriction applies to the binding")
1447 <> plural bndrs <+> ptext (sLit "for") <+> pp_bndrs,
1448 ptext (sLit "Consider giving a type signature for") <+> pp_bndrs])
1450 ; traceTc (text "tcSimplifyRestricted" <+> vcat [
1451 pprInsts wanteds, pprInsts constrained_dicts',
1453 ppr constrained_tvs', ppr tau_tvs', ppr qtvs ])
1455 -- Zonk wanteds again! The first call to reduceContext may have
1456 -- instantiated some variables.
1457 -- FIXME: If red_improve would work, we could propagate that into
1458 -- the equality solver, too, to prevent instantating any
1460 ; wanteds_zz <- zonkInsts wanteds_z
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 becuase 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, binds, irreds) <- reduceContext env wanteds_zz
1485 -- See "Notes on implicit parameters, Question 4: top level"
1486 ; ASSERT( all (isFreeWrtTyVars qtvs) irreds ) -- None should be captured
1487 if is_nested_group then
1489 else do { let (bad_ips, non_ips) = partition isIPDict irreds
1490 ; addTopIPErrs bndrs bad_ips
1491 ; extendLIEs non_ips }
1493 ; qtvs' <- zonkQuantifiedTyVars (varSetElems qtvs)
1494 ; return (qtvs', binds) }
1498 %************************************************************************
1502 %************************************************************************
1504 On the LHS of transformation rules we only simplify methods and constants,
1505 getting dictionaries. We want to keep all of them unsimplified, to serve
1506 as the available stuff for the RHS of the rule.
1508 Example. Consider the following left-hand side of a rule
1510 f (x == y) (y > z) = ...
1512 If we typecheck this expression we get constraints
1514 d1 :: Ord a, d2 :: Eq a
1516 We do NOT want to "simplify" to the LHS
1518 forall x::a, y::a, z::a, d1::Ord a.
1519 f ((==) (eqFromOrd d1) x y) ((>) d1 y z) = ...
1523 forall x::a, y::a, z::a, d1::Ord a, d2::Eq a.
1524 f ((==) d2 x y) ((>) d1 y z) = ...
1526 Here is another example:
1528 fromIntegral :: (Integral a, Num b) => a -> b
1529 {-# RULES "foo" fromIntegral = id :: Int -> Int #-}
1531 In the rule, a=b=Int, and Num Int is a superclass of Integral Int. But
1532 we *dont* want to get
1534 forall dIntegralInt.
1535 fromIntegral Int Int dIntegralInt (scsel dIntegralInt) = id Int
1537 because the scsel will mess up RULE matching. Instead we want
1539 forall dIntegralInt, dNumInt.
1540 fromIntegral Int Int dIntegralInt dNumInt = id Int
1544 g (x == y) (y == z) = ..
1546 where the two dictionaries are *identical*, we do NOT WANT
1548 forall x::a, y::a, z::a, d1::Eq a
1549 f ((==) d1 x y) ((>) d1 y z) = ...
1551 because that will only match if the dict args are (visibly) equal.
1552 Instead we want to quantify over the dictionaries separately.
1554 In short, tcSimplifyRuleLhs must *only* squash LitInst and MethInts, leaving
1555 all dicts unchanged, with absolutely no sharing. It's simpler to do this
1556 from scratch, rather than further parameterise simpleReduceLoop etc.
1557 Simpler, maybe, but alas not simple (see Trac #2494)
1559 * Type errors may give rise to an (unsatisfiable) equality constraint
1561 * Applications of a higher-rank function on the LHS may give
1562 rise to an implication constraint, esp if there are unsatisfiable
1563 equality constraints inside.
1566 tcSimplifyRuleLhs :: [Inst] -> TcM ([Inst], TcDictBinds)
1567 tcSimplifyRuleLhs wanteds
1568 = do { wanteds' <- zonkInsts wanteds
1569 ; (irreds, binds) <- go [] emptyBag wanteds'
1570 ; let (dicts, bad_irreds) = partition isDict irreds
1571 ; traceTc (text "tcSimplifyrulelhs" <+> pprInsts bad_irreds)
1572 ; addNoInstanceErrs (nub bad_irreds)
1573 -- The nub removes duplicates, which has
1574 -- not happened otherwise (see notes above)
1575 ; return (dicts, binds) }
1577 go :: [Inst] -> TcDictBinds -> [Inst] -> TcM ([Inst], TcDictBinds)
1579 = return (irreds, binds)
1580 go irreds binds (w:ws)
1582 = go (w:irreds) binds ws
1583 | isImplicInst w -- Have a go at reducing the implication
1584 = do { (binds1, irreds1) <- reduceImplication red_env w
1585 ; let (bad_irreds, ok_irreds) = partition isImplicInst irreds1
1586 ; go (bad_irreds ++ irreds)
1587 (binds `unionBags` binds1)
1590 = do { w' <- zonkInst w -- So that (3::Int) does not generate a call
1591 -- to fromInteger; this looks fragile to me
1592 ; lookup_result <- lookupSimpleInst w'
1593 ; case lookup_result of
1594 NoInstance -> go (w:irreds) binds ws
1595 GenInst ws' rhs -> go irreds binds' (ws' ++ ws)
1597 binds' = addInstToDictBind binds w rhs
1600 -- Sigh: we need to reduce inside implications
1601 red_env = mkInferRedEnv doc try_me
1602 doc = ptext (sLit "Implication constraint in RULE lhs")
1603 try_me inst | isMethodOrLit inst = ReduceMe
1604 | otherwise = Stop -- Be gentle
1607 tcSimplifyBracket is used when simplifying the constraints arising from
1608 a Template Haskell bracket [| ... |]. We want to check that there aren't
1609 any constraints that can't be satisfied (e.g. Show Foo, where Foo has no
1610 Show instance), but we aren't otherwise interested in the results.
1611 Nor do we care about ambiguous dictionaries etc. We will type check
1612 this bracket again at its usage site.
1615 tcSimplifyBracket :: [Inst] -> TcM ()
1616 tcSimplifyBracket wanteds
1617 = do { tryHardCheckLoop doc wanteds
1620 doc = text "tcSimplifyBracket"
1624 %************************************************************************
1626 \subsection{Filtering at a dynamic binding}
1628 %************************************************************************
1633 we must discharge all the ?x constraints from B. We also do an improvement
1634 step; if we have ?x::t1 and ?x::t2 we must unify t1, t2.
1636 Actually, the constraints from B might improve the types in ?x. For example
1638 f :: (?x::Int) => Char -> Char
1641 then the constraint (?x::Int) arising from the call to f will
1642 force the binding for ?x to be of type Int.
1645 tcSimplifyIPs :: [Inst] -- The implicit parameters bound here
1648 -- We need a loop so that we do improvement, and then
1649 -- (next time round) generate a binding to connect the two
1651 -- Here the two ?x's have different types, and improvement
1652 -- makes them the same.
1654 tcSimplifyIPs given_ips wanteds
1655 = do { wanteds' <- zonkInsts wanteds
1656 ; given_ips' <- zonkInsts given_ips
1657 -- Unusually for checking, we *must* zonk the given_ips
1659 ; let env = mkRedEnv doc try_me given_ips'
1660 ; (improved, binds, irreds) <- reduceContext env wanteds'
1662 ; if null irreds || not improved then
1663 ASSERT( all is_free irreds )
1664 do { extendLIEs irreds
1667 -- If improvement did some unification, we go round again.
1668 -- We start again with irreds, not wanteds
1669 -- Using an instance decl might have introduced a fresh type
1670 -- variable which might have been unified, so we'd get an
1671 -- infinite loop if we started again with wanteds!
1673 { binds1 <- tcSimplifyIPs given_ips' irreds
1674 ; return $ binds `unionBags` binds1
1677 doc = text "tcSimplifyIPs" <+> ppr given_ips
1678 ip_set = mkNameSet (ipNamesOfInsts given_ips)
1679 is_free inst = isFreeWrtIPs ip_set inst
1681 -- Simplify any methods that mention the implicit parameter
1682 try_me inst | is_free inst = Stop
1683 | otherwise = ReduceMe
1687 %************************************************************************
1689 \subsection[binds-for-local-funs]{@bindInstsOfLocalFuns@}
1691 %************************************************************************
1693 When doing a binding group, we may have @Insts@ of local functions.
1694 For example, we might have...
1696 let f x = x + 1 -- orig local function (overloaded)
1697 f.1 = f Int -- two instances of f
1702 The point is: we must drop the bindings for @f.1@ and @f.2@ here,
1703 where @f@ is in scope; those @Insts@ must certainly not be passed
1704 upwards towards the top-level. If the @Insts@ were binding-ified up
1705 there, they would have unresolvable references to @f@.
1707 We pass in an @init_lie@ of @Insts@ and a list of locally-bound @Ids@.
1708 For each method @Inst@ in the @init_lie@ that mentions one of the
1709 @Ids@, we create a binding. We return the remaining @Insts@ (in an
1710 @LIE@), as well as the @HsBinds@ generated.
1713 bindInstsOfLocalFuns :: [Inst] -> [TcId] -> TcM TcDictBinds
1714 -- Simlifies only MethodInsts, and generate only bindings of form
1716 -- We're careful not to even generate bindings of the form
1718 -- You'd think that'd be fine, but it interacts with what is
1719 -- arguably a bug in Match.tidyEqnInfo (see notes there)
1721 bindInstsOfLocalFuns wanteds local_ids
1722 | null overloaded_ids = do
1725 return emptyLHsBinds
1728 = do { (irreds, binds) <- gentleInferLoop doc for_me
1729 ; extendLIEs not_for_me
1733 doc = text "bindInsts" <+> ppr local_ids
1734 overloaded_ids = filter is_overloaded local_ids
1735 is_overloaded id = isOverloadedTy (idType id)
1736 (for_me, not_for_me) = partition (isMethodFor overloaded_set) wanteds
1738 overloaded_set = mkVarSet overloaded_ids -- There can occasionally be a lot of them
1739 -- so it's worth building a set, so that
1740 -- lookup (in isMethodFor) is faster
1744 %************************************************************************
1746 \subsection{Data types for the reduction mechanism}
1748 %************************************************************************
1750 The main control over context reduction is here
1754 = RedEnv { red_doc :: SDoc -- The context
1755 , red_try_me :: Inst -> WhatToDo
1756 , red_improve :: Bool -- True <=> do improvement
1757 , red_givens :: [Inst] -- All guaranteed rigid
1758 -- Always dicts & equalities
1759 -- but see Note [Rigidity]
1761 , red_given_scs :: Inst -> WantSCs -- See Note [Recursive instances and superclases]
1763 , red_stack :: (Int, [Inst]) -- Recursion stack (for err msg)
1764 -- See Note [RedStack]
1768 -- The red_givens are rigid so far as cmpInst is concerned.
1769 -- There is one case where they are not totally rigid, namely in tcSimplifyIPs
1770 -- let ?x = e in ...
1771 -- Here, the given is (?x::a), where 'a' is not necy a rigid type
1772 -- But that doesn't affect the comparison, which is based only on mame.
1775 -- The red_stack pair (n,insts) pair is just used for error reporting.
1776 -- 'n' is always the depth of the stack.
1777 -- The 'insts' is the stack of Insts being reduced: to produce X
1778 -- I had to produce Y, to produce Y I had to produce Z, and so on.
1781 mkRedEnv :: SDoc -> (Inst -> WhatToDo) -> [Inst] -> RedEnv
1782 mkRedEnv doc try_me givens
1783 = RedEnv { red_doc = doc, red_try_me = try_me,
1784 red_givens = givens,
1785 red_given_scs = const AddSCs,
1787 red_improve = True }
1789 mkInferRedEnv :: SDoc -> (Inst -> WhatToDo) -> RedEnv
1791 mkInferRedEnv doc try_me
1792 = RedEnv { red_doc = doc, red_try_me = try_me,
1794 red_given_scs = const AddSCs,
1796 red_improve = True }
1798 mkNoImproveRedEnv :: SDoc -> (Inst -> WhatToDo) -> RedEnv
1799 -- Do not do improvement; no givens
1800 mkNoImproveRedEnv doc try_me
1801 = RedEnv { red_doc = doc, red_try_me = try_me,
1803 red_given_scs = const AddSCs,
1805 red_improve = True }
1808 = ReduceMe -- Try to reduce this
1809 -- If there's no instance, add the inst to the
1810 -- irreductible ones, but don't produce an error
1811 -- message of any kind.
1812 -- It might be quite legitimate such as (Eq a)!
1814 | Stop -- Return as irreducible unless it can
1815 -- be reduced to a constant in one step
1816 -- Do not add superclasses; see
1818 data WantSCs = NoSCs | AddSCs -- Tells whether we should add the superclasses
1819 -- of a predicate when adding it to the avails
1820 -- The reason for this flag is entirely the super-class loop problem
1821 -- Note [SUPER-CLASS LOOP 1]
1823 zonkRedEnv :: RedEnv -> TcM RedEnv
1825 = do { givens' <- mapM zonkInst (red_givens env)
1826 ; return $ env {red_givens = givens'}
1831 %************************************************************************
1833 \subsection[reduce]{@reduce@}
1835 %************************************************************************
1837 Note [Ancestor Equalities]
1838 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1839 During context reduction, we add to the wanted equalities also those
1840 equalities that (transitively) occur in superclass contexts of wanted
1841 class constraints. Consider the following code
1843 class a ~ Int => C a
1846 If (C a) is wanted, we want to add (a ~ Int), which will be discharged by
1847 substituting Int for a. Hence, we ultimately want (C Int), which we
1848 discharge with the explicit instance.
1851 reduceContext :: RedEnv
1853 -> TcM (ImprovementDone,
1854 TcDictBinds, -- Dictionary bindings
1855 [Inst]) -- Irreducible
1857 reduceContext env wanteds0
1858 = do { traceTc (text "reduceContext" <+> (vcat [
1859 text "----------------------",
1861 text "given" <+> ppr (red_givens env),
1862 text "wanted" <+> ppr wanteds0,
1863 text "----------------------"
1866 -- We want to add as wanted equalities those that (transitively)
1867 -- occur in superclass contexts of wanted class constraints.
1868 -- See Note [Ancestor Equalities]
1869 ; ancestor_eqs <- ancestorEqualities wanteds0
1870 ; traceTc $ text "reduceContext: ancestor eqs" <+> ppr ancestor_eqs
1872 -- Normalise and solve all equality constraints as far as possible
1873 -- and normalise all dictionary constraints wrt to the reduced
1874 -- equalities. The returned wanted constraints include the
1875 -- irreducible wanted equalities.
1876 ; let wanteds = wanteds0 ++ ancestor_eqs
1877 givens = red_givens env
1881 eq_improved) <- tcReduceEqs givens wanteds
1882 ; traceTc $ text "reduceContext: tcReduceEqs result" <+> vcat
1883 [ppr givens', ppr wanteds', ppr normalise_binds]
1885 -- Build the Avail mapping from "given_dicts"
1886 ; (init_state, _) <- getLIE $ do
1887 { init_state <- foldlM (addGiven (red_given_scs env))
1892 -- Solve the *wanted* *dictionary* constraints (not implications)
1893 -- This may expose some further equational constraints in the course
1894 -- of improvement due to functional dependencies if any of the
1895 -- involved unifications gets deferred.
1896 ; let (wanted_implics, wanted_dicts) = partition isImplicInst wanteds'
1897 ; (avails, extra_eqs) <- getLIE (reduceList env wanted_dicts init_state)
1898 -- The getLIE is reqd because reduceList does improvement
1899 -- (via extendAvails) which may in turn do unification
1902 dict_irreds) <- extractResults avails wanted_dicts
1903 ; traceTc $ text "reduceContext: extractResults" <+> vcat
1904 [ppr avails, ppr wanted_dicts, ppr dict_binds]
1906 -- Solve the wanted *implications*. In doing so, we can provide
1907 -- as "given" all the dicts that were originally given,
1908 -- *or* for which we now have bindings,
1909 -- *or* which are now irreds
1910 -- NB: Equality irreds need to be converted, as the recursive
1911 -- invocation of the solver will still treat them as wanteds
1913 ; let implic_env = env { red_givens
1914 = givens ++ bound_dicts ++
1915 map wantedToLocalEqInst dict_irreds }
1916 ; (implic_binds_s, implic_irreds_s)
1917 <- mapAndUnzipM (reduceImplication implic_env) wanted_implics
1918 ; let implic_binds = unionManyBags implic_binds_s
1919 implic_irreds = concat implic_irreds_s
1921 -- Collect all irreducible instances, and determine whether we should
1922 -- go round again. We do so in either of two cases:
1923 -- (1) If dictionary reduction or equality solving led to
1924 -- improvement (i.e., instantiated type variables).
1925 -- (2) If we reduced dictionaries (i.e., got dictionary bindings),
1926 -- they may have exposed further opportunities to normalise
1927 -- family applications. See Note [Dictionary Improvement]
1929 -- NB: We do *not* go around for new extra_eqs. Morally, we should,
1930 -- but we can't without risking non-termination (see #2688). By
1931 -- not going around, we miss some legal programs mixing FDs and
1932 -- TFs, but we never claimed to support such programs in the
1933 -- current implementation anyway.
1935 ; let all_irreds = dict_irreds ++ implic_irreds ++ extra_eqs
1936 avails_improved = availsImproved avails
1937 improvedFlexible = avails_improved || eq_improved
1938 reduced_dicts = not (isEmptyBag dict_binds)
1939 improved = improvedFlexible || reduced_dicts
1941 improvedHint = (if avails_improved then " [AVAILS]" else "") ++
1942 (if eq_improved then " [EQ]" else "")
1944 ; traceTc (text "reduceContext end" <+> (vcat [
1945 text "----------------------",
1947 text "given" <+> ppr givens,
1948 text "wanted" <+> ppr wanteds0,
1950 text "avails" <+> pprAvails avails,
1951 text "improved =" <+> ppr improved <+> text improvedHint,
1952 text "(all) irreds = " <+> ppr all_irreds,
1953 text "dict-binds = " <+> ppr dict_binds,
1954 text "implic-binds = " <+> ppr implic_binds,
1955 text "----------------------"
1959 normalise_binds `unionBags` dict_binds
1960 `unionBags` implic_binds,
1964 tcImproveOne :: Avails -> Inst -> TcM ImprovementDone
1965 tcImproveOne avails inst
1966 | not (isDict inst) = return False
1968 = do { inst_envs <- tcGetInstEnvs
1969 ; let eqns = improveOne (classInstances inst_envs)
1970 (dictPred inst, pprInstArising inst)
1971 [ (dictPred p, pprInstArising p)
1972 | p <- availsInsts avails, isDict p ]
1973 -- Avails has all the superclasses etc (good)
1974 -- It also has all the intermediates of the deduction (good)
1975 -- It does not have duplicates (good)
1976 -- NB that (?x::t1) and (?x::t2) will be held separately in
1977 -- avails so that improve will see them separate
1978 ; traceTc (text "improveOne" <+> ppr inst)
1981 unifyEqns :: [(Equation, (PredType, SDoc), (PredType, SDoc))]
1982 -> TcM ImprovementDone
1983 unifyEqns [] = return False
1985 = do { traceTc (ptext (sLit "Improve:") <+> vcat (map pprEquationDoc eqns))
1986 ; improved <- mapM unify eqns
1987 ; return $ or improved
1990 unify ((qtvs, pairs), what1, what2)
1991 = addErrCtxtM (mkEqnMsg what1 what2) $
1992 do { let freeTyVars = unionVarSets (map tvs_pr pairs)
1994 ; (_, _, tenv) <- tcInstTyVars (varSetElems qtvs)
1995 ; mapM_ (unif_pr tenv) pairs
1996 ; anyM isFilledMetaTyVar $ varSetElems freeTyVars
1999 unif_pr tenv (ty1, ty2) = unifyType (substTy tenv ty1) (substTy tenv ty2)
2001 tvs_pr (ty1, ty2) = tyVarsOfType ty1 `unionVarSet` tyVarsOfType ty2
2003 pprEquationDoc :: (Equation, (PredType, SDoc), (PredType, SDoc)) -> SDoc
2004 pprEquationDoc (eqn, (p1, _), (p2, _))
2005 = vcat [pprEquation eqn, nest 2 (ppr p1), nest 2 (ppr p2)]
2007 mkEqnMsg :: (TcPredType, SDoc) -> (TcPredType, SDoc) -> TidyEnv
2008 -> TcM (TidyEnv, SDoc)
2009 mkEqnMsg (pred1,from1) (pred2,from2) tidy_env
2010 = do { pred1' <- zonkTcPredType pred1
2011 ; pred2' <- zonkTcPredType pred2
2012 ; let { pred1'' = tidyPred tidy_env pred1'
2013 ; pred2'' = tidyPred tidy_env pred2' }
2014 ; let msg = vcat [ptext (sLit "When using functional dependencies to combine"),
2015 nest 2 (sep [ppr pred1'' <> comma, nest 2 from1]),
2016 nest 2 (sep [ppr pred2'' <> comma, nest 2 from2])]
2017 ; return (tidy_env, msg) }
2020 Note [Dictionary Improvement]
2021 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2022 In reduceContext, we first reduce equalities and then class constraints.
2023 However, the letter may expose further opportunities for the former. Hence,
2024 we need to go around again if dictionary reduction produced any dictionary
2025 bindings. The following example demonstrated the point:
2027 data EX _x _y (p :: * -> *)
2032 class Base (Def p) => Prop p where
2036 instance Prop () where
2039 instance (Base (Def (p ANY))) => Base (EX _x _y p)
2040 instance (Prop (p ANY)) => Prop (EX _x _y p) where
2041 type Def (EX _x _y p) = EX _x _y p
2044 instance Prop (FOO x) where
2045 type Def (FOO x) = ()
2048 instance Prop BAR where
2049 type Def BAR = EX () () FOO
2051 During checking the last instance declaration, we need to check the superclass
2052 cosntraint Base (Def BAR), which family normalisation reduced to
2053 Base (EX () () FOO). Chasing the instance for Base (EX _x _y p), gives us
2054 Base (Def (FOO ANY)), which again requires family normalisation of Def to
2055 Base () before we can finish.
2058 The main context-reduction function is @reduce@. Here's its game plan.
2061 reduceList :: RedEnv -> [Inst] -> Avails -> TcM Avails
2062 reduceList env@(RedEnv {red_stack = (n,stk)}) wanteds state
2063 = do { traceTc (text "reduceList " <+> (ppr wanteds $$ ppr state))
2065 ; when (debugIsOn && (n > 8)) $ do
2066 debugDumpTcRn (hang (ptext (sLit "Interesting! Context reduction stack depth") <+> int n)
2067 2 (ifPprDebug (nest 2 (pprStack stk))))
2068 ; if n >= ctxtStkDepth dopts then
2069 failWithTc (reduceDepthErr n stk)
2073 go [] state = return state
2074 go (w:ws) state = do { state' <- reduce (env {red_stack = (n+1, w:stk)}) w state
2077 -- Base case: we're done!
2078 reduce :: RedEnv -> Inst -> Avails -> TcM Avails
2079 reduce env wanted avails
2081 -- We don't reduce equalities here (and they must not end up as irreds
2086 -- It's the same as an existing inst, or a superclass thereof
2087 | Just _ <- findAvail avails wanted
2088 = do { traceTc (text "reduce: found " <+> ppr wanted)
2093 = do { traceTc (text "reduce" <+> ppr wanted $$ ppr avails)
2094 ; case red_try_me env wanted of {
2095 Stop -> try_simple (addIrred NoSCs);
2096 -- See Note [No superclasses for Stop]
2098 ReduceMe -> do -- It should be reduced
2099 { (avails, lookup_result) <- reduceInst env avails wanted
2100 ; case lookup_result of
2101 NoInstance -> addIrred AddSCs avails wanted
2102 -- Add it and its superclasses
2104 GenInst [] rhs -> addWanted AddSCs avails wanted rhs []
2106 GenInst wanteds' rhs
2107 -> do { avails1 <- addIrred NoSCs avails wanted
2108 ; avails2 <- reduceList env wanteds' avails1
2109 ; addWanted AddSCs avails2 wanted rhs wanteds' } }
2110 -- Temporarily do addIrred *before* the reduceList,
2111 -- which has the effect of adding the thing we are trying
2112 -- to prove to the database before trying to prove the things it
2113 -- needs. See note [RECURSIVE DICTIONARIES]
2114 -- NB: we must not do an addWanted before, because that adds the
2115 -- superclasses too, and that can lead to a spurious loop; see
2116 -- the examples in [SUPERCLASS-LOOP]
2117 -- So we do an addIrred before, and then overwrite it afterwards with addWanted
2120 -- First, see if the inst can be reduced to a constant in one step
2121 -- Works well for literals (1::Int) and constant dictionaries (d::Num Int)
2122 -- Don't bother for implication constraints, which take real work
2123 try_simple do_this_otherwise
2124 = do { res <- lookupSimpleInst wanted
2126 GenInst [] rhs -> addWanted AddSCs avails wanted rhs []
2127 _ -> do_this_otherwise avails wanted }
2131 Note [RECURSIVE DICTIONARIES]
2132 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2134 data D r = ZeroD | SuccD (r (D r));
2136 instance (Eq (r (D r))) => Eq (D r) where
2137 ZeroD == ZeroD = True
2138 (SuccD a) == (SuccD b) = a == b
2141 equalDC :: D [] -> D [] -> Bool;
2144 We need to prove (Eq (D [])). Here's how we go:
2148 by instance decl, holds if
2152 by instance decl of Eq, holds if
2154 where d2 = dfEqList d3
2157 But now we can "tie the knot" to give
2163 and it'll even run! The trick is to put the thing we are trying to prove
2164 (in this case Eq (D []) into the database before trying to prove its
2165 contributing clauses.
2167 Note [SUPERCLASS-LOOP 2]
2168 ~~~~~~~~~~~~~~~~~~~~~~~~
2169 We need to be careful when adding "the constaint we are trying to prove".
2170 Suppose we are *given* d1:Ord a, and want to deduce (d2:C [a]) where
2172 class Ord a => C a where
2173 instance Ord [a] => C [a] where ...
2175 Then we'll use the instance decl to deduce C [a] from Ord [a], and then add the
2176 superclasses of C [a] to avails. But we must not overwrite the binding
2177 for Ord [a] (which is obtained from Ord a) with a superclass selection or we'll just
2180 Here's another variant, immortalised in tcrun020
2181 class Monad m => C1 m
2182 class C1 m => C2 m x
2183 instance C2 Maybe Bool
2184 For the instance decl we need to build (C1 Maybe), and it's no good if
2185 we run around and add (C2 Maybe Bool) and its superclasses to the avails
2186 before we search for C1 Maybe.
2188 Here's another example
2189 class Eq b => Foo a b
2190 instance Eq a => Foo [a] a
2194 we'll first deduce that it holds (via the instance decl). We must not
2195 then overwrite the Eq t constraint with a superclass selection!
2197 At first I had a gross hack, whereby I simply did not add superclass constraints
2198 in addWanted, though I did for addGiven and addIrred. This was sub-optimal,
2199 becuase it lost legitimate superclass sharing, and it still didn't do the job:
2200 I found a very obscure program (now tcrun021) in which improvement meant the
2201 simplifier got two bites a the cherry... so something seemed to be an Stop
2202 first time, but reducible next time.
2204 Now we implement the Right Solution, which is to check for loops directly
2205 when adding superclasses. It's a bit like the occurs check in unification.
2209 %************************************************************************
2211 Reducing a single constraint
2213 %************************************************************************
2216 ---------------------------------------------
2217 reduceInst :: RedEnv -> Avails -> Inst -> TcM (Avails, LookupInstResult)
2218 reduceInst _ avails other_inst
2219 = do { result <- lookupSimpleInst other_inst
2220 ; return (avails, result) }
2223 Note [Equational Constraints in Implication Constraints]
2224 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2226 An implication constraint is of the form
2228 where Given and Wanted may contain both equational and dictionary
2229 constraints. The delay and reduction of these two kinds of constraints
2232 -) In the generated code, wanted Dictionary constraints are wrapped up in an
2233 implication constraint that is created at the code site where the wanted
2234 dictionaries can be reduced via a let-binding. This let-bound implication
2235 constraint is deconstructed at the use-site of the wanted dictionaries.
2237 -) While the reduction of equational constraints is also delayed, the delay
2238 is not manifest in the generated code. The required evidence is generated
2239 in the code directly at the use-site. There is no let-binding and deconstruction
2240 necessary. The main disadvantage is that we cannot exploit sharing as the
2241 same evidence may be generated at multiple use-sites. However, this disadvantage
2242 is limited because it only concerns coercions which are erased.
2244 The different treatment is motivated by the different in representation. Dictionary
2245 constraints require manifest runtime dictionaries, while equations require coercions
2249 ---------------------------------------------
2250 reduceImplication :: RedEnv
2252 -> TcM (TcDictBinds, [Inst])
2255 Suppose we are simplifying the constraint
2256 forall bs. extras => wanted
2257 in the context of an overall simplification problem with givens 'givens'.
2260 * The 'givens' need not mention any of the quantified type variables
2261 e.g. forall {}. Eq a => Eq [a]
2262 forall {}. C Int => D (Tree Int)
2264 This happens when you have something like
2266 T1 :: Eq a => a -> T a
2269 f x = ...(case x of { T1 v -> v==v })...
2272 -- ToDo: should we instantiate tvs? I think it's not necessary
2274 -- Note on coercion variables:
2276 -- The extra given coercion variables are bound at two different
2279 -- -) in the creation context of the implication constraint
2280 -- the solved equational constraints use these binders
2282 -- -) at the solving site of the implication constraint
2283 -- the solved dictionaries use these binders;
2284 -- these binders are generated by reduceImplication
2286 -- Note [Binders for equalities]
2287 -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2288 -- To reuse the binders of local/given equalities in the binders of
2289 -- implication constraints, it is crucial that these given equalities
2290 -- always have the form
2292 -- where cotv is a simple coercion type variable (and not a more
2293 -- complex coercion term). We require that the extra_givens always
2294 -- have this form and exploit the special form when generating binders.
2295 reduceImplication env
2296 orig_implic@(ImplicInst { tci_name = name, tci_loc = inst_loc,
2298 tci_given = extra_givens, tci_wanted = wanteds
2300 = do { -- Solve the sub-problem
2301 ; let try_me _ = ReduceMe -- Note [Freeness and implications]
2302 env' = env { red_givens = extra_givens ++ red_givens env
2303 , red_doc = sep [ptext (sLit "reduceImplication for")
2305 nest 2 (parens $ ptext (sLit "within")
2307 , red_try_me = try_me }
2309 ; traceTc (text "reduceImplication" <+> vcat
2310 [ ppr (red_givens env), ppr extra_givens,
2312 ; (irreds, binds) <- checkLoop env' wanteds
2314 ; traceTc (text "reduceImplication result" <+> vcat
2315 [ppr irreds, ppr binds])
2317 ; -- extract superclass binds
2318 -- (sc_binds,_) <- extractResults avails []
2319 -- ; traceTc (text "reduceImplication sc_binds" <+> vcat
2320 -- [ppr sc_binds, ppr avails])
2323 -- SLPJ Sept 07: what if improvement happened inside the checkLoop?
2324 -- Then we must iterate the outer loop too!
2326 ; didntSolveWantedEqs <- allM wantedEqInstIsUnsolved wanteds
2327 -- we solve wanted eqs by side effect!
2329 -- Progress is no longer measered by the number of bindings
2330 -- If there are any irreds, but no bindings and no solved
2331 -- equalities, we back off and do nothing
2332 ; let backOff = isEmptyLHsBinds binds && -- no new bindings
2333 (not $ null irreds) && -- but still some irreds
2334 didntSolveWantedEqs -- no instantiated cotv
2336 ; if backOff then -- No progress
2337 return (emptyBag, [orig_implic])
2339 { (simpler_implic_insts, bind)
2340 <- makeImplicationBind inst_loc tvs extra_givens irreds
2341 -- This binding is useless if the recursive simplification
2342 -- made no progress; but currently we don't try to optimise that
2343 -- case. After all, we only try hard to reduce at top level, or
2344 -- when inferring types.
2346 ; let -- extract Id binders for dicts and CoTyVar binders for eqs;
2347 -- see Note [Binders for equalities]
2348 (extra_eq_givens, extra_dict_givens) = partition isEqInst
2350 eq_cotvs = map instToVar extra_eq_givens
2351 dict_ids = map instToId extra_dict_givens
2353 -- Note [Always inline implication constraints]
2354 wrap_inline | null dict_ids = idHsWrapper
2355 | otherwise = WpInline
2358 <.> mkWpTyLams eq_cotvs
2359 <.> mkWpLams dict_ids
2360 <.> WpLet (binds `unionBags` bind)
2361 rhs = mkLHsWrap co payload
2362 loc = instLocSpan inst_loc
2363 -- wanted equalities are solved by updating their
2364 -- cotv; we don't generate bindings for them
2365 dict_bndrs = map (L loc . HsVar . instToId)
2366 . filter (not . isEqInst)
2368 payload = mkBigLHsTup dict_bndrs
2371 ; traceTc (vcat [text "reduceImplication" <+> ppr name,
2372 ppr simpler_implic_insts,
2373 text "->" <+> ppr rhs])
2374 ; return (unitBag (L loc (VarBind (instToId orig_implic) rhs)),
2375 simpler_implic_insts)
2378 reduceImplication _ i = pprPanic "reduceImplication" (ppr i)
2381 Note [Always inline implication constraints]
2382 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2383 Suppose an implication constraint floats out of an INLINE function.
2384 Then although the implication has a single call site, it won't be
2385 inlined. And that is bad because it means that even if there is really
2386 *no* overloading (type signatures specify the exact types) there will
2387 still be dictionary passing in the resulting code. To avert this,
2388 we mark the implication constraints themselves as INLINE, at least when
2389 there is no loss of sharing as a result.
2391 Note [Freeness and implications]
2392 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2393 It's hard to say when an implication constraint can be floated out. Consider
2394 forall {} Eq a => Foo [a]
2395 The (Foo [a]) doesn't mention any of the quantified variables, but it
2396 still might be partially satisfied by the (Eq a).
2398 There is a useful special case when it *is* easy to partition the
2399 constraints, namely when there are no 'givens'. Consider
2400 forall {a}. () => Bar b
2401 There are no 'givens', and so there is no reason to capture (Bar b).
2402 We can let it float out. But if there is even one constraint we
2403 must be much more careful:
2404 forall {a}. C a b => Bar (m b)
2405 because (C a b) might have a superclass (D b), from which we might
2406 deduce (Bar [b]) when m later gets instantiated to []. Ha!
2408 Here is an even more exotic example
2410 Now consider the constraint
2411 forall b. D Int b => C Int
2412 We can satisfy the (C Int) from the superclass of D, so we don't want
2413 to float the (C Int) out, even though it mentions no type variable in
2416 One more example: the constraint
2418 instance (C a, E c) => E (a,c)
2420 constraint: forall b. D Int b => E (Int,c)
2422 You might think that the (D Int b) can't possibly contribute
2423 to solving (E (Int,c)), since the latter mentions 'c'. But
2424 in fact it can, because solving the (E (Int,c)) constraint needs
2427 and the (C Int) can be satisfied from the superclass of (D Int b).
2428 So we must still not float (E (Int,c)) out.
2430 To think about: special cases for unary type classes?
2432 Note [Pruning the givens in an implication constraint]
2433 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2434 Suppose we are about to form the implication constraint
2435 forall tvs. Eq a => Ord b
2436 The (Eq a) cannot contribute to the (Ord b), because it has no access to
2437 the type variable 'b'. So we could filter out the (Eq a) from the givens.
2438 But BE CAREFUL of the examples above in [Freeness and implications].
2440 Doing so would be a bit tidier, but all the implication constraints get
2441 simplified away by the optimiser, so it's no great win. So I don't take
2442 advantage of that at the moment.
2444 If you do, BE CAREFUL of wobbly type variables.
2447 %************************************************************************
2449 Avails and AvailHow: the pool of evidence
2451 %************************************************************************
2455 data Avails = Avails !ImprovementDone !AvailEnv
2457 type ImprovementDone = Bool -- True <=> some unification has happened
2458 -- so some Irreds might now be reducible
2459 -- keys that are now
2461 type AvailEnv = FiniteMap Inst AvailHow
2463 = IsIrred -- Used for irreducible dictionaries,
2464 -- which are going to be lambda bound
2466 | Given Inst -- Used for dictionaries for which we have a binding
2467 -- e.g. those "given" in a signature
2469 | Rhs -- Used when there is a RHS
2470 (LHsExpr TcId) -- The RHS
2471 [Inst] -- Insts free in the RHS; we need these too
2473 instance Outputable Avails where
2476 pprAvails :: Avails -> SDoc
2477 pprAvails (Avails imp avails)
2478 = vcat [ ptext (sLit "Avails") <> (if imp then ptext (sLit "[improved]") else empty)
2480 vcat [ sep [ppr inst, nest 2 (equals <+> ppr avail)]
2481 | (inst,avail) <- fmToList avails ]]
2483 instance Outputable AvailHow where
2486 -------------------------
2487 pprAvail :: AvailHow -> SDoc
2488 pprAvail IsIrred = text "Irred"
2489 pprAvail (Given x) = text "Given" <+> ppr x
2490 pprAvail (Rhs rhs bs) = sep [text "Rhs" <+> ppr bs,
2493 -------------------------
2494 extendAvailEnv :: AvailEnv -> Inst -> AvailHow -> AvailEnv
2495 extendAvailEnv env inst avail = addToFM env inst avail
2497 findAvailEnv :: AvailEnv -> Inst -> Maybe AvailHow
2498 findAvailEnv env wanted = lookupFM env wanted
2499 -- NB 1: the Ord instance of Inst compares by the class/type info
2500 -- *not* by unique. So
2501 -- d1::C Int == d2::C Int
2503 emptyAvails :: Avails
2504 emptyAvails = Avails False emptyFM
2506 findAvail :: Avails -> Inst -> Maybe AvailHow
2507 findAvail (Avails _ avails) wanted = findAvailEnv avails wanted
2509 elemAvails :: Inst -> Avails -> Bool
2510 elemAvails wanted (Avails _ avails) = wanted `elemFM` avails
2512 extendAvails :: Avails -> Inst -> AvailHow -> TcM Avails
2514 extendAvails avails@(Avails imp env) inst avail
2515 = do { imp1 <- tcImproveOne avails inst -- Do any improvement
2516 ; return (Avails (imp || imp1) (extendAvailEnv env inst avail)) }
2518 availsInsts :: Avails -> [Inst]
2519 availsInsts (Avails _ avails) = keysFM avails
2521 availsImproved :: Avails -> ImprovementDone
2522 availsImproved (Avails imp _) = imp
2525 Extracting the bindings from a bunch of Avails.
2526 The bindings do *not* come back sorted in dependency order.
2527 We assume that they'll be wrapped in a big Rec, so that the
2528 dependency analyser can sort them out later
2531 type DoneEnv = FiniteMap Inst [Id]
2532 -- Tracks which things we have evidence for
2534 extractResults :: Avails
2536 -> TcM (TcDictBinds, -- Bindings
2537 [Inst], -- The insts bound by the bindings
2538 [Inst]) -- Irreducible ones
2539 -- Note [Reducing implication constraints]
2541 extractResults (Avails _ avails) wanteds
2542 = go emptyBag [] [] emptyFM wanteds
2544 go :: TcDictBinds -- Bindings for dicts
2545 -> [Inst] -- Bound by the bindings
2547 -> DoneEnv -- Has an entry for each inst in the above three sets
2549 -> TcM (TcDictBinds, [Inst], [Inst])
2550 go binds bound_dicts irreds _ []
2551 = return (binds, bound_dicts, irreds)
2553 go binds bound_dicts irreds done (w:ws)
2555 = go binds bound_dicts (w:irreds) done' ws
2557 | Just done_ids@(done_id : rest_done_ids) <- lookupFM done w
2558 = if w_id `elem` done_ids then
2559 go binds bound_dicts irreds done ws
2561 go (add_bind (nlHsVar done_id)) bound_dicts irreds
2562 (addToFM done w (done_id : w_id : rest_done_ids)) ws
2564 | otherwise -- Not yet done
2565 = case findAvailEnv avails w of
2566 Nothing -> pprTrace "Urk: extractResults" (ppr w) $
2567 go binds bound_dicts irreds done ws
2569 Just IsIrred -> go binds bound_dicts (w:irreds) done' ws
2571 Just (Rhs rhs ws') -> go (add_bind rhs) (w:bound_dicts) irreds done' (ws' ++ ws)
2573 Just (Given g) -> go binds' bound_dicts irreds (addToFM done w [g_id]) ws
2576 binds' | w_id == g_id = binds
2577 | otherwise = add_bind (nlHsVar g_id)
2580 done' = addToFM done w [w_id]
2581 add_bind rhs = addInstToDictBind binds w rhs
2585 Note [No superclasses for Stop]
2586 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2587 When we decide not to reduce an Inst -- the 'WhatToDo' --- we still
2588 add it to avails, so that any other equal Insts will be commoned up
2589 right here. However, we do *not* add superclasses. If we have
2592 but a is not bound here, then we *don't* want to derive dn from df
2593 here lest we lose sharing.
2596 addWanted :: WantSCs -> Avails -> Inst -> LHsExpr TcId -> [Inst] -> TcM Avails
2597 addWanted want_scs avails wanted rhs_expr wanteds
2598 = addAvailAndSCs want_scs avails wanted avail
2600 avail = Rhs rhs_expr wanteds
2602 addGiven :: (Inst -> WantSCs) -> Avails -> Inst -> TcM Avails
2603 addGiven want_scs avails given = addAvailAndSCs (want_scs given) avails given (Given given)
2604 -- Conditionally add superclasses for 'givens'
2605 -- See Note [Recursive instances and superclases]
2607 -- No ASSERT( not (given `elemAvails` avails) ) because in an instance
2608 -- decl for Ord t we can add both Ord t and Eq t as 'givens',
2609 -- so the assert isn't true
2613 addIrred :: WantSCs -> Avails -> Inst -> TcM Avails
2614 addIrred want_scs avails irred = ASSERT2( not (irred `elemAvails` avails), ppr irred $$ ppr avails )
2615 addAvailAndSCs want_scs avails irred IsIrred
2617 addAvailAndSCs :: WantSCs -> Avails -> Inst -> AvailHow -> TcM Avails
2618 addAvailAndSCs want_scs avails inst avail
2619 | not (isClassDict inst) = extendAvails avails inst avail
2620 | NoSCs <- want_scs = extendAvails avails inst avail
2621 | otherwise = do { traceTc (text "addAvailAndSCs" <+> vcat [ppr inst, ppr deps])
2622 ; avails' <- extendAvails avails inst avail
2623 ; addSCs is_loop avails' inst }
2625 is_loop pred = any (`tcEqType` mkPredTy pred) dep_tys
2626 -- Note: this compares by *type*, not by Unique
2627 deps = findAllDeps (unitVarSet (instToVar inst)) avail
2628 dep_tys = map idType (varSetElems deps)
2630 findAllDeps :: IdSet -> AvailHow -> IdSet
2631 -- Find all the Insts that this one depends on
2632 -- See Note [SUPERCLASS-LOOP 2]
2633 -- Watch out, though. Since the avails may contain loops
2634 -- (see Note [RECURSIVE DICTIONARIES]), so we need to track the ones we've seen so far
2635 findAllDeps so_far (Rhs _ kids) = foldl find_all so_far kids
2636 findAllDeps so_far _ = so_far
2638 find_all :: IdSet -> Inst -> IdSet
2640 | isEqInst kid = so_far
2641 | kid_id `elemVarSet` so_far = so_far
2642 | Just avail <- findAvail avails kid = findAllDeps so_far' avail
2643 | otherwise = so_far'
2645 so_far' = extendVarSet so_far kid_id -- Add the new kid to so_far
2646 kid_id = instToId kid
2648 addSCs :: (TcPredType -> Bool) -> Avails -> Inst -> TcM Avails
2649 -- Add all the superclasses of the Inst to Avails
2650 -- The first param says "don't do this because the original thing
2651 -- depends on this one, so you'd build a loop"
2652 -- Invariant: the Inst is already in Avails.
2654 addSCs is_loop avails dict
2655 = ASSERT( isDict dict )
2656 do { sc_dicts <- newDictBndrs (instLoc dict) sc_theta'
2657 ; foldlM add_sc avails (zipEqual "add_scs" sc_dicts sc_sels) }
2659 (clas, tys) = getDictClassTys dict
2660 (tyvars, sc_theta, sc_sels, _) = classBigSig clas
2661 sc_theta' = filter (not . isEqPred) $
2662 substTheta (zipTopTvSubst tyvars tys) sc_theta
2664 add_sc avails (sc_dict, sc_sel)
2665 | is_loop (dictPred sc_dict) = return avails -- See Note [SUPERCLASS-LOOP 2]
2666 | is_given sc_dict = return avails
2667 | otherwise = do { avails' <- extendAvails avails sc_dict (Rhs sc_sel_rhs [dict])
2668 ; addSCs is_loop avails' sc_dict }
2670 sc_sel_rhs = L (instSpan dict) (HsWrap co_fn (HsVar sc_sel))
2671 co_fn = WpApp (instToVar dict) <.> mkWpTyApps tys
2673 is_given :: Inst -> Bool
2674 is_given sc_dict = case findAvail avails sc_dict of
2675 Just (Given _) -> True -- Given is cheaper than superclass selection
2678 -- From the a set of insts obtain all equalities that (transitively) occur in
2679 -- superclass contexts of class constraints (aka the ancestor equalities).
2681 ancestorEqualities :: [Inst] -> TcM [Inst]
2683 = mapM mkWantedEqInst -- turn only equality predicates..
2684 . filter isEqPred -- ..into wanted equality insts
2686 . addAEsToBag emptyBag -- collect the superclass constraints..
2687 . map dictPred -- ..of all predicates in a bag
2688 . filter isClassDict
2690 addAEsToBag :: Bag PredType -> [PredType] -> Bag PredType
2691 addAEsToBag bag [] = bag
2692 addAEsToBag bag (pred:preds)
2693 | pred `elemBag` bag = addAEsToBag bag preds
2694 | isEqPred pred = addAEsToBag bagWithPred preds
2695 | isClassPred pred = addAEsToBag bagWithPred predsWithSCs
2696 | otherwise = addAEsToBag bag preds
2698 bagWithPred = bag `snocBag` pred
2699 predsWithSCs = preds ++ substTheta (zipTopTvSubst tyvars tys) sc_theta
2701 (tyvars, sc_theta, _, _) = classBigSig clas
2702 (clas, tys) = getClassPredTys pred
2706 %************************************************************************
2708 \section{tcSimplifyTop: defaulting}
2710 %************************************************************************
2713 @tcSimplifyTop@ is called once per module to simplify all the constant
2714 and ambiguous Insts.
2716 We need to be careful of one case. Suppose we have
2718 instance Num a => Num (Foo a b) where ...
2720 and @tcSimplifyTop@ is given a constraint (Num (Foo x y)). Then it'll simplify
2721 to (Num x), and default x to Int. But what about y??
2723 It's OK: the final zonking stage should zap y to (), which is fine.
2727 tcSimplifyTop, tcSimplifyInteractive :: [Inst] -> TcM TcDictBinds
2728 tcSimplifyTop wanteds
2729 = tc_simplify_top doc False wanteds
2731 doc = text "tcSimplifyTop"
2733 tcSimplifyInteractive wanteds
2734 = tc_simplify_top doc True wanteds
2736 doc = text "tcSimplifyInteractive"
2738 -- The TcLclEnv should be valid here, solely to improve
2739 -- error message generation for the monomorphism restriction
2740 tc_simplify_top :: SDoc -> Bool -> [Inst] -> TcM (Bag (LHsBind TcId))
2741 tc_simplify_top doc interactive wanteds
2742 = do { dflags <- getDOpts
2743 ; wanteds <- zonkInsts wanteds
2744 ; mapM_ zonkTopTyVar (varSetElems (tyVarsOfInsts wanteds))
2746 ; traceTc (text "tc_simplify_top 0: " <+> ppr wanteds)
2747 ; (irreds1, binds1) <- tryHardCheckLoop doc1 wanteds
2748 -- ; (irreds1, binds1) <- gentleInferLoop doc1 wanteds
2749 ; traceTc (text "tc_simplify_top 1: " <+> ppr irreds1)
2750 ; (irreds2, binds2) <- approximateImplications doc2 (\_ -> True) irreds1
2751 ; traceTc (text "tc_simplify_top 2: " <+> ppr irreds2)
2753 -- Use the defaulting rules to do extra unification
2754 -- NB: irreds2 are already zonked
2755 ; (irreds3, binds3) <- disambiguate doc3 interactive dflags irreds2
2757 -- Deal with implicit parameters
2758 ; let (bad_ips, non_ips) = partition isIPDict irreds3
2759 (ambigs, others) = partition isTyVarDict non_ips
2761 ; topIPErrs bad_ips -- Can arise from f :: Int -> Int
2763 ; addNoInstanceErrs others
2764 ; addTopAmbigErrs ambigs
2766 ; return (binds1 `unionBags` binds2 `unionBags` binds3) }
2768 doc1 = doc <+> ptext (sLit "(first round)")
2769 doc2 = doc <+> ptext (sLit "(approximate)")
2770 doc3 = doc <+> ptext (sLit "(disambiguate)")
2773 If a dictionary constrains a type variable which is
2774 * not mentioned in the environment
2775 * and not mentioned in the type of the expression
2776 then it is ambiguous. No further information will arise to instantiate
2777 the type variable; nor will it be generalised and turned into an extra
2778 parameter to a function.
2780 It is an error for this to occur, except that Haskell provided for
2781 certain rules to be applied in the special case of numeric types.
2783 * at least one of its classes is a numeric class, and
2784 * all of its classes are numeric or standard
2785 then the type variable can be defaulted to the first type in the
2786 default-type list which is an instance of all the offending classes.
2788 So here is the function which does the work. It takes the ambiguous
2789 dictionaries and either resolves them (producing bindings) or
2790 complains. It works by splitting the dictionary list by type
2791 variable, and using @disambigOne@ to do the real business.
2793 @disambigOne@ assumes that its arguments dictionaries constrain all
2794 the same type variable.
2796 ADR Comment 20/6/94: I've changed the @CReturnable@ case to default to
2797 @()@ instead of @Int@. I reckon this is the Right Thing to do since
2798 the most common use of defaulting is code like:
2800 _ccall_ foo `seqPrimIO` bar
2802 Since we're not using the result of @foo@, the result if (presumably)
2806 disambiguate :: SDoc -> Bool -> DynFlags -> [Inst] -> TcM ([Inst], TcDictBinds)
2807 -- Just does unification to fix the default types
2808 -- The Insts are assumed to be pre-zonked
2809 disambiguate doc interactive dflags insts
2811 = return (insts, emptyBag)
2813 | null defaultable_groups
2814 = do { traceTc (text "disambigutate, no defaultable groups" <+> vcat [ppr unaries, ppr insts, ppr bad_tvs, ppr defaultable_groups])
2815 ; return (insts, emptyBag) }
2818 = do { -- Figure out what default types to use
2819 default_tys <- getDefaultTys extended_defaulting ovl_strings
2821 ; traceTc (text "disambiguate1" <+> vcat [ppr insts, ppr unaries, ppr bad_tvs, ppr defaultable_groups])
2822 ; mapM_ (disambigGroup default_tys) defaultable_groups
2824 -- disambigGroup does unification, hence try again
2825 ; tryHardCheckLoop doc insts }
2828 extended_defaulting = interactive || dopt Opt_ExtendedDefaultRules dflags
2829 ovl_strings = dopt Opt_OverloadedStrings dflags
2831 unaries :: [(Inst, Class, TcTyVar)] -- (C tv) constraints
2832 bad_tvs :: TcTyVarSet -- Tyvars mentioned by *other* constraints
2833 (unaries, bad_tvs_s) = partitionWith find_unary insts
2834 bad_tvs = unionVarSets bad_tvs_s
2836 -- Finds unary type-class constraints
2837 find_unary d@(Dict {tci_pred = ClassP cls [ty]})
2838 | Just tv <- tcGetTyVar_maybe ty = Left (d,cls,tv)
2839 find_unary inst = Right (tyVarsOfInst inst)
2841 -- Group by type variable
2842 defaultable_groups :: [[(Inst,Class,TcTyVar)]]
2843 defaultable_groups = filter defaultable_group (equivClasses cmp_tv unaries)
2844 cmp_tv (_,_,tv1) (_,_,tv2) = tv1 `compare` tv2
2846 defaultable_group :: [(Inst,Class,TcTyVar)] -> Bool
2847 defaultable_group ds@((_,_,tv):_)
2848 = isTyConableTyVar tv -- Note [Avoiding spurious errors]
2849 && not (tv `elemVarSet` bad_tvs)
2850 && defaultable_classes [c | (_,c,_) <- ds]
2851 defaultable_group [] = panic "defaultable_group"
2853 defaultable_classes clss
2854 | extended_defaulting = any isInteractiveClass clss
2855 | otherwise = all is_std_class clss && (any is_num_class clss)
2857 -- In interactive mode, or with -XExtendedDefaultRules,
2858 -- we default Show a to Show () to avoid graututious errors on "show []"
2859 isInteractiveClass cls
2860 = is_num_class cls || (classKey cls `elem` [showClassKey, eqClassKey, ordClassKey])
2862 is_num_class cls = isNumericClass cls || (ovl_strings && (cls `hasKey` isStringClassKey))
2863 -- is_num_class adds IsString to the standard numeric classes,
2864 -- when -foverloaded-strings is enabled
2866 is_std_class cls = isStandardClass cls || (ovl_strings && (cls `hasKey` isStringClassKey))
2867 -- Similarly is_std_class
2869 -----------------------
2870 disambigGroup :: [Type] -- The default types
2871 -> [(Inst,Class,TcTyVar)] -- All standard classes of form (C a)
2872 -> TcM () -- Just does unification, to fix the default types
2874 disambigGroup default_tys dicts
2875 = try_default default_tys
2877 (_,_,tyvar) = ASSERT(not (null dicts)) head dicts -- Should be non-empty
2878 classes = [c | (_,c,_) <- dicts]
2880 try_default [] = return ()
2881 try_default (default_ty : default_tys)
2882 = tryTcLIE_ (try_default default_tys) $
2883 do { tcSimplifyDefault [mkClassPred clas [default_ty] | clas <- classes]
2884 -- This may fail; then the tryTcLIE_ kicks in
2885 -- Failure here is caused by there being no type in the
2886 -- default list which can satisfy all the ambiguous classes.
2887 -- For example, if Real a is reqd, but the only type in the
2888 -- default list is Int.
2890 -- After this we can't fail
2891 ; warnDefault dicts default_ty
2892 ; unifyType default_ty (mkTyVarTy tyvar)
2893 ; return () -- TOMDO: do something with the coercion
2897 -----------------------
2898 getDefaultTys :: Bool -> Bool -> TcM [Type]
2899 getDefaultTys extended_deflts ovl_strings
2900 = do { mb_defaults <- getDeclaredDefaultTys
2901 ; case mb_defaults of {
2902 Just tys -> return tys ; -- User-supplied defaults
2905 -- No use-supplied default
2906 -- Use [Integer, Double], plus modifications
2907 { integer_ty <- tcMetaTy integerTyConName
2908 ; checkWiredInTyCon doubleTyCon
2909 ; string_ty <- tcMetaTy stringTyConName
2910 ; return (opt_deflt extended_deflts unitTy
2911 -- Note [Default unitTy]
2913 [integer_ty,doubleTy]
2915 opt_deflt ovl_strings string_ty) } } }
2917 opt_deflt True ty = [ty]
2918 opt_deflt False _ = []
2921 Note [Default unitTy]
2922 ~~~~~~~~~~~~~~~~~~~~~
2923 In interative mode (or with -XExtendedDefaultRules) we add () as the first type we
2924 try when defaulting. This has very little real impact, except in the following case.
2926 Text.Printf.printf "hello"
2927 This has type (forall a. IO a); it prints "hello", and returns 'undefined'. We don't
2928 want the GHCi repl loop to try to print that 'undefined'. The neatest thing is to
2929 default the 'a' to (), rather than to Integer (which is what would otherwise happen;
2930 and then GHCi doesn't attempt to print the (). So in interactive mode, we add
2931 () to the list of defaulting types. See Trac #1200.
2933 Note [Avoiding spurious errors]
2934 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2935 When doing the unification for defaulting, we check for skolem
2936 type variables, and simply don't default them. For example:
2937 f = (*) -- Monomorphic
2938 g :: Num a => a -> a
2940 Here, we get a complaint when checking the type signature for g,
2941 that g isn't polymorphic enough; but then we get another one when
2942 dealing with the (Num a) context arising from f's definition;
2943 we try to unify a with Int (to default it), but find that it's
2944 already been unified with the rigid variable from g's type sig
2947 %************************************************************************
2949 \subsection[simple]{@Simple@ versions}
2951 %************************************************************************
2953 Much simpler versions when there are no bindings to make!
2955 @tcSimplifyThetas@ simplifies class-type constraints formed by
2956 @deriving@ declarations and when specialising instances. We are
2957 only interested in the simplified bunch of class/type constraints.
2959 It simplifies to constraints of the form (C a b c) where
2960 a,b,c are type variables. This is required for the context of
2961 instance declarations.
2964 tcSimplifyDeriv :: InstOrigin
2966 -> ThetaType -- Wanted
2967 -> TcM ThetaType -- Needed
2968 -- Given instance (wanted) => C inst_ty
2969 -- Simplify 'wanted' as much as possible
2971 tcSimplifyDeriv orig tyvars theta
2972 = do { (tvs, _, tenv) <- tcInstTyVars tyvars
2973 -- The main loop may do unification, and that may crash if
2974 -- it doesn't see a TcTyVar, so we have to instantiate. Sigh
2975 -- ToDo: what if two of them do get unified?
2976 ; wanteds <- newDictBndrsO orig (substTheta tenv theta)
2977 ; (irreds, _) <- tryHardCheckLoop doc wanteds
2979 ; let (tv_dicts, others) = partition ok irreds
2980 ; addNoInstanceErrs others
2981 -- See Note [Exotic derived instance contexts] in TcMType
2983 ; let rev_env = zipTopTvSubst tvs (mkTyVarTys tyvars)
2984 simpl_theta = substTheta rev_env (map dictPred tv_dicts)
2985 -- This reverse-mapping is a pain, but the result
2986 -- should mention the original TyVars not TcTyVars
2988 ; return simpl_theta }
2990 doc = ptext (sLit "deriving classes for a data type")
2992 ok dict | isDict dict = validDerivPred (dictPred dict)
2997 @tcSimplifyDefault@ just checks class-type constraints, essentially;
2998 used with \tr{default} declarations. We are only interested in
2999 whether it worked or not.
3002 tcSimplifyDefault :: ThetaType -- Wanted; has no type variables in it
3005 tcSimplifyDefault theta = do
3006 wanteds <- newDictBndrsO DefaultOrigin theta
3007 (irreds, _) <- tryHardCheckLoop doc wanteds
3008 addNoInstanceErrs irreds
3012 traceTc (ptext (sLit "tcSimplifyDefault failing")) >> failM
3014 doc = ptext (sLit "default declaration")
3018 %************************************************************************
3020 \section{Errors and contexts}
3022 %************************************************************************
3024 ToDo: for these error messages, should we note the location as coming
3025 from the insts, or just whatever seems to be around in the monad just
3029 groupErrs :: ([Inst] -> TcM ()) -- Deal with one group
3030 -> [Inst] -- The offending Insts
3032 -- Group together insts with the same origin
3033 -- We want to report them together in error messages
3037 groupErrs report_err (inst:insts)
3038 = do { do_one (inst:friends)
3039 ; groupErrs report_err others }
3041 -- (It may seem a bit crude to compare the error messages,
3042 -- but it makes sure that we combine just what the user sees,
3043 -- and it avoids need equality on InstLocs.)
3044 (friends, others) = partition is_friend insts
3045 loc_msg = showSDoc (pprInstLoc (instLoc inst))
3046 is_friend friend = showSDoc (pprInstLoc (instLoc friend)) == loc_msg
3047 do_one insts = addInstCtxt (instLoc (head insts)) (report_err insts)
3048 -- Add location and context information derived from the Insts
3050 -- Add the "arising from..." part to a message about bunch of dicts
3051 addInstLoc :: [Inst] -> Message -> Message
3052 addInstLoc insts msg = msg $$ nest 2 (pprInstArising (head insts))
3054 addTopIPErrs :: [Name] -> [Inst] -> TcM ()
3057 addTopIPErrs bndrs ips
3058 = do { dflags <- getDOpts
3059 ; addErrTcM (tidy_env, mk_msg dflags tidy_ips) }
3061 (tidy_env, tidy_ips) = tidyInsts ips
3063 = vcat [sep [ptext (sLit "Implicit parameters escape from"),
3064 nest 2 (ptext (sLit "the monomorphic top-level binding")
3065 <> plural bndrs <+> ptext (sLit "of")
3066 <+> pprBinders bndrs <> colon)],
3067 nest 2 (vcat (map ppr_ip ips)),
3068 monomorphism_fix dflags]
3069 ppr_ip ip = pprPred (dictPred ip) <+> pprInstArising ip
3071 topIPErrs :: [Inst] -> TcM ()
3073 = groupErrs report tidy_dicts
3075 (tidy_env, tidy_dicts) = tidyInsts dicts
3076 report dicts = addErrTcM (tidy_env, mk_msg dicts)
3077 mk_msg dicts = addInstLoc dicts (ptext (sLit "Unbound implicit parameter") <>
3078 plural tidy_dicts <+> pprDictsTheta tidy_dicts)
3080 addNoInstanceErrs :: [Inst] -- Wanted (can include implications)
3082 addNoInstanceErrs insts
3083 = do { let (tidy_env, tidy_insts) = tidyInsts insts
3084 ; reportNoInstances tidy_env Nothing tidy_insts }
3088 -> Maybe (InstLoc, [Inst]) -- Context
3089 -- Nothing => top level
3090 -- Just (d,g) => d describes the construct
3092 -> [Inst] -- What is wanted (can include implications)
3095 reportNoInstances tidy_env mb_what insts
3096 = groupErrs (report_no_instances tidy_env mb_what) insts
3098 report_no_instances :: TidyEnv -> Maybe (InstLoc, [Inst]) -> [Inst] -> TcM ()
3099 report_no_instances tidy_env mb_what insts
3100 = do { inst_envs <- tcGetInstEnvs
3101 ; let (implics, insts1) = partition isImplicInst insts
3102 (insts2, overlaps) = partitionWith (check_overlap inst_envs) insts1
3103 (eqInsts, insts3) = partition isEqInst insts2
3104 ; traceTc (text "reportNoInstances" <+> vcat
3105 [ppr insts, ppr implics, ppr insts1, ppr insts2])
3106 ; mapM_ complain_implic implics
3107 ; mapM_ (\doc -> addErrTcM (tidy_env, doc)) overlaps
3108 ; groupErrs complain_no_inst insts3
3109 ; mapM_ (addErrTcM . mk_eq_err) eqInsts
3112 complain_no_inst insts = addErrTcM (tidy_env, mk_no_inst_err insts)
3114 complain_implic inst -- Recurse!
3115 = reportNoInstances tidy_env
3116 (Just (tci_loc inst, tci_given inst))
3119 check_overlap :: (InstEnv,InstEnv) -> Inst -> Either Inst SDoc
3120 -- Right msg => overlap message
3121 -- Left inst => no instance
3122 check_overlap inst_envs wanted
3123 | not (isClassDict wanted) = Left wanted
3125 = case lookupInstEnv inst_envs clas tys of
3126 ([], _) -> Left wanted -- No match
3127 -- The case of exactly one match and no unifiers means a
3128 -- successful lookup. That can't happen here, because dicts
3129 -- only end up here if they didn't match in Inst.lookupInst
3131 | debugIsOn -> pprPanic "reportNoInstance" (ppr wanted)
3132 res -> Right (mk_overlap_msg wanted res)
3134 (clas,tys) = getDictClassTys wanted
3136 mk_overlap_msg dict (matches, unifiers)
3137 = ASSERT( not (null matches) )
3138 vcat [ addInstLoc [dict] ((ptext (sLit "Overlapping instances for")
3139 <+> pprPred (dictPred dict))),
3140 sep [ptext (sLit "Matching instances") <> colon,
3141 nest 2 (vcat [pprInstances ispecs, pprInstances unifiers])],
3142 if not (isSingleton matches)
3143 then -- Two or more matches
3145 else -- One match, plus some unifiers
3146 ASSERT( not (null unifiers) )
3147 parens (vcat [ptext (sLit "The choice depends on the instantiation of") <+>
3148 quotes (pprWithCommas ppr (varSetElems (tyVarsOfInst dict))),
3149 ptext (sLit "To pick the first instance above, use -XIncoherentInstances"),
3150 ptext (sLit "when compiling the other instance declarations")])]
3152 ispecs = [ispec | (ispec, _) <- matches]
3154 mk_eq_err :: Inst -> (TidyEnv, SDoc)
3155 mk_eq_err inst = misMatchMsg tidy_env (eqInstTys inst)
3157 mk_no_inst_err insts
3158 | null insts = empty
3160 | Just (loc, givens) <- mb_what, -- Nested (type signatures, instance decls)
3161 not (isEmptyVarSet (tyVarsOfInsts insts))
3162 = vcat [ addInstLoc insts $
3163 sep [ ptext (sLit "Could not deduce") <+> pprDictsTheta insts
3164 , nest 2 $ ptext (sLit "from the context") <+> pprDictsTheta givens]
3165 , show_fixes (fix1 loc : fixes2) ]
3167 | otherwise -- Top level
3168 = vcat [ addInstLoc insts $
3169 ptext (sLit "No instance") <> plural insts
3170 <+> ptext (sLit "for") <+> pprDictsTheta insts
3171 , show_fixes fixes2 ]
3174 fix1 loc = sep [ ptext (sLit "add") <+> pprDictsTheta insts
3175 <+> ptext (sLit "to the context of"),
3176 nest 2 (ppr (instLocOrigin loc)) ]
3177 -- I'm not sure it helps to add the location
3178 -- nest 2 (ptext (sLit "at") <+> ppr (instLocSpan loc)) ]
3180 fixes2 | null instance_dicts = []
3181 | otherwise = [sep [ptext (sLit "add an instance declaration for"),
3182 pprDictsTheta instance_dicts]]
3183 instance_dicts = [d | d <- insts, isClassDict d, not (isTyVarDict d)]
3184 -- Insts for which it is worth suggesting an adding an instance declaration
3185 -- Exclude implicit parameters, and tyvar dicts
3187 show_fixes :: [SDoc] -> SDoc
3188 show_fixes [] = empty
3189 show_fixes (f:fs) = sep [ptext (sLit "Possible fix:"),
3190 nest 2 (vcat (f : map (ptext (sLit "or") <+>) fs))]
3192 addTopAmbigErrs :: [Inst] -> TcRn ()
3193 addTopAmbigErrs dicts
3194 -- Divide into groups that share a common set of ambiguous tyvars
3195 = ifErrsM (return ()) $ -- Only report ambiguity if no other errors happened
3196 -- See Note [Avoiding spurious errors]
3197 mapM_ report (equivClasses cmp [(d, tvs_of d) | d <- tidy_dicts])
3199 (tidy_env, tidy_dicts) = tidyInsts dicts
3201 tvs_of :: Inst -> [TcTyVar]
3202 tvs_of d = varSetElems (tyVarsOfInst d)
3203 cmp (_,tvs1) (_,tvs2) = tvs1 `compare` tvs2
3205 report :: [(Inst,[TcTyVar])] -> TcM ()
3206 report pairs@((inst,tvs) : _) = do -- The pairs share a common set of ambiguous tyvars
3207 (tidy_env, mono_msg) <- mkMonomorphismMsg tidy_env tvs
3208 setSrcSpan (instSpan inst) $
3209 -- the location of the first one will do for the err message
3210 addErrTcM (tidy_env, msg $$ mono_msg)
3212 dicts = map fst pairs
3213 msg = sep [text "Ambiguous type variable" <> plural tvs <+>
3214 pprQuotedList tvs <+> in_msg,
3215 nest 2 (pprDictsInFull dicts)]
3216 in_msg = text "in the constraint" <> plural dicts <> colon
3217 report [] = panic "addTopAmbigErrs"
3220 mkMonomorphismMsg :: TidyEnv -> [TcTyVar] -> TcM (TidyEnv, Message)
3221 -- There's an error with these Insts; if they have free type variables
3222 -- it's probably caused by the monomorphism restriction.
3223 -- Try to identify the offending variable
3224 -- ASSUMPTION: the Insts are fully zonked
3225 mkMonomorphismMsg tidy_env inst_tvs
3226 = do { dflags <- getDOpts
3227 ; (tidy_env, docs) <- findGlobals (mkVarSet inst_tvs) tidy_env
3228 ; return (tidy_env, mk_msg dflags docs) }
3230 mk_msg _ _ | any isRuntimeUnk inst_tvs
3231 = vcat [ptext (sLit "Cannot resolve unknown runtime types:") <+>
3232 (pprWithCommas ppr inst_tvs),
3233 ptext (sLit "Use :print or :force to determine these types")]
3234 mk_msg _ [] = ptext (sLit "Probable fix: add a type signature that fixes these type variable(s)")
3235 -- This happens in things like
3236 -- f x = show (read "foo")
3237 -- where monomorphism doesn't play any role
3239 = vcat [ptext (sLit "Possible cause: the monomorphism restriction applied to the following:"),
3241 monomorphism_fix dflags]
3243 monomorphism_fix :: DynFlags -> SDoc
3244 monomorphism_fix dflags
3245 = ptext (sLit "Probable fix:") <+> vcat
3246 [ptext (sLit "give these definition(s) an explicit type signature"),
3247 if dopt Opt_MonomorphismRestriction dflags
3248 then ptext (sLit "or use -XNoMonomorphismRestriction")
3249 else empty] -- Only suggest adding "-XNoMonomorphismRestriction"
3250 -- if it is not already set!
3252 warnDefault :: [(Inst, Class, Var)] -> Type -> TcM ()
3253 warnDefault ups default_ty = do
3254 warn_flag <- doptM Opt_WarnTypeDefaults
3255 addInstCtxt (instLoc (head (dicts))) (warnTc warn_flag warn_msg)
3257 dicts = [d | (d,_,_) <- ups]
3260 (_, tidy_dicts) = tidyInsts dicts
3261 warn_msg = vcat [ptext (sLit "Defaulting the following constraint(s) to type") <+>
3262 quotes (ppr default_ty),
3263 pprDictsInFull tidy_dicts]
3265 reduceDepthErr :: Int -> [Inst] -> SDoc
3266 reduceDepthErr n stack
3267 = vcat [ptext (sLit "Context reduction stack overflow; size =") <+> int n,
3268 ptext (sLit "Use -fcontext-stack=N to increase stack size to N"),
3269 nest 4 (pprStack stack)]
3271 pprStack :: [Inst] -> SDoc
3272 pprStack stack = vcat (map pprInstInFull stack)