TcSimplify.lhs 52.1 KB
Newer Older
1
\begin{code}
Ian Lynagh's avatar
Ian Lynagh committed
2 3 4 5 6 7 8
{-# OPTIONS -fno-warn-tabs #-}
-- The above warning supression flag is a temporary kludge.
-- While working on this module you are encouraged to remove it and
-- detab the module (please do the detabbing in a separate patch). See
--     http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#TabsvsSpaces
-- for details

9
module TcSimplify( 
10
       simplifyInfer, simplifyAmbiguityCheck,
11
       simplifyDefault, simplifyDeriv, 
12 13
       simplifyRule, simplifyTop, simplifyInteractive,
       solveWantedsTcM
14
  ) where
15

16
#include "HsVersions.h"
17

18
import TcRnTypes
19
import TcRnMonad
20
import TcErrors
21
import TcMType
22 23
import TcType 
import TcSMonad 
24
import TcInteract 
25
import Inst
26 27
import Type     ( classifyPredType, PredTree(..), getClassPredTys_maybe )
import Class    ( Class )
28
import Var
29
import Unique
30
import VarSet
31
import VarEnv 
32
import TcEvidence
33
import TypeRep
34
import Name
35
import Bag
36 37
import ListSetOps
import Util
38 39 40
import PrelInfo
import PrelNames
import Class		( classKey )
41
import BasicTypes       ( RuleName )
42
import Outputable
43
import FastString
dimitris's avatar
dimitris committed
44
import TrieMap () -- DV: for now
45 46 47
\end{code}


48 49 50 51 52
*********************************************************************************
*                                                                               * 
*                           External interface                                  *
*                                                                               *
*********************************************************************************
53

54 55 56
\begin{code}
simplifyTop :: WantedConstraints -> TcM (Bag EvBind)
-- Simplify top-level constraints
57 58 59
-- Usually these will be implications,
-- but when there is nothing to quantify we don't wrap
-- in a degenerate implication, so we do that here instead
60 61 62 63 64
simplifyTop wanteds
  = do { traceTc "simplifyTop {" $ text "wanted = " <+> ppr wanteds 
       ; ev_binds_var <- newTcEvBinds
       ; zonked_final_wc <- solveWantedsTcMWithEvBinds ev_binds_var wanteds simpl_top
       ; binds1 <- TcRnMonad.getTcEvBinds ev_binds_var
65 66 67
       ; traceTc "End simplifyTop }" empty

       ; traceTc "reportUnsolved {" empty
68
       ; binds2 <- reportUnsolved zonked_final_wc
69
       ; traceTc "reportUnsolved }" empty
70
         
71
       ; return (binds1 `unionBags` binds2) }
72

73 74 75
  where
    -- See Note [Top-level Defaulting Plan]
    simpl_top wanteds
76
      = do { wc_first_go <- nestTcS (solve_wanteds_and_drop wanteds)
77 78
           ; applyTyVarDefaulting wc_first_go 
           ; simpl_top_loop wc_first_go }
79
    
80 81 82 83
    simpl_top_loop wc
      | isEmptyWC wc 
      = return wc
      | otherwise
84
      = do { wc_residual <- nestTcS (solve_wanteds_and_drop wc)
85 86 87 88 89 90 91
           ; let wc_flat_approximate = approximateWC wc_residual
           ; something_happened <- applyDefaultingRules wc_flat_approximate
                                        -- See Note [Top-level Defaulting Plan]
           ; if something_happened then 
               simpl_top_loop wc_residual 
             else 
               return wc_residual }
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
\end{code}

Note [Top-level Defaulting Plan]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

We have considered two design choices for where/when to apply defaulting.   
   (i) Do it in SimplCheck mode only /whenever/ you try to solve some 
       flat constraints, maybe deep inside the context of implications.
       This used to be the case in GHC 7.4.1.
   (ii) Do it in a tight loop at simplifyTop, once all other constraint has 
        finished. This is the current story.

Option (i) had many disadvantages: 
   a) First it was deep inside the actual solver, 
   b) Second it was dependent on the context (Infer a type signature, 
      or Check a type signature, or Interactive) since we did not want 
      to always start defaulting when inferring (though there is an exception to  
      this see Note [Default while Inferring])
   c) It plainly did not work. Consider typecheck/should_compile/DfltProb2.hs:
          f :: Int -> Bool
          f x = const True (\y -> let w :: a -> a
                                      w a = const a (y+1)
                                  in w y)
      We will get an implication constraint (for beta the type of y):
               [untch=beta] forall a. 0 => Num beta
      which we really cannot default /while solving/ the implication, since beta is
      untouchable.

Instead our new defaulting story is to pull defaulting out of the solver loop and
go with option (i), implemented at SimplifyTop. Namely:
     - First have a go at solving the residual constraint of the whole program
     - Try to approximate it with a flat constraint
     - Figure out derived defaulting equations for that flat constraint
     - Go round the loop again if you did manage to get some equations

Now, that has to do with class defaulting. However there exists type variable /kind/
defaulting. Again this is done at the top-level and the plan is:
     - At the top-level, once you had a go at solving the constraint, do 
       figure out /all/ the touchable unification variables of the wanted contraints.
     - Apply defaulting to their kinds

More details in Note [DefaultTyVar].

\begin{code}
136

137 138 139
------------------
simplifyAmbiguityCheck :: Name -> WantedConstraints -> TcM (Bag EvBind)
simplifyAmbiguityCheck name wanteds
140
  = traceTc "simplifyAmbiguityCheck" (text "name =" <+> ppr name) >> 
141
    simplifyTop wanteds  -- NB: must be simplifyTop so that we
142 143
                         --     do ambiguity resolution.  
                         -- See Note [Impedence matching] in TcBinds.
144
 
145 146 147
------------------
simplifyInteractive :: WantedConstraints -> TcM (Bag EvBind)
simplifyInteractive wanteds 
148 149
  = traceTc "simplifyInteractive" empty >>
    simplifyTop wanteds 
150 151 152 153 154

------------------
simplifyDefault :: ThetaType	-- Wanted; has no type variables in it
                -> TcM ()	-- Succeeds iff the constraint is soluble
simplifyDefault theta
155 156
  = do { traceTc "simplifyInteractive" empty
       ; wanted <- newFlatWanteds DefaultOrigin theta
157 158 159 160
       ; (unsolved, _binds) <- solveWantedsTcM (mkFlatWC wanted)

       ; traceTc "reportUnsolved {" empty
       -- See Note [Deferring coercion errors to runtime]
161
       ; reportAllUnsolved unsolved 
162 163 164
         -- Postcondition of solveWantedsTcM is that returned
         -- constraints are zonked. So Precondition of reportUnsolved
         -- is true.
165 166
       ; traceTc "reportUnsolved }" empty

167 168
       ; return () }
\end{code}
169

170

171
***********************************************************************************
172
*                                                                                 * 
173
*                            Deriving                                             *
174 175
*                                                                                 *
***********************************************************************************
176

177 178
\begin{code}
simplifyDeriv :: CtOrigin
179 180 181 182
              -> PredType
	      -> [TyVar]	
	      -> ThetaType		-- Wanted
	      -> TcM ThetaType	-- Needed
183 184
-- Given  instance (wanted) => C inst_ty 
-- Simplify 'wanted' as much as possibles
185
-- Fail if not possible
186
simplifyDeriv orig pred tvs theta 
187
  = do { (skol_subst, tvs_skols) <- tcInstSkolTyVars tvs -- Skolemize
simonpj@microsoft.com's avatar
simonpj@microsoft.com committed
188 189 190 191
      	 	-- The constraint solving machinery 
		-- expects *TcTyVars* not TyVars.  
		-- We use *non-overlappable* (vanilla) skolems
		-- See Note [Overlap and deriving]
192

193
       ; let subst_skol = zipTopTvSubst tvs_skols $ map mkTyVarTy tvs
194
             skol_set   = mkVarSet tvs_skols
195
	     doc = ptext (sLit "deriving") <+> parens (ppr pred)
196 197 198

       ; wanted <- newFlatWanteds orig (substTheta skol_subst theta)

199 200
       ; traceTc "simplifyDeriv" $ 
         vcat [ pprTvBndrs tvs $$ ppr theta $$ ppr wanted, doc ]
201
       ; (residual_wanted, _ev_binds1)
202
             <- solveWantedsTcM (mkFlatWC wanted)
203
                -- Post: residual_wanted are already zonked
204

205 206
       ; let (good, bad) = partitionBagWith get_good (wc_flat residual_wanted)
                         -- See Note [Exotic derived instance contexts]
207
             get_good :: Ct -> Either PredType Ct
208 209 210 211 212 213
             get_good ct | validDerivPred skol_set p 
                         , isWantedCt ct  = Left p 
                         -- NB: residual_wanted may contain unsolved
                         -- Derived and we stick them into the bad set
                         -- so that reportUnsolved may decide what to do with them
                         | otherwise = Right ct
214
                         where p = ctPred ct
215

216 217
       -- We never want to defer these errors because they are errors in the
       -- compiler! Hence the `False` below
218
       ; reportAllUnsolved (residual_wanted { wc_flat = bad })
219

220 221
       ; let min_theta = mkMinimalBySCs (bagToList good)
       ; return (substTheta subst_skol min_theta) }
222
\end{code}
223

simonpj@microsoft.com's avatar
simonpj@microsoft.com committed
224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
Note [Overlap and deriving]
~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider some overlapping instances:
  data Show a => Show [a] where ..
  data Show [Char] where ...

Now a data type with deriving:
  data T a = MkT [a] deriving( Show )

We want to get the derived instance
  instance Show [a] => Show (T a) where...
and NOT
  instance Show a => Show (T a) where...
so that the (Show (T Char)) instance does the Right Thing

It's very like the situation when we're inferring the type
of a function
   f x = show [x]
and we want to infer
   f :: Show [a] => a -> String

BOTTOM LINE: use vanilla, non-overlappable skolems when inferring
             the context for the derived instance. 
	     Hence tcInstSkolTyVars not tcInstSuperSkolTyVars

249 250 251 252 253 254 255
Note [Exotic derived instance contexts]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In a 'derived' instance declaration, we *infer* the context.  It's a
bit unclear what rules we should apply for this; the Haskell report is
silent.  Obviously, constraints like (Eq a) are fine, but what about
	data T f a = MkT (f a) deriving( Eq )
where we'd get an Eq (f a) constraint.  That's probably fine too.
256

257 258 259
One could go further: consider
	data T a b c = MkT (Foo a b c) deriving( Eq )
	instance (C Int a, Eq b, Eq c) => Eq (Foo a b c)
260

261 262
Notice that this instance (just) satisfies the Paterson termination 
conditions.  Then we *could* derive an instance decl like this:
263

264 265 266 267
	instance (C Int a, Eq b, Eq c) => Eq (T a b c) 
even though there is no instance for (C Int a), because there just
*might* be an instance for, say, (C Int Bool) at a site where we
need the equality instance for T's.  
268

269 270 271
However, this seems pretty exotic, and it's quite tricky to allow
this, and yet give sensible error messages in the (much more common)
case where we really want that instance decl for C.
272

273 274
So for now we simply require that the derived instance context
should have only type-variable constraints.
275

276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304
Here is another example:
	data Fix f = In (f (Fix f)) deriving( Eq )
Here, if we are prepared to allow -XUndecidableInstances we
could derive the instance
	instance Eq (f (Fix f)) => Eq (Fix f)
but this is so delicate that I don't think it should happen inside
'deriving'. If you want this, write it yourself!

NB: if you want to lift this condition, make sure you still meet the
termination conditions!  If not, the deriving mechanism generates
larger and larger constraints.  Example:
  data Succ a = S a
  data Seq a = Cons a (Seq (Succ a)) | Nil deriving Show

Note the lack of a Show instance for Succ.  First we'll generate
  instance (Show (Succ a), Show a) => Show (Seq a)
and then
  instance (Show (Succ (Succ a)), Show (Succ a), Show a) => Show (Seq a)
and so on.  Instead we want to complain of no instance for (Show (Succ a)).

The bottom line
~~~~~~~~~~~~~~~
Allow constraints which consist only of type variables, with no repeats.

*********************************************************************************
*                                                                                 * 
*                            Inference
*                                                                                 *
***********************************************************************************
305

dreixel's avatar
dreixel committed
306 307 308 309 310 311 312 313 314 315 316 317
Note [Which variables to quantify]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Suppose the inferred type of a function is
   T kappa (alpha:kappa) -> Int
where alpha is a type unification variable and 
      kappa is a kind unification variable
Then we want to quantify over *both* alpha and kappa.  But notice that
kappa appears "at top level" of the type, as well as inside the kind
of alpha.  So it should be fine to just look for the "top level"
kind/type variables of the type, without looking transitively into the
kinds of those type variables.

318
\begin{code}
319
simplifyInfer :: Bool
320 321 322
              -> Bool                  -- Apply monomorphism restriction
              -> [(Name, TcTauType)]   -- Variables to be generalised,
                                       -- and their tau-types
323
              -> WantedConstraints
324 325
              -> TcM ([TcTyVar],    -- Quantify over these type variables
                      [EvVar],      -- ... and these constraints
326 327 328
		      Bool,	    -- The monomorphism restriction did something
		      		    --   so the results type is not as general as
				    --   it could be
329
                      TcEvBinds)    -- ... binding these evidence variables
330
simplifyInfer _top_lvl apply_mr name_taus wanteds
331 332 333
  | isEmptyWC wanteds
  = do { gbl_tvs     <- tcGetGlobalTyVars            -- Already zonked
       ; zonked_taus <- zonkTcTypes (map snd name_taus)
Simon Peyton Jones's avatar
Simon Peyton Jones committed
334
       ; let tvs_to_quantify = varSetElems (tyVarsOfTypes zonked_taus `minusVarSet` gbl_tvs)
dreixel's avatar
dreixel committed
335 336 337
       	     		       -- tvs_to_quantify can contain both kind and type vars
       	                       -- See Note [Which variables to quantify]
       ; qtvs <- zonkQuantifiedTyVars tvs_to_quantify
338
       ; return (qtvs, [], False, emptyTcEvBinds) }
339

340
  | otherwise
341
  = do { zonked_tau_tvs <- zonkTyVarsAndFV (tyVarsOfTypes (map snd name_taus))
342

343
       ; ev_binds_var <- newTcEvBinds
344
       ; traceTc "simplifyInfer {"  $ vcat
345
             [ ptext (sLit "names =") <+> ppr (map fst name_taus)
346 347
             , ptext (sLit "taus =") <+> ppr (map snd name_taus)
             , ptext (sLit "tau_tvs (zonked) =") <+> ppr zonked_tau_tvs
348 349
             , ptext (sLit "closed =") <+> ppr _top_lvl
             , ptext (sLit "apply_mr =") <+> ppr apply_mr
350
             , ptext (sLit "(unzonked) wanted =") <+> ppr wanteds
351 352
             ]

353 354 355 356 357
              -- Historical note: Before step 2 we used to have a
              -- HORRIBLE HACK described in Note [Avoid unecessary
              -- constraint simplification] but, as described in Trac
              -- #4361, we have taken in out now.  That's why we start
              -- with step 2!
358

359 360 361 362 363 364 365 366
              -- Step 2) First try full-blown solving 

              -- NB: we must gather up all the bindings from doing
              -- this solving; hence (runTcSWithEvBinds ev_binds_var).
              -- And note that since there are nested implications,
              -- calling solveWanteds will side-effect their evidence
              -- bindings, so we can't just revert to the input
              -- constraint.
367 368 369
       ; wanted_transformed <- solveWantedsTcMWithEvBinds ev_binds_var wanteds $
                               solve_wanteds_and_drop
                               -- Post: wanted_transformed are zonked
370 371

              -- Step 4) Candidates for quantification are an approximation of wanted_transformed
372 373 374
              -- NB: Already the fixpoint of any unifications that may have happened                                
              -- NB: We do not do any defaulting when inferring a type, this can lead
              -- to less polymorphic types, see Note [Default while Inferring]
375
 
376 377
              -- Step 5) Minimize the quantification candidates                             
              -- Step 6) Final candidates for quantification                
378
              -- We discard bindings, insolubles etc, because all we are
379 380
              -- care aout it

381
       ; (quant_pred_candidates, _extra_binds)   
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396
             <- if insolubleWC wanted_transformed 
                then return ([], emptyBag)   -- See Note [Quantification with errors]
                else runTcS $ 
                do { let quant_candidates = approximateWC wanted_transformed
                   ; traceTcS "simplifyWithApprox" $
                     text "quant_candidates = " <+> ppr quant_candidates
                   ; promoteTyVars quant_candidates
                   ; _implics <- solveInteract quant_candidates
                   ; (flats, _insols) <- getInertUnsolved
                   -- NB: Dimitrios is slightly worried that we will get
                   -- family equalities (F Int ~ alpha) in the quantification
                   -- candidates, as we have performed no further unflattening
                   -- at this point. Nothing bad, but inferred contexts might
                   -- look complicated.
                   ; return (map ctPred $ filter isWantedCt (bagToList flats)) }
397 398 399

             -- NB: quant_pred_candidates is already the fixpoint of any 
             --     unifications that may have happened
400 401
                  
       ; gbl_tvs        <- tcGetGlobalTyVars -- TODO: can we just use untch instead of gbl_tvs?
402
       ; zonked_tau_tvs <- zonkTyVarsAndFV zonked_tau_tvs
403
       
404
       ; let init_tvs  = zonked_tau_tvs `minusVarSet` gbl_tvs
405
             poly_qtvs = growThetaTyVars quant_pred_candidates init_tvs 
406
                         `minusVarSet` gbl_tvs
407
             pbound    = filter (quantifyPred poly_qtvs) quant_pred_candidates
408
             
409
	     -- Monomorphism restriction
410 411
             mr_qtvs  	     = init_tvs `minusVarSet` constrained_tvs
             constrained_tvs = tyVarsOfTypes quant_pred_candidates
412
	     mr_bites        = apply_mr && not (null pbound)
413

414 415
             (qtvs, bound) | mr_bites  = (mr_qtvs,   [])
                           | otherwise = (poly_qtvs, pbound)
416
             
417 418 419 420 421 422 423
       ; traceTc "simplifyWithApprox" $
         vcat [ ptext (sLit "quant_pred_candidates =") <+> ppr quant_pred_candidates
              , ptext (sLit "gbl_tvs=") <+> ppr gbl_tvs
              , ptext (sLit "zonked_tau_tvs=") <+> ppr zonked_tau_tvs
              , ptext (sLit "pbound =") <+> ppr pbound
              , ptext (sLit "init_qtvs =") <+> ppr init_tvs 
              , ptext (sLit "poly_qtvs =") <+> ppr poly_qtvs ]
424

425
       ; if isEmptyVarSet qtvs && null bound
426 427 428 429
         then do { traceTc "} simplifyInfer/no quantification" empty                   
                 ; emitConstraints wanted_transformed
                    -- Includes insolubles (if -fdefer-type-errors)
                    -- as well as flats and implications
430
                 ; return ([], [], mr_bites, TcEvBinds ev_binds_var) }
431 432
         else do

433 434 435
       { traceTc "simplifyApprox" $ 
         ptext (sLit "bound are =") <+> ppr bound 
         
436
            -- Step 4, zonk quantified variables 
437
       ; let minimal_flat_preds = mkMinimalBySCs bound
438 439
             skol_info = InferSkol [ (name, mkSigmaTy [] minimal_flat_preds ty)
                                   | (name, ty) <- name_taus ]
440 441 442 443
                        -- Don't add the quantified variables here, because
                        -- they are also bound in ic_skols and we want them to be
                        -- tidied uniformly

Simon Peyton Jones's avatar
Simon Peyton Jones committed
444
       ; qtvs_to_return <- zonkQuantifiedTyVars (varSetElems qtvs)
445

446
            -- Step 7) Emit an implication
447
       ; minimal_bound_ev_vars <- mapM TcMType.newEvVar minimal_flat_preds
448 449
       ; lcl_env <- TcRnMonad.getLclEnv
       ; let implic = Implic { ic_untch    = pushUntouchables (tcl_untch lcl_env)
450
                             , ic_skols    = qtvs_to_return
451 452
                             , ic_fsks     = []  -- wanted_tansformed arose only from solveWanteds
                                                 -- hence no flatten-skolems (which come from givens)
453
                             , ic_given    = minimal_bound_ev_vars
454
                             , ic_wanted   = wanted_transformed 
455 456
                             , ic_insol    = False
                             , ic_binds    = ev_binds_var
457 458
                             , ic_info     = skol_info
                             , ic_env      = lcl_env }
459
       ; emitImplication implic
460
         
461 462 463
       ; traceTc "} simplifyInfer/produced residual implication for quantification" $
             vcat [ ptext (sLit "implic =") <+> ppr implic
                       -- ic_skols, ic_given give rest of result
464
                  , ptext (sLit "qtvs =") <+> ppr qtvs_to_return
465
                  , ptext (sLit "spb =") <+> ppr quant_pred_candidates
466 467
                  , ptext (sLit "bound =") <+> ppr bound ]

468 469
       ; return ( qtvs_to_return, minimal_bound_ev_vars
                , mr_bites,  TcEvBinds ev_binds_var) } }
470
\end{code}
471

472 473 474 475 476 477 478 479 480
Note [Quantification with errors]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
If we find that the RHS of the definition has some absolutely-insoluble
constraints, we abandon all attempts to find a context to quantify
over, and instead make the function fully-polymorphic in whatever
type we have found.  For two reasons
  a) Minimise downstream errors
  b) Avoid spurious errors from this function
   
481

482 483
Note [Default while Inferring]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517
Our current plan is that defaulting only happens at simplifyTop and
not simplifyInfer.  This may lead to some insoluble deferred constraints
Example:

instance D g => C g Int b 

constraint inferred = (forall b. 0 => C gamma alpha b) /\ Num alpha
type inferred       = gamma -> gamma 

Now, if we try to default (alpha := Int) we will be able to refine the implication to 
  (forall b. 0 => C gamma Int b) 
which can then be simplified further to 
  (forall b. 0 => D gamma)
Finally we /can/ approximate this implication with (D gamma) and infer the quantified
type:  forall g. D g => g -> g

Instead what will currently happen is that we will get a quantified type 
(forall g. g -> g) and an implication:
       forall g. 0 => (forall b. 0 => C g alpha b) /\ Num alpha

which, even if the simplifyTop defaults (alpha := Int) we will still be left with an 
unsolvable implication:
       forall g. 0 => (forall b. 0 => D g)

The concrete example would be: 
       h :: C g a s => g -> a -> ST s a
       f (x::gamma) = (\_ -> x) (runST (h x (undefined::alpha)) + 1)

But it is quite tedious to do defaulting and resolve the implication constraints and
we have not observed code breaking because of the lack of defaulting in inference so 
we don't do it for now.



518 519 520 521 522 523 524 525 526
Note [Minimize by Superclasses]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
When we quantify over a constraint, in simplifyInfer we need to
quantify over a constraint that is minimal in some sense: For
instance, if the final wanted constraint is (Eq alpha, Ord alpha),
we'd like to quantify over Ord alpha, because we can just get Eq alpha
from superclass selection from Ord alpha. This minimization is what
mkMinimalBySCs does. Then, simplifyInfer uses the minimal constraint
to check the original wanted.
527

528

529 530
Note [Avoid unecessary constraint simplification]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
531 532 533 534
    -------- NB NB NB (Jun 12) ------------- 
    This note not longer applies; see the notes with Trac #4361.
    But I'm leaving it in here so we remember the issue.)
    ----------------------------------------
535
When inferring the type of a let-binding, with simplifyInfer,
536
try to avoid unnecessarily simplifying class constraints.
537 538
Doing so aids sharing, but it also helps with delicate 
situations like
539

540
   instance C t => C [t] where ..
541

542 543 544 545 546 547 548 549 550 551 552
   f :: C [t] => ....
   f x = let g y = ...(constraint C [t])... 
         in ...
When inferring a type for 'g', we don't want to apply the
instance decl, because then we can't satisfy (C t).  So we
just notice that g isn't quantified over 't' and partition
the contraints before simplifying.

This only half-works, but then let-generalisation only half-works.


553 554 555 556 557
*********************************************************************************
*                                                                                 * 
*                             RULES                                               *
*                                                                                 *
***********************************************************************************
558

559
See note [Simplifying RULE consraints] in TcRule
560

561 562 563 564 565 566 567 568 569 570 571 572 573 574
Note [RULE quanfification over equalities]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Decideing which equalities to quantify over is tricky:
 * We do not want to quantify over insoluble equalities (Int ~ Bool)
    (a) because we prefer to report a LHS type error
    (b) because if such things end up in 'givens' we get a bogus
        "inaccessible code" error

 * But we do want to quantify over things like (a ~ F b), where
   F is a type function.

The difficulty is that it's hard to tell what is insoluble!
So we see whether the simplificaiotn step yielded any type errors,
and if so refrain from quantifying over *any* equalites.
575 576

\begin{code}
577 578 579
simplifyRule :: RuleName 
             -> WantedConstraints	-- Constraints from LHS
             -> WantedConstraints	-- Constraints from RHS
580 581 582
             -> TcM ([EvVar], WantedConstraints)   -- LHS evidence varaibles
-- See Note [Simplifying RULE constraints] in TcRule
simplifyRule name lhs_wanted rhs_wanted
583
  = do {      	 -- We allow ourselves to unify environment 
584
		 -- variables: runTcS runs with NoUntouchables
585
         (resid_wanted, _) <- solveWantedsTcM (lhs_wanted `andWC` rhs_wanted)
586
                              -- Post: these are zonked and unflattened
587

588 589 590 591
       -- Dimitrios would be happy if we could avoid this zonking here. But
       -- I am afraid that if we do not zonk, we will quantify over the wrong things.
       ; _ev_binds_var <- newTcEvBinds 
       ; zonked_lhs <- zonkWC _ev_binds_var lhs_wanted -- Don't care about binds
592

593 594 595 596 597 598 599 600 601 602 603 604 605
       ; let (q_cts, non_q_cts) = partitionBag quantify_me (wc_flat zonked_lhs)
             quantify_me  -- Note [RULE quantification over equalities]
               | insolubleWC resid_wanted = quantify_insol
               | otherwise                = quantify_normal

             quantify_insol ct = not (isEqPred (ctPred ct))

             quantify_normal ct
               | EqPred t1 t2 <- classifyPredType (ctPred ct)
               = not (t1 `eqType` t2)
               | otherwise
               = True
             
606
       ; traceTc "simplifyRule" $
607
         vcat [ ptext (sLit "LHS of rule") <+> doubleQuotes (ftext name)
608
              , text "zonked_lhs" <+> ppr zonked_lhs 
609 610
              , text "q_cts"      <+> ppr q_cts ]

611 612
       ; return ( map (ctEvId . ctEvidence) (bagToList q_cts)
                , zonked_lhs { wc_flat = non_q_cts }) }
613 614 615
\end{code}


616 617 618 619 620
*********************************************************************************
*                                                                                 * 
*                                 Main Simplifier                                 *
*                                                                                 *
***********************************************************************************
621

622 623 624 625 626 627
Note [Deferring coercion errors to runtime]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
While developing, sometimes it is desirable to allow compilation to succeed even
if there are type errors in the code. Consider the following case:

  module Main where
628

629 630
  a :: Int
  a = 'a'
631

632
  main = print "b"
633

634 635
Even though `a` is ill-typed, it is not used in the end, so if all that we're
interested in is `main` it is handy to be able to ignore the problems in `a`.
636

637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659
Since we treat type equalities as evidence, this is relatively simple. Whenever
we run into a type mismatch in TcUnify, we normally just emit an error. But it
is always safe to defer the mismatch to the main constraint solver. If we do
that, `a` will get transformed into

  co :: Int ~ Char
  co = ...

  a :: Int
  a = 'a' `cast` co

The constraint solver would realize that `co` is an insoluble constraint, and
emit an error with `reportUnsolved`. But we can also replace the right-hand side
of `co` with `error "Deferred type error: Int ~ Char"`. This allows the program
to compile, and it will run fine unless we evaluate `a`. This is what
`deferErrorsToRuntime` does.

It does this by keeping track of which errors correspond to which coercion
in TcErrors (with ErrEnv). TcErrors.reportTidyWanteds does not print the errors
and does not fail if -fwarn-type-errors is on, so that we can continue
compilation. The errors are turned into warnings in `reportUnsolved`.

\begin{code}
660 661 662 663
solveWantedsTcMWithEvBinds :: EvBindsVar
                           -> WantedConstraints
                           -> (WantedConstraints -> TcS WantedConstraints)
                           -> TcM WantedConstraints
664 665 666 667 668 669
-- Returns a *zonked* result
-- We zonk when we finish primarily to un-flatten out any
-- flatten-skolems etc introduced by canonicalisation of
-- types involving type funuctions.  Happily the result 
-- is typically much smaller than the input, indeed it is 
-- often empty.
670
solveWantedsTcMWithEvBinds ev_binds_var wc tcs_action
671 672
  = do { traceTc "solveWantedsTcMWithEvBinds" $ text "wanted=" <+> ppr wc
       ; wc2 <- runTcSWithEvBinds ev_binds_var (tcs_action wc)
673 674
       ; zonkWC ev_binds_var wc2 }

675
solveWantedsTcM :: WantedConstraints -> TcM (WantedConstraints, Bag EvBind)
676
-- Zonk the input constraints, and simplify them
677
-- Return the evidence binds in the BagEvBinds result
678
-- Discards all Derived stuff in result
679
-- Postcondition: fully zonked and unflattened constraints
680
solveWantedsTcM wanted 
681 682 683
  = do { ev_binds_var <- newTcEvBinds
       ; wanteds' <- solveWantedsTcMWithEvBinds ev_binds_var wanted solve_wanteds_and_drop
       ; binds <- TcRnMonad.getTcEvBinds ev_binds_var
684
       ; return (wanteds', binds) }
685

686 687 688 689 690
solve_wanteds_and_drop :: WantedConstraints -> TcS (WantedConstraints)
-- Since solve_wanteds returns the residual WantedConstraints,
-- it should alway be called within a runTcS or something similar,
solve_wanteds_and_drop wanted = do { wc <- solve_wanteds wanted 
                                   ; return (dropDerivedWC wc) }
691 692

solve_wanteds :: WantedConstraints -> TcS WantedConstraints 
693
-- so that the inert set doesn't mindlessly propagate.
694
-- NB: wc_flats may be wanted /or/ derived now
695
solve_wanteds wanted@(WC { wc_flat = flats, wc_impl = implics, wc_insol = insols }) 
696 697
  = do { traceTcS "solveWanteds {" (ppr wanted)

698 699
         -- Try the flat bit, including insolubles. Solving insolubles a 
         -- second time round is a bit of a waste but the code is simple 
700 701 702
         -- and the program is wrong anyway, and we don't run the danger 
         -- of adding Derived insolubles twice; see 
         -- TcSMonad Note [Do not add duplicate derived insolubles] 
703
       ; traceTcS "solveFlats {" empty
704
       ; let all_flats = flats `unionBags` insols
705 706
       ; impls_from_flats <- solveInteract all_flats
       ; traceTcS "solveFlats end }" (ppr impls_from_flats)
707

708 709
       -- solve_wanteds iterates when it is able to float equalities 
       -- out of one or more of the implications. 
710
       ; unsolved_implics <- simpl_loop 1 (implics `unionBags` impls_from_flats)
711

712 713
       ; (unsolved_flats, insoluble_flats) <- getInertUnsolved

714 715 716 717 718 719
        -- We used to unflatten here but now we only do it once at top-level
        -- during zonking -- see Note [Unflattening while zonking] in TcMType
       ; let wc = WC { wc_flat  = unsolved_flats   
                     , wc_impl  = unsolved_implics 
                     , wc_insol = insoluble_flats }
                  
720
       ; bb <- getTcEvBindsMap
721
       ; tb <- getTcSTyBindsMap
722
       ; traceTcS "solveWanteds }" $
723
                 vcat [ text "unsolved_flats   =" <+> ppr unsolved_flats
724
                      , text "unsolved_implics =" <+> ppr unsolved_implics
725
                      , text "current evbinds  =" <+> ppr (evBindMapBinds bb)
726
                      , text "current tybinds  =" <+> vcat (map ppr (varEnvElts tb))
727
                      , text "final wc =" <+> ppr wc ]
728

729
       ; return wc }
730 731 732 733 734 735 736 737

simpl_loop :: Int
           -> Bag Implication
           -> TcS (Bag Implication)
simpl_loop n implics
  | n > 10 
  = traceTcS "solveWanteds: loop!" empty >> return implics
  | otherwise 
738 739 740 741 742 743 744
  = do { (floated_eqs, unsolved_implics) <- solveNestedImplications implics
       ; if isEmptyBag floated_eqs 
         then return unsolved_implics 
         else 
    do {   -- Put floated_eqs into the current inert set before looping
         impls_from_eqs <- solveInteract floated_eqs
       ; simpl_loop (n+1) (unsolved_implics `unionBags` impls_from_eqs)} }
745

746

747 748 749 750 751 752 753 754 755
solveNestedImplications :: Bag Implication
                        -> TcS (Cts, Bag Implication)
-- Precondition: the TcS inerts may contain unsolved flats which have 
-- to be converted to givens before we go inside a nested implication.
solveNestedImplications implics
  | isEmptyBag implics
  = return (emptyBag, emptyBag)
  | otherwise 
  = do { inerts <- getTcSInerts
756 757
       ; let thinner_inerts = prepareInertsForImplications inerts
                 -- See Note [Preparing inert set for implications]
758
  
759
       ; traceTcS "solveNestedImplications starting {" $ 
760
         vcat [ text "original inerts = " <+> ppr inerts
761 762
              , text "thinner_inerts  = " <+> ppr thinner_inerts ]
         
763
       ; (floated_eqs, unsolved_implics)
764
           <- flatMapBagPairM (solveImplication thinner_inerts) implics
765 766 767 768

       -- ... and we are back in the original TcS inerts 
       -- Notice that the original includes the _insoluble_flats so it was safe to ignore
       -- them in the beginning of this function.
769
       ; traceTcS "solveNestedImplications end }" $
770
                  vcat [ text "all floated_eqs ="  <+> ppr floated_eqs
771 772
                       , text "unsolved_implics =" <+> ppr unsolved_implics ]

773
       ; return (floated_eqs, unsolved_implics) }
774

775
solveImplication :: InertSet
776 777 778 779 780
                 -> Implication    -- Wanted
                 -> TcS (Cts,      -- All wanted or derived floated equalities: var = type
                         Bag Implication) -- Unsolved rest (always empty or singleton)
-- Precondition: The TcS monad contains an empty worklist and given-only inerts 
-- which after trying to solve this implication we must restore to their original value
781
solveImplication inerts
782
     imp@(Implic { ic_untch  = untch
783 784
                 , ic_binds  = ev_binds
                 , ic_skols  = skols 
785
                 , ic_fsks   = old_fsks
786
                 , ic_given  = givens
787
                 , ic_wanted = wanteds
788 789
                 , ic_info   = info
                 , ic_env    = env })
790
  = 
791 792
    do { traceTcS "solveImplication {" (ppr imp) 

793
         -- Solve the nested constraints
794 795 796
         -- NB: 'inerts' has empty inert_fsks
       ; (new_fsks, residual_wanted) 
            <- nestImplicTcS ev_binds untch inerts $
797
               do { solveInteractGiven (mkGivenLoc info env) old_fsks givens 
798 799 800 801 802 803 804 805
                  ; residual_wanted <- solve_wanteds wanteds
                  ; more_fsks <- getFlattenSkols
                  ; return (more_fsks ++ old_fsks, residual_wanted) }

       ; (floated_eqs, final_wanted)
             <- floatEqualities (skols ++ new_fsks) givens residual_wanted

       ; let res_implic | isEmptyWC final_wanted 
806 807
                        = emptyBag
                        | otherwise
808 809 810
                        = unitBag (imp { ic_fsks   = new_fsks
                                       , ic_wanted = dropDerivedWC final_wanted
                                       , ic_insol  = insolubleWC final_wanted })
811

812
       ; evbinds <- getTcEvBindsMap
813
       ; traceTcS "solveImplication end }" $ vcat
814
             [ text "floated_eqs =" <+> ppr floated_eqs
815
             , text "new_fsks =" <+> ppr new_fsks
816 817
             , text "res_implic =" <+> ppr res_implic
             , text "implication evbinds = " <+> ppr (evBindMapBinds evbinds) ]
818

819
       ; return (floated_eqs, res_implic) }
820 821 822 823
\end{code}


\begin{code}
824 825
floatEqualities :: [TcTyVar] -> [EvVar] -> WantedConstraints 
                -> TcS (Cts, WantedConstraints)
826 827
-- Post: The returned FlavoredEvVar's are only Wanted or Derived
-- and come from the input wanted ev vars or deriveds 
828 829
-- Also performs some unifications, adding to monadically-carried ty_binds
-- These will be used when processing floated_eqs later
830 831
floatEqualities skols can_given wanteds@(WC { wc_flat = flats })
  | hasEqualities can_given 
832
  = return (emptyBag, wanteds)   -- Note [Float Equalities out of Implications]
833
  | otherwise 
834 835
  = do { let (float_eqs, remaining_flats) = partitionBag is_floatable flats
       ; promoteTyVars float_eqs
836
       ; ty_binds <- getTcSTyBindsMap
837
       ; traceTcS "floatEqualities" (vcat [ text "Floated eqs =" <+> ppr float_eqs
838 839
                                          , text "Ty binds =" <+> ppr ty_binds])
       ; return (float_eqs, wanteds { wc_flat = remaining_flats }) }
840 841 842 843 844 845 846 847 848
  where 
    skol_set = growSkols wanteds (mkVarSet skols)

    is_floatable :: Ct -> Bool
    is_floatable ct
       = isEqPred pred && skol_set `disjointVarSet` tyVarsOfType pred
       where
         pred = ctPred ct

849
promoteTyVars :: Cts -> TcS ()
850 851 852
-- When we float a constraint out of an implication we
-- must restore (MetaTvInv) in Note [Untouchable type variables]
-- in TcType
853 854 855 856
promoteTyVars cts
  = do { untch <- TcSMonad.getUntouchables
       ; mapM_ (promote_tv untch) (varSetElems (tyVarsOfCts cts)) }
  where
857 858 859 860 861 862 863 864
    promote_tv untch tv 
      | isFloatedTouchableMetaTyVar untch tv
      = do { cloned_tv <- TcSMonad.cloneMetaTyVar tv
           ; let rhs_tv = setMetaTyVarUntouchables cloned_tv untch
           ; setWantedTyBind tv (mkTyVarTy rhs_tv) }
      | otherwise
      = return ()

865 866 867 868 869 870 871 872 873
growSkols :: WantedConstraints -> VarSet -> VarSet
-- Find all the type variables that might possibly be unified
-- with a type that mentions a skolem.  This test is very conservative.
-- I don't *think* we need look inside the implications, because any 
-- relevant unification variables in there are untouchable.
growSkols (WC { wc_flat = flats }) skols
  = growThetaTyVars theta skols
  where
    theta = foldrBag ((:) . ctPred) [] flats
874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898

approximateWC :: WantedConstraints -> Cts
-- Postcondition: Wanted or Derived Cts 
approximateWC wc = float_wc emptyVarSet wc
  where 
    float_wc :: TcTyVarSet -> WantedConstraints -> Cts
    float_wc skols (WC { wc_flat = flat, wc_impl = implic }) = floats1 `unionBags` floats2
      where floats1 = do_bag (float_flat skols) flat
            floats2 = do_bag (float_implic skols) implic
                                 
    float_implic :: TcTyVarSet -> Implication -> Cts
    float_implic skols imp
      = float_wc skols' (ic_wanted imp)
      where
        skols' = skols `extendVarSetList` ic_skols imp `extendVarSetList` ic_fsks imp
            
    float_flat :: TcTyVarSet -> Ct -> Cts
    float_flat skols ct
      | tyVarsOfCt ct `disjointVarSet` skols 
      = singleCt ct
      | otherwise = emptyCts
        
    do_bag :: (a -> Bag c) -> Bag a -> Bag c
    do_bag f = foldrBag (unionBags.f) emptyBag
\end{code}
899

simonpj@microsoft.com's avatar
simonpj@microsoft.com committed
900 901
Note [Float Equalities out of Implications]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944
For ordinary pattern matches (including existentials) we float 
equalities out of implications, for instance: 
     data T where 
       MkT :: Eq a => a -> T 
     f x y = case x of MkT _ -> (y::Int)
We get the implication constraint (x::T) (y::alpha): 
     forall a. [untouchable=alpha] Eq a => alpha ~ Int
We want to float out the equality into a scope where alpha is no
longer untouchable, to solve the implication!  

But we cannot float equalities out of implications whose givens may
yield or contain equalities:

      data T a where 
        T1 :: T Int
        T2 :: T Bool
        T3 :: T a 
        
      h :: T a -> a -> Int
      
      f x y = case x of 
                T1 -> y::Int
                T2 -> y::Bool
                T3 -> h x y

We generate constraint, for (x::T alpha) and (y :: beta): 
   [untouchables = beta] (alpha ~ Int => beta ~ Int)   -- From 1st branch
   [untouchables = beta] (alpha ~ Bool => beta ~ Bool) -- From 2nd branch
   (alpha ~ beta)                                      -- From 3rd branch 

If we float the equality (beta ~ Int) outside of the first implication and 
the equality (beta ~ Bool) out of the second we get an insoluble constraint.
But if we just leave them inside the implications we unify alpha := beta and
solve everything.

Principle: 
    We do not want to float equalities out which may need the given *evidence*
    to become soluble.

Consequence: classes with functional dependencies don't matter (since there is 
no evidence for a fundep equality), but equality superclasses do matter (since 
they carry evidence).

945 946 947 948 949 950 951 952 953
Note [Promoting unification variables]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When we float an equality out of an implication we must "promote" free
unification variables of the equality, in order to maintain Invariant
(MetaTvInv) from Note [Untouchable type variables] in TcType.  for the
leftover implication.

This is absolutely necessary. Consider the following example. We start
with two implications and a class with a functional dependency.
954

955 956 957 958 959
    class C x y | x -> y
    instance C [a] [a]
          
    (I1)      [untch=beta]forall b. 0 => F Int ~ [beta]
    (I2)      [untch=beta]forall c. 0 => F Int ~ [[alpha]] /\ C beta [c]
960 961 962 963 964 965 966

We float (F Int ~ [beta]) out of I1, and we float (F Int ~ [[alpha]]) out of I2. 
They may react to yield that (beta := [alpha]) which can then be pushed inwards 
the leftover of I2 to get (C [alpha] [a]) which, using the FunDep, will mean that
(alpha := a). In the end we will have the skolem 'b' escaping in the untouchable
beta! Concrete example is in indexed_types/should_fail/ExtraTcsUntch.hs:

967 968 969 970 971 972 973 974 975 976 977 978 979
    class C x y | x -> y where 
     op :: x -> y -> ()

    instance C [a] [a]

    type family F a :: *

    h :: F Int -> ()
    h = undefined

    data TEx where 
      TEx :: a -> TEx 

980

981 982 983 984 985 986
    f (x::beta) = 
        let g1 :: forall b. b -> ()
            g1 _ = h [x]
            g2 z = case z of TEx y -> (h [[undefined]], op x [y])
        in (g1 '3', g2 undefined)

987

988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006

Note [Solving Family Equations] 
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 
After we are done with simplification we may be left with constraints of the form:
     [Wanted] F xis ~ beta 
If 'beta' is a touchable unification variable not already bound in the TyBinds 
then we'd like to create a binding for it, effectively "defaulting" it to be 'F xis'.

When is it ok to do so? 
    1) 'beta' must not already be defaulted to something. Example: 

           [Wanted] F Int  ~ beta   <~ Will default [beta := F Int]
           [Wanted] F Char ~ beta   <~ Already defaulted, can't default again. We 
                                       have to report this as unsolved.

    2) However, we must still do an occurs check when defaulting (F xis ~ beta), to 
       set [beta := F xis] only if beta is not among the free variables of xis.

    3) Notice that 'beta' can't be bound in ty binds already because we rewrite RHS 
1007 1008 1009 1010
       of type family equations. See Inert Set invariants in TcInteract.

This solving is now happening during zonking, see Note [Unflattening during zonking]
in TcMType.
1011 1012


1013 1014 1015 1016 1017
*********************************************************************************
*                                                                               * 
*                          Defaulting and disamgiguation                        *
*                                                                               *
*********************************************************************************
1018
\begin{code}
1019 1020 1021
applyDefaultingRules :: Cts -> TcS Bool
  -- True <=> I did some defaulting, reflected in ty_binds
                 
1022 1023
-- Return some extra derived equalities, which express the
-- type-class default choice. 
1024
applyDefaultingRules wanteds
1025
  | isEmptyBag wanteds 
1026
  = return False
1027
  | otherwise
1028 1029
  = do { traceTcS "applyDefaultingRules { " $ 
                  text "wanteds =" <+> ppr wanteds
1030 1031 1032
                  
       ; info@(default_tys, _) <- getDefaultInfo
       ; let groups = findDefaultableGroups info wanteds
1033 1034
       ; traceTcS "findDefaultableGroups" $ vcat [ text "groups=" <+> ppr groups
                                                 , text "info=" <+> ppr info ]
1035
       ; something_happeneds <- mapM (disambigGroup default_tys) groups
1036

1037
       ; traceTcS "applyDefaultingRules }" (ppr something_happeneds)
1038

1039
       ; return (or something_happeneds) }
1040