2 % (c) The University of Glasgow 2006
3 % (c) The AQUA Project, Glasgow University, 1998
6 This module contains definitions for the IdInfo for things that
7 have a standard form, namely:
11 * method and superclass selectors
12 * primitive operations
16 mkDictFunId, mkDefaultMethodId,
21 mkPrimOpId, mkFCallId, mkTickBoxOpId, mkBreakPointOpId,
23 mkReboxingAlt, wrapNewTypeBody, unwrapNewTypeBody,
24 mkUnpackCase, mkProductBox,
26 -- And some particular Ids; see below for why they are wired in
27 wiredInIds, ghcPrimIds,
28 unsafeCoerceId, realWorldPrimId, voidArgId, nullAddrId, seqId,
29 lazyId, lazyIdUnfolding, lazyIdKey,
32 rEC_CON_ERROR_ID, iRREFUT_PAT_ERROR_ID, rUNTIME_ERROR_ID,
33 nON_EXHAUSTIVE_GUARDS_ERROR_ID, nO_METHOD_BINDING_ERROR_ID,
34 pAT_ERROR_ID, eRROR_ID,
39 #include "HsVersions.h"
61 import Var ( Var, TyVar)
69 import BasicTypes hiding ( SuccessFlag(..) )
77 %************************************************************************
79 \subsection{Wired in Ids}
81 %************************************************************************
85 = [ -- These error-y things are wired in because we don't yet have
86 -- a way to express in an interface file that the result type variable
87 -- is 'open'; that is can be unified with an unboxed type
89 -- [The interface file format now carry such information, but there's
90 -- no way yet of expressing at the definition site for these
91 -- error-reporting functions that they have an 'open'
92 -- result type. -- sof 1/99]
94 eRROR_ID, -- This one isn't used anywhere else in the compiler
95 -- But we still need it in wiredInIds so that when GHC
96 -- compiles a program that mentions 'error' we don't
97 -- import its type from the interface file; we just get
98 -- the Id defined here. Which has an 'open-tyvar' type.
101 iRREFUT_PAT_ERROR_ID,
102 nON_EXHAUSTIVE_GUARDS_ERROR_ID,
103 nO_METHOD_BINDING_ERROR_ID,
110 -- These Ids are exported from GHC.Prim
112 = [ -- These can't be defined in Haskell, but they have
113 -- perfectly reasonable unfoldings in Core
121 %************************************************************************
123 \subsection{Data constructors}
125 %************************************************************************
127 The wrapper for a constructor is an ordinary top-level binding that evaluates
128 any strict args, unboxes any args that are going to be flattened, and calls
131 We're going to build a constructor that looks like:
133 data (Data a, C b) => T a b = T1 !a !Int b
136 \d1::Data a, d2::C b ->
137 \p q r -> case p of { p ->
139 Con T1 [a,b] [p,q,r]}}
143 * d2 is thrown away --- a context in a data decl is used to make sure
144 one *could* construct dictionaries at the site the constructor
145 is used, but the dictionary isn't actually used.
147 * We have to check that we can construct Data dictionaries for
148 the types a and Int. Once we've done that we can throw d1 away too.
150 * We use (case p of q -> ...) to evaluate p, rather than "seq" because
151 all that matters is that the arguments are evaluated. "seq" is
152 very careful to preserve evaluation order, which we don't need
155 You might think that we could simply give constructors some strictness
156 info, like PrimOps, and let CoreToStg do the let-to-case transformation.
157 But we don't do that because in the case of primops and functions strictness
158 is a *property* not a *requirement*. In the case of constructors we need to
159 do something active to evaluate the argument.
161 Making an explicit case expression allows the simplifier to eliminate
162 it in the (common) case where the constructor arg is already evaluated.
164 Note [Wrappers for data instance tycons]
165 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
166 In the case of data instances, the wrapper also applies the coercion turning
167 the representation type into the family instance type to cast the result of
168 the wrapper. For example, consider the declarations
170 data family Map k :: * -> *
171 data instance Map (a, b) v = MapPair (Map a (Pair b v))
173 The tycon to which the datacon MapPair belongs gets a unique internal
174 name of the form :R123Map, and we call it the representation tycon.
175 In contrast, Map is the family tycon (accessible via
176 tyConFamInst_maybe). A coercion allows you to move between
177 representation and family type. It is accessible from :R123Map via
178 tyConFamilyCoercion_maybe and has kind
180 Co123Map a b v :: {Map (a, b) v :=: :R123Map a b v}
182 The wrapper and worker of MapPair get the types
185 $WMapPair :: forall a b v. Map a (Map a b v) -> Map (a, b) v
186 $WMapPair a b v = MapPair a b v `cast` sym (Co123Map a b v)
189 MapPair :: forall a b v. Map a (Map a b v) -> :R123Map a b v
191 This coercion is conditionally applied by wrapFamInstBody.
193 It's a bit more complicated if the data instance is a GADT as well!
195 data instance T [a] where
196 T1 :: forall b. b -> T [Maybe b]
198 Co7T a :: T [a] ~ :R7T a
203 $WT1 :: forall b. b -> T [Maybe b]
204 $WT1 b v = T1 (Maybe b) b (Maybe b) v
205 `cast` sym (Co7T (Maybe b))
208 T1 :: forall c b. (c ~ Maybe b) => b -> :R7T c
211 mkDataConIds :: Name -> Name -> DataCon -> DataConIds
212 mkDataConIds wrap_name wkr_name data_con
213 | isNewTyCon tycon -- Newtype, only has a worker
214 = DCIds Nothing nt_work_id
216 | any isMarkedStrict all_strict_marks -- Algebraic, needs wrapper
217 || not (null eq_spec) -- NB: LoadIface.ifaceDeclSubBndrs
218 || isFamInstTyCon tycon -- depends on this test
219 = DCIds (Just alg_wrap_id) wrk_id
221 | otherwise -- Algebraic, no wrapper
222 = DCIds Nothing wrk_id
224 (univ_tvs, ex_tvs, eq_spec,
225 theta, orig_arg_tys, res_ty) = dataConFullSig data_con
226 tycon = dataConTyCon data_con -- The representation TyCon (not family)
228 ----------- Worker (algebraic data types only) --------------
229 -- The *worker* for the data constructor is the function that
230 -- takes the representation arguments and builds the constructor.
231 wrk_id = mkGlobalId (DataConWorkId data_con) wkr_name
232 (dataConRepType data_con) wkr_info
234 wkr_arity = dataConRepArity data_con
235 wkr_info = noCafIdInfo
236 `setArityInfo` wkr_arity
237 `setAllStrictnessInfo` Just wkr_sig
238 `setUnfoldingInfo` evaldUnfolding -- Record that it's evaluated,
241 wkr_sig = mkStrictSig (mkTopDmdType (replicate wkr_arity topDmd) cpr_info)
242 -- Note [Data-con worker strictness]
243 -- Notice that we do *not* say the worker is strict
244 -- even if the data constructor is declared strict
245 -- e.g. data T = MkT !(Int,Int)
246 -- Why? Because the *wrapper* is strict (and its unfolding has case
247 -- expresssions that do the evals) but the *worker* itself is not.
248 -- If we pretend it is strict then when we see
249 -- case x of y -> $wMkT y
250 -- the simplifier thinks that y is "sure to be evaluated" (because
251 -- $wMkT is strict) and drops the case. No, $wMkT is not strict.
253 -- When the simplifer sees a pattern
254 -- case e of MkT x -> ...
255 -- it uses the dataConRepStrictness of MkT to mark x as evaluated;
256 -- but that's fine... dataConRepStrictness comes from the data con
257 -- not from the worker Id.
259 cpr_info | isProductTyCon tycon &&
262 wkr_arity <= mAX_CPR_SIZE = retCPR
264 -- RetCPR is only true for products that are real data types;
265 -- that is, not unboxed tuples or [non-recursive] newtypes
267 ----------- Workers for newtypes --------------
268 nt_work_id = mkGlobalId (DataConWrapId data_con) wkr_name wrap_ty nt_work_info
269 nt_work_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
270 `setArityInfo` 1 -- Arity 1
271 `setUnfoldingInfo` newtype_unf
272 newtype_unf = ASSERT( isVanillaDataCon data_con &&
273 isSingleton orig_arg_tys )
274 -- No existentials on a newtype, but it can have a context
275 -- e.g. newtype Eq a => T a = MkT (...)
276 mkCompulsoryUnfolding $
277 mkLams wrap_tvs $ Lam id_arg1 $
278 wrapNewTypeBody tycon res_ty_args
281 id_arg1 = ASSERT( not (null orig_arg_tys) ) mkTemplateLocal 1 (head orig_arg_tys)
283 ----------- Wrapper --------------
284 -- We used to include the stupid theta in the wrapper's args
285 -- but now we don't. Instead the type checker just injects these
286 -- extra constraints where necessary.
287 wrap_tvs = (univ_tvs `minusList` map fst eq_spec) ++ ex_tvs
288 res_ty_args = substTyVars (mkTopTvSubst eq_spec) univ_tvs
289 dict_tys = mkPredTys theta
290 wrap_ty = mkForAllTys wrap_tvs $ mkFunTys dict_tys $
291 mkFunTys orig_arg_tys $ res_ty
292 -- NB: watch out here if you allow user-written equality
293 -- constraints in data constructor signatures
295 ----------- Wrappers for algebraic data types --------------
296 alg_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty alg_wrap_info
297 alg_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
298 `setArityInfo` wrap_arity
299 -- It's important to specify the arity, so that partial
300 -- applications are treated as values
301 `setUnfoldingInfo` wrap_unf
302 `setAllStrictnessInfo` Just wrap_sig
304 all_strict_marks = dataConExStricts data_con ++ dataConStrictMarks data_con
305 wrap_sig = mkStrictSig (mkTopDmdType arg_dmds cpr_info)
306 arg_dmds = map mk_dmd all_strict_marks
307 mk_dmd str | isMarkedStrict str = evalDmd
308 | otherwise = lazyDmd
309 -- The Cpr info can be important inside INLINE rhss, where the
310 -- wrapper constructor isn't inlined.
311 -- And the argument strictness can be important too; we
312 -- may not inline a contructor when it is partially applied.
314 -- data W = C !Int !Int !Int
315 -- ...(let w = C x in ...(w p q)...)...
316 -- we want to see that w is strict in its two arguments
318 wrap_unf = mkTopUnfolding $ Note InlineMe $
320 mkLams dict_args $ mkLams id_args $
321 foldr mk_case con_app
322 (zip (dict_args ++ id_args) all_strict_marks)
325 con_app _ rep_ids = wrapFamInstBody tycon res_ty_args $
326 Var wrk_id `mkTyApps` res_ty_args
328 `mkTyApps` map snd eq_spec -- Equality evidence
329 `mkVarApps` reverse rep_ids
331 (dict_args,i2) = mkLocals 1 dict_tys
332 (id_args,i3) = mkLocals i2 orig_arg_tys
336 :: (Id, StrictnessMark) -- Arg, strictness
337 -> (Int -> [Id] -> CoreExpr) -- Body
338 -> Int -- Next rep arg id
339 -> [Id] -- Rep args so far, reversed
341 mk_case (arg,strict) body i rep_args
343 NotMarkedStrict -> body i (arg:rep_args)
345 | isUnLiftedType (idType arg) -> body i (arg:rep_args)
347 Case (Var arg) arg res_ty [(DEFAULT,[], body i (arg:rep_args))]
350 -> unboxProduct i (Var arg) (idType arg) the_body
352 the_body i con_args = body i (reverse con_args ++ rep_args)
354 mAX_CPR_SIZE :: Arity
356 -- We do not treat very big tuples as CPR-ish:
357 -- a) for a start we get into trouble because there aren't
358 -- "enough" unboxed tuple types (a tiresome restriction,
360 -- b) more importantly, big unboxed tuples get returned mainly
361 -- on the stack, and are often then allocated in the heap
362 -- by the caller. So doing CPR for them may in fact make
365 mkLocals i tys = (zipWith mkTemplateLocal [i..i+n-1] tys, i+n)
371 %************************************************************************
373 \subsection{Record selectors}
375 %************************************************************************
377 We're going to build a record selector unfolding that looks like this:
379 data T a b c = T1 { ..., op :: a, ...}
380 | T2 { ..., op :: a, ...}
383 sel = /\ a b c -> \ d -> case d of
388 Similarly for newtypes
390 newtype N a = MkN { unN :: a->a }
393 unN n = coerce (a->a) n
395 We need to take a little care if the field has a polymorphic type:
397 data R = R { f :: forall a. a->a }
401 f :: forall a. R -> a -> a
402 f = /\ a \ r = case r of
405 (not f :: R -> forall a. a->a, which gives the type inference mechanism
406 problems at call sites)
408 Similarly for (recursive) newtypes
410 newtype N = MkN { unN :: forall a. a->a }
412 unN :: forall b. N -> b -> b
413 unN = /\b -> \n:N -> (coerce (forall a. a->a) n)
416 Note [Naughty record selectors]
417 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
418 A "naughty" field is one for which we can't define a record
419 selector, because an existential type variable would escape. For example:
420 data T = forall a. MkT { x,y::a }
421 We obviously can't define
423 Nevertheless we *do* put a RecordSelId into the type environment
424 so that if the user tries to use 'x' as a selector we can bleat
425 helpfully, rather than saying unhelpfully that 'x' is not in scope.
426 Hence the sel_naughty flag, to identify record selectors that don't really exist.
428 In general, a field is naughty if its type mentions a type variable that
429 isn't in the result type of the constructor.
431 Note [GADT record selectors]
432 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
433 For GADTs, we require that all constructors with a common field 'f' have the same
434 result type (modulo alpha conversion). [Checked in TcTyClsDecls.checkValidTyCon]
437 T1 { f :: a } :: T [a]
438 T2 { f :: a, y :: b } :: T [a]
439 and now the selector takes that type as its argument:
440 f :: forall a. T [a] -> a
444 Note the forall'd tyvars of the selector are just the free tyvars
445 of the result type; there may be other tyvars in the constructor's
446 type (e.g. 'b' in T2).
448 Note [Selector running example]
449 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
450 It's OK to combine GADTs and type families. Here's a running example:
452 data instance T [a] where
453 T1 { fld :: b } :: T [Maybe b]
455 The representation type looks like this
457 T1 { fld :: b } :: :R7T (Maybe b)
459 and there's coercion from the family type to the representation type
460 :CoR7T a :: T [a] ~ :R7T a
462 The selector we want for fld looks like this:
464 fld :: forall b. T [Maybe b] -> b
465 fld = /\b. \(d::T [Maybe b]).
466 case d `cast` :CoR7T (Maybe b) of
469 The scrutinee of the case has type :R7T (Maybe b), which can be
470 gotten by appying the eq_spec to the univ_tvs of the data con.
473 mkRecordSelId :: TyCon -> FieldLabel -> Id
474 mkRecordSelId tycon field_label
475 -- Assumes that all fields with the same field label have the same type
476 | is_naughty = naughty_id
479 is_naughty = not (tyVarsOfType field_ty `subVarSet` data_tv_set)
480 sel_id_details = RecordSelId { sel_tycon = tycon, sel_label = field_label, sel_naughty = is_naughty }
481 -- For a data type family, the tycon is the *instance* TyCon
483 -- Escapist case here for naughty constructors
484 -- We give it no IdInfo, and a type of forall a.a (never looked at)
485 naughty_id = mkGlobalId sel_id_details field_label forall_a_a noCafIdInfo
486 forall_a_a = mkForAllTy alphaTyVar (mkTyVarTy alphaTyVar)
488 -- Normal case starts here
489 sel_id = mkGlobalId sel_id_details field_label selector_ty info
490 data_cons = tyConDataCons tycon
491 data_cons_w_field = filter has_field data_cons -- Can't be empty!
492 has_field con = field_label `elem` dataConFieldLabels con
494 con1 = ASSERT( not (null data_cons_w_field) ) head data_cons_w_field
495 (univ_tvs, _, eq_spec, _, _, data_ty) = dataConFullSig con1
496 -- For a data type family, the data_ty (and hence selector_ty) mentions
497 -- only the family TyCon, not the instance TyCon
498 data_tv_set = tyVarsOfType data_ty
499 data_tvs = varSetElems data_tv_set
500 field_ty = dataConFieldType con1 field_label
502 -- *Very* tiresomely, the selectors are (unnecessarily!) overloaded over
503 -- just the dictionaries in the types of the constructors that contain
504 -- the relevant field. [The Report says that pattern matching on a
505 -- constructor gives the same constraints as applying it.] Urgh.
507 -- However, not all data cons have all constraints (because of
508 -- BuildTyCl.mkDataConStupidTheta). So we need to find all the data cons
509 -- involved in the pattern match and take the union of their constraints.
510 stupid_dict_tys = mkPredTys (dataConsStupidTheta data_cons_w_field)
511 n_stupid_dicts = length stupid_dict_tys
513 (field_tyvars,pre_field_theta,field_tau) = tcSplitSigmaTy field_ty
514 field_theta = filter (not . isEqPred) pre_field_theta
515 field_dict_tys = mkPredTys field_theta
516 n_field_dict_tys = length field_dict_tys
517 -- If the field has a universally quantified type we have to
518 -- be a bit careful. Suppose we have
519 -- data R = R { op :: forall a. Foo a => a -> a }
520 -- Then we can't give op the type
521 -- op :: R -> forall a. Foo a => a -> a
522 -- because the typechecker doesn't understand foralls to the
523 -- right of an arrow. The "right" type to give it is
524 -- op :: forall a. Foo a => R -> a -> a
525 -- But then we must generate the right unfolding too:
526 -- op = /\a -> \dfoo -> \ r ->
529 -- Note that this is exactly the type we'd infer from a user defn
533 selector_ty = mkForAllTys data_tvs $ mkForAllTys field_tyvars $
534 mkFunTys stupid_dict_tys $ mkFunTys field_dict_tys $
535 mkFunTy data_ty field_tau
537 arity = 1 + n_stupid_dicts + n_field_dict_tys
539 (strict_sig, rhs_w_str) = dmdAnalTopRhs sel_rhs
540 -- Use the demand analyser to work out strictness.
541 -- With all this unpackery it's not easy!
544 `setCafInfo` caf_info
546 `setUnfoldingInfo` mkTopUnfolding rhs_w_str
547 `setAllStrictnessInfo` Just strict_sig
549 -- Allocate Ids. We do it a funny way round because field_dict_tys is
550 -- almost always empty. Also note that we use max_dict_tys
551 -- rather than n_dict_tys, because the latter gives an infinite loop:
552 -- n_dict tys depends on the_alts, which depens on arg_ids, which depends
553 -- on arity, which depends on n_dict tys. Sigh! Mega sigh!
554 stupid_dict_ids = mkTemplateLocalsNum 1 stupid_dict_tys
555 max_stupid_dicts = length (tyConStupidTheta tycon)
556 field_dict_base = max_stupid_dicts + 1
557 field_dict_ids = mkTemplateLocalsNum field_dict_base field_dict_tys
558 dict_id_base = field_dict_base + n_field_dict_tys
559 data_id = mkTemplateLocal dict_id_base data_ty
560 scrut_id = mkTemplateLocal (dict_id_base+1) scrut_ty
561 arg_base = dict_id_base + 2
563 the_alts :: [CoreAlt]
564 the_alts = map mk_alt data_cons_w_field -- Already sorted by data-con
565 no_default = length data_cons == length data_cons_w_field -- No default needed
567 default_alt | no_default = []
568 | otherwise = [(DEFAULT, [], error_expr)]
570 -- The default branch may have CAF refs, because it calls recSelError etc.
571 caf_info | no_default = NoCafRefs
572 | otherwise = MayHaveCafRefs
574 sel_rhs = mkLams data_tvs $ mkLams field_tyvars $
575 mkLams stupid_dict_ids $ mkLams field_dict_ids $
576 Lam data_id $ mk_result sel_body
578 scrut_ty_args = substTyVars (mkTopTvSubst eq_spec) univ_tvs
579 scrut_ty = mkTyConApp tycon scrut_ty_args
580 scrut = unwrapFamInstScrut tycon scrut_ty_args (Var data_id)
581 -- First coerce from the type family to the representation type
583 -- NB: A newtype always has a vanilla DataCon; no existentials etc
584 -- data_tys will simply be the dataConUnivTyVars
585 sel_body | isNewTyCon tycon = unwrapNewTypeBody tycon scrut_ty_args scrut
586 | otherwise = Case scrut scrut_id field_ty (default_alt ++ the_alts)
588 mk_result poly_result = mkVarApps (mkVarApps poly_result field_tyvars) field_dict_ids
589 -- We pull the field lambdas to the top, so we need to
590 -- apply them in the body. For example:
591 -- data T = MkT { foo :: forall a. a->a }
593 -- foo :: forall a. T -> a -> a
594 -- foo = /\a. \t:T. case t of { MkT f -> f a }
597 = ASSERT2( data_ty `tcEqType` field_ty, ppr data_con $$ ppr data_ty $$ ppr field_ty )
598 mkReboxingAlt rebox_uniqs data_con (ex_tvs ++ co_tvs ++ arg_vs) rhs
600 -- get pattern binders with types appropriately instantiated
601 arg_uniqs = map mkBuiltinUnique [arg_base..]
602 (ex_tvs, co_tvs, arg_vs) = dataConOrigInstPat arg_uniqs data_con scrut_ty_args
604 rebox_base = arg_base + length ex_tvs + length co_tvs + length arg_vs
605 rebox_uniqs = map mkBuiltinUnique [rebox_base..]
607 -- data T :: *->* where T1 { fld :: Maybe b } -> T [b]
608 -- Hence T1 :: forall a b. (a=[b]) => b -> T a
609 -- fld :: forall b. T [b] -> Maybe b
610 -- fld = /\b.\(t:T[b]). case t of
611 -- T1 b' (c : [b]=[b']) (x:Maybe b')
612 -- -> x `cast` Maybe (sym (right c))
615 -- Generate the refinement for b'=b,
616 -- and apply to (Maybe b'), to get (Maybe b)
617 Succeeded refinement = gadtRefine emptyRefinement ex_tvs co_tvs
618 the_arg_id_ty = idType the_arg_id
619 (rhs, data_ty) = case refineType refinement the_arg_id_ty of
620 Just (co, data_ty) -> (Cast (Var the_arg_id) co, data_ty)
621 Nothing -> (Var the_arg_id, the_arg_id_ty)
623 field_vs = filter (not . isPredTy . idType) arg_vs
624 the_arg_id = assoc "mkRecordSelId:mk_alt" (field_lbls `zip` field_vs) field_label
625 field_lbls = dataConFieldLabels data_con
627 error_expr = mkRuntimeErrorApp rEC_SEL_ERROR_ID field_ty full_msg
628 full_msg = showSDoc (sep [text "No match in record selector", ppr sel_id])
630 -- unbox a product type...
631 -- we will recurse into newtypes, casting along the way, and unbox at the
632 -- first product data constructor we find. e.g.
634 -- data PairInt = PairInt Int Int
635 -- newtype S = MkS PairInt
638 -- If we have e = MkT (MkS (PairInt 0 1)) and some body expecting a list of
639 -- ids, we get (modulo int passing)
641 -- case (e `cast` CoT) `cast` CoS of
642 -- PairInt a b -> body [a,b]
644 -- The Ints passed around are just for creating fresh locals
645 unboxProduct :: Int -> CoreExpr -> Type -> (Int -> [Id] -> CoreExpr) -> CoreExpr
646 unboxProduct i arg arg_ty body
649 result = mkUnpackCase the_id arg con_args boxing_con rhs
650 (_tycon, _tycon_args, boxing_con, tys) = deepSplitProductType "unboxProduct" arg_ty
651 ([the_id], i') = mkLocals i [arg_ty]
652 (con_args, i'') = mkLocals i' tys
653 rhs = body i'' con_args
655 mkUnpackCase :: Id -> CoreExpr -> [Id] -> DataCon -> CoreExpr -> CoreExpr
656 -- (mkUnpackCase x e args Con body)
658 -- case (e `cast` ...) of bndr { Con args -> body }
660 -- the type of the bndr passed in is irrelevent
661 mkUnpackCase bndr arg unpk_args boxing_con body
662 = Case cast_arg (setIdType bndr bndr_ty) (exprType body) [(DataAlt boxing_con, unpk_args, body)]
664 (cast_arg, bndr_ty) = go (idType bndr) arg
666 | (tycon, tycon_args, _, _) <- splitProductType "mkUnpackCase" ty
667 , isNewTyCon tycon && not (isRecursiveTyCon tycon)
668 = go (newTyConInstRhs tycon tycon_args)
669 (unwrapNewTypeBody tycon tycon_args arg)
670 | otherwise = (arg, ty)
673 reboxProduct :: [Unique] -- uniques to create new local binders
674 -> Type -- type of product to box
675 -> ([Unique], -- remaining uniques
676 CoreExpr, -- boxed product
677 [Id]) -- Ids being boxed into product
680 (_tycon, _tycon_args, _pack_con, con_arg_tys) = deepSplitProductType "reboxProduct" ty
682 us' = dropList con_arg_tys us
684 arg_ids = zipWith (mkSysLocal FSLIT("rb")) us con_arg_tys
686 bind_rhs = mkProductBox arg_ids ty
689 (us', bind_rhs, arg_ids)
691 mkProductBox :: [Id] -> Type -> CoreExpr
692 mkProductBox arg_ids ty
695 (tycon, tycon_args, pack_con, _con_arg_tys) = splitProductType "mkProductBox" ty
698 | isNewTyCon tycon && not (isRecursiveTyCon tycon)
699 = wrap (mkProductBox arg_ids (newTyConInstRhs tycon tycon_args))
700 | otherwise = mkConApp pack_con (map Type tycon_args ++ map Var arg_ids)
702 wrap expr = wrapNewTypeBody tycon tycon_args expr
705 -- (mkReboxingAlt us con xs rhs) basically constructs the case
706 -- alternative (con, xs, rhs)
707 -- but it does the reboxing necessary to construct the *source*
708 -- arguments, xs, from the representation arguments ys.
710 -- data T = MkT !(Int,Int) Bool
712 -- mkReboxingAlt MkT [x,b] r
713 -- = (DataAlt MkT, [y::Int,z::Int,b], let x = (y,z) in r)
715 -- mkDataAlt should really be in DataCon, but it can't because
716 -- it manipulates CoreSyn.
719 :: [Unique] -- Uniques for the new Ids
721 -> [Var] -- Source-level args, including existential dicts
725 mkReboxingAlt us con args rhs
726 | not (any isMarkedUnboxed stricts)
727 = (DataAlt con, args, rhs)
731 (binds, args') = go args stricts us
733 (DataAlt con, args', mkLets binds rhs)
736 stricts = dataConExStricts con ++ dataConStrictMarks con
738 go [] _stricts _us = ([], [])
740 -- Type variable case
741 go (arg:args) stricts us
743 = let (binds, args') = go args stricts us
744 in (binds, arg:args')
746 -- Term variable case
747 go (arg:args) (str:stricts) us
748 | isMarkedUnboxed str
750 let (binds, unpacked_args') = go args stricts us'
751 (us', bind_rhs, unpacked_args) = reboxProduct us (idType arg)
753 (NonRec arg bind_rhs : binds, unpacked_args ++ unpacked_args')
755 = let (binds, args') = go args stricts us
756 in (binds, arg:args')
760 %************************************************************************
762 \subsection{Dictionary selectors}
764 %************************************************************************
766 Selecting a field for a dictionary. If there is just one field, then
767 there's nothing to do.
769 Dictionary selectors may get nested forall-types. Thus:
772 op :: forall b. Ord b => a -> b -> b
774 Then the top-level type for op is
776 op :: forall a. Foo a =>
780 This is unlike ordinary record selectors, which have all the for-alls
781 at the outside. When dealing with classes it's very convenient to
782 recover the original type signature from the class op selector.
785 mkDictSelId :: Name -> Class -> Id
786 mkDictSelId name clas
787 = mkGlobalId (ClassOpId clas) name sel_ty info
789 sel_ty = mkForAllTys tyvars (mkFunTy (idType dict_id) (idType the_arg_id))
790 -- We can't just say (exprType rhs), because that would give a type
792 -- for a single-op class (after all, the selector is the identity)
793 -- But it's type must expose the representation of the dictionary
794 -- to gat (say) C a -> (a -> a)
798 `setUnfoldingInfo` mkTopUnfolding rhs
799 `setAllStrictnessInfo` Just strict_sig
801 -- We no longer use 'must-inline' on record selectors. They'll
802 -- inline like crazy if they scrutinise a constructor
804 -- The strictness signature is of the form U(AAAVAAAA) -> T
805 -- where the V depends on which item we are selecting
806 -- It's worth giving one, so that absence info etc is generated
807 -- even if the selector isn't inlined
808 strict_sig = mkStrictSig (mkTopDmdType [arg_dmd] TopRes)
809 arg_dmd | isNewTyCon tycon = evalDmd
810 | otherwise = Eval (Prod [ if the_arg_id == id then evalDmd else Abs
813 tycon = classTyCon clas
814 [data_con] = tyConDataCons tycon
815 tyvars = dataConUnivTyVars data_con
816 arg_tys = ASSERT( isVanillaDataCon data_con ) dataConRepArgTys data_con
817 the_arg_id = assoc "MkId.mkDictSelId" (map idName (classSelIds clas) `zip` arg_ids) name
819 pred = mkClassPred clas (mkTyVarTys tyvars)
820 (dict_id:arg_ids) = mkTemplateLocals (mkPredTy pred : arg_tys)
822 rhs = mkLams tyvars (Lam dict_id rhs_body)
823 rhs_body | isNewTyCon tycon = unwrapNewTypeBody tycon (map mkTyVarTy tyvars) (Var dict_id)
824 | otherwise = Case (Var dict_id) dict_id (idType the_arg_id)
825 [(DataAlt data_con, arg_ids, Var the_arg_id)]
829 %************************************************************************
831 Wrapping and unwrapping newtypes and type families
833 %************************************************************************
836 wrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
837 -- The wrapper for the data constructor for a newtype looks like this:
838 -- newtype T a = MkT (a,Int)
839 -- MkT :: forall a. (a,Int) -> T a
840 -- MkT = /\a. \(x:(a,Int)). x `cast` sym (CoT a)
841 -- where CoT is the coercion TyCon assoicated with the newtype
843 -- The call (wrapNewTypeBody T [a] e) returns the
844 -- body of the wrapper, namely
845 -- e `cast` (CoT [a])
847 -- If a coercion constructor is provided in the newtype, then we use
848 -- it, otherwise the wrap/unwrap are both no-ops
850 -- If the we are dealing with a newtype *instance*, we have a second coercion
851 -- identifying the family instance with the constructor of the newtype
852 -- instance. This coercion is applied in any case (ie, composed with the
853 -- coercion constructor of the newtype or applied by itself).
855 wrapNewTypeBody tycon args result_expr
856 = wrapFamInstBody tycon args inner
859 | Just co_con <- newTyConCo_maybe tycon
860 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) result_expr
864 -- When unwrapping, we do *not* apply any family coercion, because this will
865 -- be done via a CoPat by the type checker. We have to do it this way as
866 -- computing the right type arguments for the coercion requires more than just
867 -- a spliting operation (cf, TcPat.tcConPat).
869 unwrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
870 unwrapNewTypeBody tycon args result_expr
871 | Just co_con <- newTyConCo_maybe tycon
872 = mkCoerce (mkTyConApp co_con args) result_expr
876 -- If the type constructor is a representation type of a data instance, wrap
877 -- the expression into a cast adjusting the expression type, which is an
878 -- instance of the representation type, to the corresponding instance of the
879 -- family instance type.
880 -- See Note [Wrappers for data instance tycons]
881 wrapFamInstBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
882 wrapFamInstBody tycon args body
883 | Just co_con <- tyConFamilyCoercion_maybe tycon
884 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) body
888 unwrapFamInstScrut :: TyCon -> [Type] -> CoreExpr -> CoreExpr
889 unwrapFamInstScrut tycon args scrut
890 | Just co_con <- tyConFamilyCoercion_maybe tycon
891 = mkCoerce (mkTyConApp co_con args) scrut
897 %************************************************************************
899 \subsection{Primitive operations
901 %************************************************************************
904 mkPrimOpId :: PrimOp -> Id
908 (tyvars,arg_tys,res_ty, arity, strict_sig) = primOpSig prim_op
909 ty = mkForAllTys tyvars (mkFunTys arg_tys res_ty)
910 name = mkWiredInName gHC_PRIM (primOpOcc prim_op)
911 (mkPrimOpIdUnique (primOpTag prim_op))
913 id = mkGlobalId (PrimOpId prim_op) name ty info
916 `setSpecInfo` mkSpecInfo (primOpRules prim_op name)
918 `setAllStrictnessInfo` Just strict_sig
920 -- For each ccall we manufacture a separate CCallOpId, giving it
921 -- a fresh unique, a type that is correct for this particular ccall,
922 -- and a CCall structure that gives the correct details about calling
925 -- The *name* of this Id is a local name whose OccName gives the full
926 -- details of the ccall, type and all. This means that the interface
927 -- file reader can reconstruct a suitable Id
929 mkFCallId :: Unique -> ForeignCall -> Type -> Id
930 mkFCallId uniq fcall ty
931 = ASSERT( isEmptyVarSet (tyVarsOfType ty) )
932 -- A CCallOpId should have no free type variables;
933 -- when doing substitutions won't substitute over it
934 mkGlobalId (FCallId fcall) name ty info
936 occ_str = showSDoc (braces (ppr fcall <+> ppr ty))
937 -- The "occurrence name" of a ccall is the full info about the
938 -- ccall; it is encoded, but may have embedded spaces etc!
940 name = mkFCallName uniq occ_str
944 `setAllStrictnessInfo` Just strict_sig
946 (_, tau) = tcSplitForAllTys ty
947 (arg_tys, _) = tcSplitFunTys tau
948 arity = length arg_tys
949 strict_sig = mkStrictSig (mkTopDmdType (replicate arity evalDmd) TopRes)
951 -- Tick boxes and breakpoints are both represented as TickBoxOpIds,
952 -- except for the type:
954 -- a plain HPC tick box has type (State# RealWorld)
955 -- a breakpoint Id has type forall a.a
957 -- The breakpoint Id will be applied to a list of arbitrary free variables,
958 -- which is why it needs a polymorphic type.
960 mkTickBoxOpId :: Unique -> Module -> TickBoxId -> Id
961 mkTickBoxOpId uniq mod ix = mkTickBox' uniq mod ix realWorldStatePrimTy
963 mkBreakPointOpId :: Unique -> Module -> TickBoxId -> Id
964 mkBreakPointOpId uniq mod ix = mkTickBox' uniq mod ix ty
965 where ty = mkSigmaTy [openAlphaTyVar] [] openAlphaTy
967 mkTickBox' uniq mod ix ty = mkGlobalId (TickBoxOpId tickbox) name ty info
969 tickbox = TickBox mod ix
970 occ_str = showSDoc (braces (ppr tickbox))
971 name = mkTickBoxOpName uniq occ_str
976 %************************************************************************
978 \subsection{DictFuns and default methods}
980 %************************************************************************
982 Important notes about dict funs and default methods
983 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
984 Dict funs and default methods are *not* ImplicitIds. Their definition
985 involves user-written code, so we can't figure out their strictness etc
986 based on fixed info, as we can for constructors and record selectors (say).
988 We build them as LocalIds, but with External Names. This ensures that
989 they are taken to account by free-variable finding and dependency
990 analysis (e.g. CoreFVs.exprFreeVars).
992 Why shouldn't they be bound as GlobalIds? Because, in particular, if
993 they are globals, the specialiser floats dict uses above their defns,
994 which prevents good simplifications happening. Also the strictness
995 analyser treats a occurrence of a GlobalId as imported and assumes it
996 contains strictness in its IdInfo, which isn't true if the thing is
997 bound in the same module as the occurrence.
999 It's OK for dfuns to be LocalIds, because we form the instance-env to
1000 pass on to the next module (md_insts) in CoreTidy, afer tidying
1001 and globalising the top-level Ids.
1003 BUT make sure they are *exported* LocalIds (mkExportedLocalId) so
1004 that they aren't discarded by the occurrence analyser.
1007 mkDefaultMethodId dm_name ty = mkExportedLocalId dm_name ty
1009 mkDictFunId :: Name -- Name to use for the dict fun;
1016 mkDictFunId dfun_name inst_tyvars dfun_theta clas inst_tys
1017 = mkExportedLocalId dfun_name dfun_ty
1019 dfun_ty = mkSigmaTy inst_tyvars dfun_theta (mkDictTy clas inst_tys)
1021 {- 1 dec 99: disable the Mark Jones optimisation for the sake
1022 of compatibility with Hugs.
1023 See `types/InstEnv' for a discussion related to this.
1025 (class_tyvars, sc_theta, _, _) = classBigSig clas
1026 not_const (clas, tys) = not (isEmptyVarSet (tyVarsOfTypes tys))
1027 sc_theta' = substClasses (zipTopTvSubst class_tyvars inst_tys) sc_theta
1028 dfun_theta = case inst_decl_theta of
1029 [] -> [] -- If inst_decl_theta is empty, then we don't
1030 -- want to have any dict arguments, so that we can
1031 -- expose the constant methods.
1033 other -> nub (inst_decl_theta ++ filter not_const sc_theta')
1034 -- Otherwise we pass the superclass dictionaries to
1035 -- the dictionary function; the Mark Jones optimisation.
1037 -- NOTE the "nub". I got caught by this one:
1038 -- class Monad m => MonadT t m where ...
1039 -- instance Monad m => MonadT (EnvT env) m where ...
1040 -- Here, the inst_decl_theta has (Monad m); but so
1041 -- does the sc_theta'!
1043 -- NOTE the "not_const". I got caught by this one too:
1044 -- class Foo a => Baz a b where ...
1045 -- instance Wob b => Baz T b where..
1046 -- Now sc_theta' has Foo T
1051 %************************************************************************
1053 \subsection{Un-definable}
1055 %************************************************************************
1057 These Ids can't be defined in Haskell. They could be defined in
1058 unfoldings in the wired-in GHC.Prim interface file, but we'd have to
1059 ensure that they were definitely, definitely inlined, because there is
1060 no curried identifier for them. That's what mkCompulsoryUnfolding
1061 does. If we had a way to get a compulsory unfolding from an interface
1062 file, we could do that, but we don't right now.
1064 unsafeCoerce# isn't so much a PrimOp as a phantom identifier, that
1065 just gets expanded into a type coercion wherever it occurs. Hence we
1066 add it as a built-in Id with an unfolding here.
1068 The type variables we use here are "open" type variables: this means
1069 they can unify with both unlifted and lifted types. Hence we provide
1070 another gun with which to shoot yourself in the foot.
1073 mkWiredInIdName mod fs uniq id
1074 = mkWiredInName mod (mkOccNameFS varName fs) uniq (AnId id) UserSyntax
1076 unsafeCoerceName = mkWiredInIdName gHC_PRIM FSLIT("unsafeCoerce#") unsafeCoerceIdKey unsafeCoerceId
1077 nullAddrName = mkWiredInIdName gHC_PRIM FSLIT("nullAddr#") nullAddrIdKey nullAddrId
1078 seqName = mkWiredInIdName gHC_PRIM FSLIT("seq") seqIdKey seqId
1079 realWorldName = mkWiredInIdName gHC_PRIM FSLIT("realWorld#") realWorldPrimIdKey realWorldPrimId
1080 lazyIdName = mkWiredInIdName gHC_BASE FSLIT("lazy") lazyIdKey lazyId
1082 errorName = mkWiredInIdName gHC_ERR FSLIT("error") errorIdKey eRROR_ID
1083 recSelErrorName = mkWiredInIdName gHC_ERR FSLIT("recSelError") recSelErrorIdKey rEC_SEL_ERROR_ID
1084 runtimeErrorName = mkWiredInIdName gHC_ERR FSLIT("runtimeError") runtimeErrorIdKey rUNTIME_ERROR_ID
1085 irrefutPatErrorName = mkWiredInIdName gHC_ERR FSLIT("irrefutPatError") irrefutPatErrorIdKey iRREFUT_PAT_ERROR_ID
1086 recConErrorName = mkWiredInIdName gHC_ERR FSLIT("recConError") recConErrorIdKey rEC_CON_ERROR_ID
1087 patErrorName = mkWiredInIdName gHC_ERR FSLIT("patError") patErrorIdKey pAT_ERROR_ID
1088 noMethodBindingErrorName = mkWiredInIdName gHC_ERR FSLIT("noMethodBindingError")
1089 noMethodBindingErrorIdKey nO_METHOD_BINDING_ERROR_ID
1090 nonExhaustiveGuardsErrorName
1091 = mkWiredInIdName gHC_ERR FSLIT("nonExhaustiveGuardsError")
1092 nonExhaustiveGuardsErrorIdKey nON_EXHAUSTIVE_GUARDS_ERROR_ID
1096 -- unsafeCoerce# :: forall a b. a -> b
1098 = pcMiscPrelId unsafeCoerceName ty info
1100 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1103 ty = mkForAllTys [openAlphaTyVar,openBetaTyVar]
1104 (mkFunTy openAlphaTy openBetaTy)
1105 [x] = mkTemplateLocals [openAlphaTy]
1106 rhs = mkLams [openAlphaTyVar,openBetaTyVar,x] $
1107 Cast (Var x) (mkUnsafeCoercion openAlphaTy openBetaTy)
1109 -- nullAddr# :: Addr#
1110 -- The reason is is here is because we don't provide
1111 -- a way to write this literal in Haskell.
1113 = pcMiscPrelId nullAddrName addrPrimTy info
1115 info = noCafIdInfo `setUnfoldingInfo`
1116 mkCompulsoryUnfolding (Lit nullAddrLit)
1119 = pcMiscPrelId seqName ty info
1121 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1124 ty = mkForAllTys [alphaTyVar,openBetaTyVar]
1125 (mkFunTy alphaTy (mkFunTy openBetaTy openBetaTy))
1126 [x,y] = mkTemplateLocals [alphaTy, openBetaTy]
1127 rhs = mkLams [alphaTyVar,openBetaTyVar,x,y] (Case (Var x) x openBetaTy [(DEFAULT, [], Var y)])
1129 -- lazy :: forall a?. a? -> a? (i.e. works for unboxed types too)
1130 -- Used to lazify pseq: pseq a b = a `seq` lazy b
1132 -- Also, no strictness: by being a built-in Id, all the info about lazyId comes from here,
1133 -- not from GHC.Base.hi. This is important, because the strictness
1134 -- analyser will spot it as strict!
1136 -- Also no unfolding in lazyId: it gets "inlined" by a HACK in the worker/wrapperpass
1137 -- (see WorkWrap.wwExpr)
1138 -- We could use inline phases to do this, but that would be vulnerable to changes in
1139 -- phase numbering....we must inline precisely after strictness analysis.
1141 = pcMiscPrelId lazyIdName ty info
1144 ty = mkForAllTys [alphaTyVar] (mkFunTy alphaTy alphaTy)
1146 lazyIdUnfolding :: CoreExpr -- Used to expand 'lazyId' after strictness anal
1147 lazyIdUnfolding = mkLams [openAlphaTyVar,x] (Var x)
1149 [x] = mkTemplateLocals [openAlphaTy]
1152 @realWorld#@ used to be a magic literal, \tr{void#}. If things get
1153 nasty as-is, change it back to a literal (@Literal@).
1155 voidArgId is a Local Id used simply as an argument in functions
1156 where we just want an arg to avoid having a thunk of unlifted type.
1158 x = \ void :: State# RealWorld -> (# p, q #)
1160 This comes up in strictness analysis
1163 realWorldPrimId -- :: State# RealWorld
1164 = pcMiscPrelId realWorldName realWorldStatePrimTy
1165 (noCafIdInfo `setUnfoldingInfo` evaldUnfolding)
1166 -- The evaldUnfolding makes it look that realWorld# is evaluated
1167 -- which in turn makes Simplify.interestingArg return True,
1168 -- which in turn makes INLINE things applied to realWorld# likely
1171 voidArgId -- :: State# RealWorld
1172 = mkSysLocal FSLIT("void") voidArgIdKey realWorldStatePrimTy
1176 %************************************************************************
1178 \subsection[PrelVals-error-related]{@error@ and friends; @trace@}
1180 %************************************************************************
1182 GHC randomly injects these into the code.
1184 @patError@ is just a version of @error@ for pattern-matching
1185 failures. It knows various ``codes'' which expand to longer
1186 strings---this saves space!
1188 @absentErr@ is a thing we put in for ``absent'' arguments. They jolly
1189 well shouldn't be yanked on, but if one is, then you will get a
1190 friendly message from @absentErr@ (rather than a totally random
1193 @parError@ is a special version of @error@ which the compiler does
1194 not know to be a bottoming Id. It is used in the @_par_@ and @_seq_@
1195 templates, but we don't ever expect to generate code for it.
1199 :: Id -- Should be of type (forall a. Addr# -> a)
1200 -- where Addr# points to a UTF8 encoded string
1201 -> Type -- The type to instantiate 'a'
1202 -> String -- The string to print
1205 mkRuntimeErrorApp err_id res_ty err_msg
1206 = mkApps (Var err_id) [Type res_ty, err_string]
1208 err_string = Lit (mkStringLit err_msg)
1210 rEC_SEL_ERROR_ID = mkRuntimeErrorId recSelErrorName
1211 rUNTIME_ERROR_ID = mkRuntimeErrorId runtimeErrorName
1212 iRREFUT_PAT_ERROR_ID = mkRuntimeErrorId irrefutPatErrorName
1213 rEC_CON_ERROR_ID = mkRuntimeErrorId recConErrorName
1214 pAT_ERROR_ID = mkRuntimeErrorId patErrorName
1215 nO_METHOD_BINDING_ERROR_ID = mkRuntimeErrorId noMethodBindingErrorName
1216 nON_EXHAUSTIVE_GUARDS_ERROR_ID = mkRuntimeErrorId nonExhaustiveGuardsErrorName
1218 -- The runtime error Ids take a UTF8-encoded string as argument
1219 mkRuntimeErrorId name = pc_bottoming_Id name runtimeErrorTy
1220 runtimeErrorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTy addrPrimTy openAlphaTy)
1224 eRROR_ID = pc_bottoming_Id errorName errorTy
1227 errorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTys [mkListTy charTy] openAlphaTy)
1228 -- Notice the openAlphaTyVar. It says that "error" can be applied
1229 -- to unboxed as well as boxed types. This is OK because it never
1230 -- returns, so the return type is irrelevant.
1234 %************************************************************************
1236 \subsection{Utilities}
1238 %************************************************************************
1241 pcMiscPrelId :: Name -> Type -> IdInfo -> Id
1242 pcMiscPrelId name ty info
1243 = mkVanillaGlobal name ty info
1244 -- We lie and say the thing is imported; otherwise, we get into
1245 -- a mess with dependency analysis; e.g., core2stg may heave in
1246 -- random calls to GHCbase.unpackPS__. If GHCbase is the module
1247 -- being compiled, then it's just a matter of luck if the definition
1248 -- will be in "the right place" to be in scope.
1250 pc_bottoming_Id name ty
1251 = pcMiscPrelId name ty bottoming_info
1253 bottoming_info = vanillaIdInfo `setAllStrictnessInfo` Just strict_sig
1254 -- Do *not* mark them as NoCafRefs, because they can indeed have
1255 -- CAF refs. For example, pAT_ERROR_ID calls GHC.Err.untangle,
1256 -- which has some CAFs
1257 -- In due course we may arrange that these error-y things are
1258 -- regarded by the GC as permanently live, in which case we
1259 -- can give them NoCaf info. As it is, any function that calls
1260 -- any pc_bottoming_Id will itself have CafRefs, which bloats
1263 strict_sig = mkStrictSig (mkTopDmdType [evalDmd] BotRes)
1264 -- These "bottom" out, no matter what their arguments