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,
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"
62 import Var ( Var, TyVar)
70 import BasicTypes hiding ( SuccessFlag(..) )
78 %************************************************************************
80 \subsection{Wired in Ids}
82 %************************************************************************
86 = [ -- These error-y things are wired in because we don't yet have
87 -- a way to express in an interface file that the result type variable
88 -- is 'open'; that is can be unified with an unboxed type
90 -- [The interface file format now carry such information, but there's
91 -- no way yet of expressing at the definition site for these
92 -- error-reporting functions that they have an 'open'
93 -- result type. -- sof 1/99]
95 eRROR_ID, -- This one isn't used anywhere else in the compiler
96 -- But we still need it in wiredInIds so that when GHC
97 -- compiles a program that mentions 'error' we don't
98 -- import its type from the interface file; we just get
99 -- the Id defined here. Which has an 'open-tyvar' type.
102 iRREFUT_PAT_ERROR_ID,
103 nON_EXHAUSTIVE_GUARDS_ERROR_ID,
104 nO_METHOD_BINDING_ERROR_ID,
111 -- These Ids are exported from GHC.Prim
113 = [ -- These can't be defined in Haskell, but they have
114 -- perfectly reasonable unfoldings in Core
122 %************************************************************************
124 \subsection{Data constructors}
126 %************************************************************************
128 The wrapper for a constructor is an ordinary top-level binding that evaluates
129 any strict args, unboxes any args that are going to be flattened, and calls
132 We're going to build a constructor that looks like:
134 data (Data a, C b) => T a b = T1 !a !Int b
137 \d1::Data a, d2::C b ->
138 \p q r -> case p of { p ->
140 Con T1 [a,b] [p,q,r]}}
144 * d2 is thrown away --- a context in a data decl is used to make sure
145 one *could* construct dictionaries at the site the constructor
146 is used, but the dictionary isn't actually used.
148 * We have to check that we can construct Data dictionaries for
149 the types a and Int. Once we've done that we can throw d1 away too.
151 * We use (case p of q -> ...) to evaluate p, rather than "seq" because
152 all that matters is that the arguments are evaluated. "seq" is
153 very careful to preserve evaluation order, which we don't need
156 You might think that we could simply give constructors some strictness
157 info, like PrimOps, and let CoreToStg do the let-to-case transformation.
158 But we don't do that because in the case of primops and functions strictness
159 is a *property* not a *requirement*. In the case of constructors we need to
160 do something active to evaluate the argument.
162 Making an explicit case expression allows the simplifier to eliminate
163 it in the (common) case where the constructor arg is already evaluated.
165 [Wrappers for data instance tycons]
166 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
167 In the case of data instances, the wrapper also applies the coercion turning
168 the representation type into the family instance type to cast the result of
169 the wrapper. For example, consider the declarations
171 data family Map k :: * -> *
172 data instance Map (a, b) v = MapPair (Map a (Pair b v))
174 The tycon to which the datacon MapPair belongs gets a unique internal name of
175 the form :R123Map, and we call it the representation tycon. In contrast, Map
176 is the family tycon (accessible via tyConFamInst_maybe). The wrapper and work
177 of MapPair get the types
179 $WMapPair :: forall a b v. Map a (Map a b v) -> Map (a, b) v
180 $wMapPair :: forall a b v. Map a (Map a b v) -> :R123Map a b v
182 which implies that the wrapper code will have to apply the coercion moving
183 between representation and family type. It is accessible via
184 tyConFamilyCoercion_maybe and has kind
186 Co123Map a b v :: {Map (a, b) v :=: :R123Map a b v}
188 This coercion is conditionally applied by wrapFamInstBody.
191 mkDataConIds :: Name -> Name -> DataCon -> DataConIds
192 mkDataConIds wrap_name wkr_name data_con
194 = DCIds Nothing nt_work_id -- Newtype, only has a worker
196 | any isMarkedStrict all_strict_marks -- Algebraic, needs wrapper
197 || not (null eq_spec) -- NB: LoadIface.ifaceDeclSubBndrs
198 || isFamInstTyCon tycon -- depends on this test
199 = DCIds (Just alg_wrap_id) wrk_id
201 | otherwise -- Algebraic, no wrapper
202 = DCIds Nothing wrk_id
204 (univ_tvs, ex_tvs, eq_spec,
205 theta, orig_arg_tys) = dataConFullSig data_con
206 tycon = dataConTyCon data_con
208 ----------- Wrapper --------------
209 -- We used to include the stupid theta in the wrapper's args
210 -- but now we don't. Instead the type checker just injects these
211 -- extra constraints where necessary.
212 wrap_tvs = (univ_tvs `minusList` map fst eq_spec) ++ ex_tvs
213 subst = mkTopTvSubst eq_spec
214 famSubst = ASSERT( length (tyConTyVars tycon ) ==
215 length (mkTyVarTys univ_tvs) )
216 zipTopTvSubst (tyConTyVars tycon) (mkTyVarTys univ_tvs)
217 -- substitution mapping the type constructor's type
218 -- arguments to the universals of the data constructor
219 -- (crucial when type checking interfaces)
220 dict_tys = mkPredTys theta
221 result_ty_args = substTyVars subst univ_tvs
222 result_ty = case tyConFamInst_maybe tycon of
223 -- ordinary constructor
224 Nothing -> mkTyConApp tycon result_ty_args
225 -- family instance constructor
228 mkTyConApp familyTyCon ( substTys subst
231 wrap_ty = mkForAllTys wrap_tvs $ mkFunTys dict_tys $
232 mkFunTys orig_arg_tys $ result_ty
233 -- NB: watch out here if you allow user-written equality
234 -- constraints in data constructor signatures
236 ----------- Worker (algebraic data types only) --------------
237 -- The *worker* for the data constructor is the function that
238 -- takes the representation arguments and builds the constructor.
239 wrk_id = mkGlobalId (DataConWorkId data_con) wkr_name
240 (dataConRepType data_con) wkr_info
242 wkr_arity = dataConRepArity data_con
243 wkr_info = noCafIdInfo
244 `setArityInfo` wkr_arity
245 `setAllStrictnessInfo` Just wkr_sig
246 `setUnfoldingInfo` evaldUnfolding -- Record that it's evaluated,
249 wkr_sig = mkStrictSig (mkTopDmdType (replicate wkr_arity topDmd) cpr_info)
250 -- Note [Data-con worker strictness]
251 -- Notice that we do *not* say the worker is strict
252 -- even if the data constructor is declared strict
253 -- e.g. data T = MkT !(Int,Int)
254 -- Why? Because the *wrapper* is strict (and its unfolding has case
255 -- expresssions that do the evals) but the *worker* itself is not.
256 -- If we pretend it is strict then when we see
257 -- case x of y -> $wMkT y
258 -- the simplifier thinks that y is "sure to be evaluated" (because
259 -- $wMkT is strict) and drops the case. No, $wMkT is not strict.
261 -- When the simplifer sees a pattern
262 -- case e of MkT x -> ...
263 -- it uses the dataConRepStrictness of MkT to mark x as evaluated;
264 -- but that's fine... dataConRepStrictness comes from the data con
265 -- not from the worker Id.
267 cpr_info | isProductTyCon tycon &&
270 wkr_arity <= mAX_CPR_SIZE = retCPR
272 -- RetCPR is only true for products that are real data types;
273 -- that is, not unboxed tuples or [non-recursive] newtypes
275 ----------- Workers for newtypes --------------
276 nt_work_id = mkGlobalId (DataConWrapId data_con) wkr_name wrap_ty nt_work_info
277 nt_work_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
278 `setArityInfo` 1 -- Arity 1
279 `setUnfoldingInfo` newtype_unf
280 newtype_unf = ASSERT( isVanillaDataCon data_con &&
281 isSingleton orig_arg_tys )
282 -- No existentials on a newtype, but it can have a context
283 -- e.g. newtype Eq a => T a = MkT (...)
284 mkCompulsoryUnfolding $
285 mkLams wrap_tvs $ Lam id_arg1 $
286 wrapNewTypeBody tycon result_ty_args
289 id_arg1 = mkTemplateLocal 1 (head orig_arg_tys)
291 ----------- Wrappers for algebraic data types --------------
292 alg_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty alg_wrap_info
293 alg_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
294 `setArityInfo` alg_arity
295 -- It's important to specify the arity, so that partial
296 -- applications are treated as values
297 `setUnfoldingInfo` alg_unf
298 `setAllStrictnessInfo` Just wrap_sig
300 all_strict_marks = dataConExStricts data_con ++ dataConStrictMarks data_con
301 wrap_sig = mkStrictSig (mkTopDmdType arg_dmds cpr_info)
302 arg_dmds = map mk_dmd all_strict_marks
303 mk_dmd str | isMarkedStrict str = evalDmd
304 | otherwise = lazyDmd
305 -- The Cpr info can be important inside INLINE rhss, where the
306 -- wrapper constructor isn't inlined.
307 -- And the argument strictness can be important too; we
308 -- may not inline a contructor when it is partially applied.
310 -- data W = C !Int !Int !Int
311 -- ...(let w = C x in ...(w p q)...)...
312 -- we want to see that w is strict in its two arguments
314 alg_unf = mkTopUnfolding $ Note InlineMe $
316 mkLams dict_args $ mkLams id_args $
317 foldr mk_case con_app
318 (zip (dict_args ++ id_args) all_strict_marks)
321 con_app _ rep_ids = wrapFamInstBody tycon result_ty_args $
322 Var wrk_id `mkTyApps` result_ty_args
324 `mkTyApps` map snd eq_spec
325 `mkVarApps` reverse rep_ids
327 (dict_args,i2) = mkLocals 1 dict_tys
328 (id_args,i3) = mkLocals i2 orig_arg_tys
332 :: (Id, StrictnessMark) -- Arg, strictness
333 -> (Int -> [Id] -> CoreExpr) -- Body
334 -> Int -- Next rep arg id
335 -> [Id] -- Rep args so far, reversed
337 mk_case (arg,strict) body i rep_args
339 NotMarkedStrict -> body i (arg:rep_args)
341 | isUnLiftedType (idType arg) -> body i (arg:rep_args)
343 Case (Var arg) arg result_ty [(DEFAULT,[], body i (arg:rep_args))]
346 -> unboxProduct i (Var arg) (idType arg) the_body
348 the_body i con_args = body i (reverse con_args ++ rep_args)
350 mAX_CPR_SIZE :: Arity
352 -- We do not treat very big tuples as CPR-ish:
353 -- a) for a start we get into trouble because there aren't
354 -- "enough" unboxed tuple types (a tiresome restriction,
356 -- b) more importantly, big unboxed tuples get returned mainly
357 -- on the stack, and are often then allocated in the heap
358 -- by the caller. So doing CPR for them may in fact make
361 mkLocals i tys = (zipWith mkTemplateLocal [i..i+n-1] tys, i+n)
365 -- If the type constructor is a representation type of a data instance, wrap
366 -- the expression into a cast adjusting the expression type, which is an
367 -- instance of the representation type, to the corresponding instance of the
368 -- family instance type.
370 wrapFamInstBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
371 wrapFamInstBody tycon args result_expr
372 | Just co_con <- tyConFamilyCoercion_maybe tycon
373 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) result_expr
379 %************************************************************************
381 \subsection{Record selectors}
383 %************************************************************************
385 We're going to build a record selector unfolding that looks like this:
387 data T a b c = T1 { ..., op :: a, ...}
388 | T2 { ..., op :: a, ...}
391 sel = /\ a b c -> \ d -> case d of
396 Similarly for newtypes
398 newtype N a = MkN { unN :: a->a }
401 unN n = coerce (a->a) n
403 We need to take a little care if the field has a polymorphic type:
405 data R = R { f :: forall a. a->a }
409 f :: forall a. R -> a -> a
410 f = /\ a \ r = case r of
413 (not f :: R -> forall a. a->a, which gives the type inference mechanism
414 problems at call sites)
416 Similarly for (recursive) newtypes
418 newtype N = MkN { unN :: forall a. a->a }
420 unN :: forall b. N -> b -> b
421 unN = /\b -> \n:N -> (coerce (forall a. a->a) n)
424 Note [Naughty record selectors]
425 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
426 A "naughty" field is one for which we can't define a record
427 selector, because an existential type variable would escape. For example:
428 data T = forall a. MkT { x,y::a }
429 We obviously can't define
431 Nevertheless we *do* put a RecordSelId into the type environment
432 so that if the user tries to use 'x' as a selector we can bleat
433 helpfully, rather than saying unhelpfully that 'x' is not in scope.
434 Hence the sel_naughty flag, to identify record selectors that don't really exist.
436 In general, a field is naughty if its type mentions a type variable that
437 isn't in the result type of the constructor.
439 Note [GADT record selectors]
440 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
441 For GADTs, we require that all constructors with a common field 'f' have the same
442 result type (modulo alpha conversion). [Checked in TcTyClsDecls.checkValidTyCon]
445 T1 { f :: a } :: T [a]
446 T2 { f :: a, y :: b } :: T [a]
447 and now the selector takes that type as its argument:
448 f :: forall a. T [a] -> a
452 Note the forall'd tyvars of the selector are just the free tyvars
453 of the result type; there may be other tyvars in the constructor's
454 type (e.g. 'b' in T2).
458 -- Steps for handling "naughty" vs "non-naughty" selectors:
459 -- 1. Determine naughtiness by comparing field type vs result type
460 -- 2. Install naughty ones with selector_ty of type _|_ and fill in mzero for info
461 -- 3. If it's not naughty, do the normal plan.
463 mkRecordSelId :: TyCon -> FieldLabel -> Id
464 mkRecordSelId tycon field_label
465 -- Assumes that all fields with the same field label have the same type
466 | is_naughty = naughty_id
469 is_naughty = not (tyVarsOfType field_ty `subVarSet` res_tv_set)
470 sel_id_details = RecordSelId tycon field_label is_naughty
472 -- Escapist case here for naughty construcotrs
473 -- We give it no IdInfo, and a type of forall a.a (never looked at)
474 naughty_id = mkGlobalId sel_id_details field_label forall_a_a noCafIdInfo
475 forall_a_a = mkForAllTy alphaTyVar (mkTyVarTy alphaTyVar)
477 -- Normal case starts here
478 sel_id = mkGlobalId sel_id_details field_label selector_ty info
479 data_cons = tyConDataCons tycon
480 data_cons_w_field = filter has_field data_cons -- Can't be empty!
481 has_field con = field_label `elem` dataConFieldLabels con
483 con1 = head data_cons_w_field
484 res_tys = dataConResTys con1
485 res_tv_set = tyVarsOfTypes res_tys
486 res_tvs = varSetElems res_tv_set
487 data_ty = mkTyConApp tycon res_tys
488 field_ty = dataConFieldType con1 field_label
490 -- *Very* tiresomely, the selectors are (unnecessarily!) overloaded over
491 -- just the dictionaries in the types of the constructors that contain
492 -- the relevant field. [The Report says that pattern matching on a
493 -- constructor gives the same constraints as applying it.] Urgh.
495 -- However, not all data cons have all constraints (because of
496 -- BuildTyCl.mkDataConStupidTheta). So we need to find all the data cons
497 -- involved in the pattern match and take the union of their constraints.
498 stupid_dict_tys = mkPredTys (dataConsStupidTheta data_cons_w_field)
499 n_stupid_dicts = length stupid_dict_tys
501 (field_tyvars,pre_field_theta,field_tau) = tcSplitSigmaTy field_ty
503 field_theta = filter (not . isEqPred) pre_field_theta
504 field_dict_tys = mkPredTys field_theta
505 n_field_dict_tys = length field_dict_tys
506 -- If the field has a universally quantified type we have to
507 -- be a bit careful. Suppose we have
508 -- data R = R { op :: forall a. Foo a => a -> a }
509 -- Then we can't give op the type
510 -- op :: R -> forall a. Foo a => a -> a
511 -- because the typechecker doesn't understand foralls to the
512 -- right of an arrow. The "right" type to give it is
513 -- op :: forall a. Foo a => R -> a -> a
514 -- But then we must generate the right unfolding too:
515 -- op = /\a -> \dfoo -> \ r ->
518 -- Note that this is exactly the type we'd infer from a user defn
522 selector_ty = mkForAllTys res_tvs $ mkForAllTys field_tyvars $
523 mkFunTys stupid_dict_tys $ mkFunTys field_dict_tys $
524 mkFunTy data_ty field_tau
526 arity = 1 + n_stupid_dicts + n_field_dict_tys
528 (strict_sig, rhs_w_str) = dmdAnalTopRhs sel_rhs
529 -- Use the demand analyser to work out strictness.
530 -- With all this unpackery it's not easy!
533 `setCafInfo` caf_info
535 `setUnfoldingInfo` mkTopUnfolding rhs_w_str
536 `setAllStrictnessInfo` Just strict_sig
538 -- Allocate Ids. We do it a funny way round because field_dict_tys is
539 -- almost always empty. Also note that we use max_dict_tys
540 -- rather than n_dict_tys, because the latter gives an infinite loop:
541 -- n_dict tys depends on the_alts, which depens on arg_ids, which depends
542 -- on arity, which depends on n_dict tys. Sigh! Mega sigh!
543 stupid_dict_ids = mkTemplateLocalsNum 1 stupid_dict_tys
544 max_stupid_dicts = length (tyConStupidTheta tycon)
545 field_dict_base = max_stupid_dicts + 1
546 field_dict_ids = mkTemplateLocalsNum field_dict_base field_dict_tys
547 dict_id_base = field_dict_base + n_field_dict_tys
548 data_id = mkTemplateLocal dict_id_base data_ty
549 arg_base = dict_id_base + 1
551 the_alts :: [CoreAlt]
552 the_alts = map mk_alt data_cons_w_field -- Already sorted by data-con
553 no_default = length data_cons == length data_cons_w_field -- No default needed
555 default_alt | no_default = []
556 | otherwise = [(DEFAULT, [], error_expr)]
558 -- The default branch may have CAF refs, because it calls recSelError etc.
559 caf_info | no_default = NoCafRefs
560 | otherwise = MayHaveCafRefs
562 sel_rhs = mkLams res_tvs $ mkLams field_tyvars $
563 mkLams stupid_dict_ids $ mkLams field_dict_ids $
564 Lam data_id $ mk_result sel_body
566 -- NB: A newtype always has a vanilla DataCon; no existentials etc
567 -- res_tys will simply be the dataConUnivTyVars
568 sel_body | isNewTyCon tycon = unwrapNewTypeBody tycon res_tys (Var data_id)
569 | otherwise = Case (Var data_id) data_id field_ty (default_alt ++ the_alts)
571 mk_result poly_result = mkVarApps (mkVarApps poly_result field_tyvars) field_dict_ids
572 -- We pull the field lambdas to the top, so we need to
573 -- apply them in the body. For example:
574 -- data T = MkT { foo :: forall a. a->a }
576 -- foo :: forall a. T -> a -> a
577 -- foo = /\a. \t:T. case t of { MkT f -> f a }
580 = ASSERT2( res_ty `tcEqType` field_ty, ppr data_con $$ ppr res_ty $$ ppr field_ty )
581 mkReboxingAlt rebox_uniqs data_con (ex_tvs ++ co_tvs ++ arg_vs) rhs
583 -- get pattern binders with types appropriately instantiated
584 arg_uniqs = map mkBuiltinUnique [arg_base..]
585 (ex_tvs, co_tvs, arg_vs) = dataConOrigInstPat arg_uniqs data_con res_tys
587 rebox_base = arg_base + length ex_tvs + length co_tvs + length arg_vs
588 rebox_uniqs = map mkBuiltinUnique [rebox_base..]
590 -- data T :: *->* where T1 { fld :: Maybe b } -> T [b]
591 -- Hence T1 :: forall a b. (a=[b]) => b -> T a
592 -- fld :: forall b. T [b] -> Maybe b
593 -- fld = /\b.\(t:T[b]). case t of
594 -- T1 b' (c : [b]=[b']) (x:Maybe b')
595 -- -> x `cast` Maybe (sym (right c))
598 -- Generate the refinement for b'=b,
599 -- and apply to (Maybe b'), to get (Maybe b)
600 Succeeded refinement = gadtRefine emptyRefinement ex_tvs co_tvs
601 the_arg_id_ty = idType the_arg_id
602 (rhs, res_ty) = case refineType refinement the_arg_id_ty of
603 Just (co, res_ty) -> (Cast (Var the_arg_id) co, res_ty)
604 Nothing -> (Var the_arg_id, the_arg_id_ty)
606 field_vs = filter (not . isPredTy . idType) arg_vs
607 the_arg_id = assoc "mkRecordSelId:mk_alt" (field_lbls `zip` field_vs) field_label
608 field_lbls = dataConFieldLabels data_con
610 error_expr = mkRuntimeErrorApp rEC_SEL_ERROR_ID field_ty full_msg
611 full_msg = showSDoc (sep [text "No match in record selector", ppr sel_id])
613 -- unbox a product type...
614 -- we will recurse into newtypes, casting along the way, and unbox at the
615 -- first product data constructor we find. e.g.
617 -- data PairInt = PairInt Int Int
618 -- newtype S = MkS PairInt
621 -- If we have e = MkT (MkS (PairInt 0 1)) and some body expecting a list of
622 -- ids, we get (modulo int passing)
624 -- case (e `cast` CoT) `cast` CoS of
625 -- PairInt a b -> body [a,b]
627 -- The Ints passed around are just for creating fresh locals
628 unboxProduct :: Int -> CoreExpr -> Type -> (Int -> [Id] -> CoreExpr) -> CoreExpr
629 unboxProduct i arg arg_ty body
632 result = mkUnpackCase the_id arg con_args boxing_con rhs
633 (_tycon, _tycon_args, boxing_con, tys) = deepSplitProductType "unboxProduct" arg_ty
634 ([the_id], i') = mkLocals i [arg_ty]
635 (con_args, i'') = mkLocals i' tys
636 rhs = body i'' con_args
638 mkUnpackCase :: Id -> CoreExpr -> [Id] -> DataCon -> CoreExpr -> CoreExpr
639 -- (mkUnpackCase x e args Con body)
641 -- case (e `cast` ...) of bndr { Con args -> body }
643 -- the type of the bndr passed in is irrelevent
644 mkUnpackCase bndr arg unpk_args boxing_con body
645 = Case cast_arg (setIdType bndr bndr_ty) (exprType body) [(DataAlt boxing_con, unpk_args, body)]
647 (cast_arg, bndr_ty) = go (idType bndr) arg
649 | (tycon, tycon_args, _, _) <- splitProductType "mkUnpackCase" ty
650 , isNewTyCon tycon && not (isRecursiveTyCon tycon)
651 = go (newTyConInstRhs tycon tycon_args)
652 (unwrapNewTypeBody tycon tycon_args arg)
653 | otherwise = (arg, ty)
656 reboxProduct :: [Unique] -- uniques to create new local binders
657 -> Type -- type of product to box
658 -> ([Unique], -- remaining uniques
659 CoreExpr, -- boxed product
660 [Id]) -- Ids being boxed into product
663 (_tycon, _tycon_args, _pack_con, con_arg_tys) = deepSplitProductType "reboxProduct" ty
665 us' = dropList con_arg_tys us
667 arg_ids = zipWith (mkSysLocal FSLIT("rb")) us con_arg_tys
669 bind_rhs = mkProductBox arg_ids ty
672 (us', bind_rhs, arg_ids)
674 mkProductBox :: [Id] -> Type -> CoreExpr
675 mkProductBox arg_ids ty
678 (tycon, tycon_args, pack_con, _con_arg_tys) = splitProductType "mkProductBox" ty
681 | isNewTyCon tycon && not (isRecursiveTyCon tycon)
682 = wrap (mkProductBox arg_ids (newTyConInstRhs tycon tycon_args))
683 | otherwise = mkConApp pack_con (map Type tycon_args ++ map Var arg_ids)
685 wrap expr = wrapNewTypeBody tycon tycon_args expr
688 -- (mkReboxingAlt us con xs rhs) basically constructs the case
689 -- alternative (con, xs, rhs)
690 -- but it does the reboxing necessary to construct the *source*
691 -- arguments, xs, from the representation arguments ys.
693 -- data T = MkT !(Int,Int) Bool
695 -- mkReboxingAlt MkT [x,b] r
696 -- = (DataAlt MkT, [y::Int,z::Int,b], let x = (y,z) in r)
698 -- mkDataAlt should really be in DataCon, but it can't because
699 -- it manipulates CoreSyn.
702 :: [Unique] -- Uniques for the new Ids
704 -> [Var] -- Source-level args, including existential dicts
708 mkReboxingAlt us con args rhs
709 | not (any isMarkedUnboxed stricts)
710 = (DataAlt con, args, rhs)
714 (binds, args') = go args stricts us
716 (DataAlt con, args', mkLets binds rhs)
719 stricts = dataConExStricts con ++ dataConStrictMarks con
721 go [] _stricts _us = ([], [])
723 -- Type variable case
724 go (arg:args) stricts us
726 = let (binds, args') = go args stricts us
727 in (binds, arg:args')
729 -- Term variable case
730 go (arg:args) (str:stricts) us
731 | isMarkedUnboxed str
733 let (binds, unpacked_args') = go args stricts us'
734 (us', bind_rhs, unpacked_args) = reboxProduct us (idType arg)
736 (NonRec arg bind_rhs : binds, unpacked_args ++ unpacked_args')
738 = let (binds, args') = go args stricts us
739 in (binds, arg:args')
743 %************************************************************************
745 \subsection{Dictionary selectors}
747 %************************************************************************
749 Selecting a field for a dictionary. If there is just one field, then
750 there's nothing to do.
752 Dictionary selectors may get nested forall-types. Thus:
755 op :: forall b. Ord b => a -> b -> b
757 Then the top-level type for op is
759 op :: forall a. Foo a =>
763 This is unlike ordinary record selectors, which have all the for-alls
764 at the outside. When dealing with classes it's very convenient to
765 recover the original type signature from the class op selector.
768 mkDictSelId :: Name -> Class -> Id
769 mkDictSelId name clas
770 = mkGlobalId (ClassOpId clas) name sel_ty info
772 sel_ty = mkForAllTys tyvars (mkFunTy (idType dict_id) (idType the_arg_id))
773 -- We can't just say (exprType rhs), because that would give a type
775 -- for a single-op class (after all, the selector is the identity)
776 -- But it's type must expose the representation of the dictionary
777 -- to gat (say) C a -> (a -> a)
781 `setUnfoldingInfo` mkTopUnfolding rhs
782 `setAllStrictnessInfo` Just strict_sig
784 -- We no longer use 'must-inline' on record selectors. They'll
785 -- inline like crazy if they scrutinise a constructor
787 -- The strictness signature is of the form U(AAAVAAAA) -> T
788 -- where the V depends on which item we are selecting
789 -- It's worth giving one, so that absence info etc is generated
790 -- even if the selector isn't inlined
791 strict_sig = mkStrictSig (mkTopDmdType [arg_dmd] TopRes)
792 arg_dmd | isNewTyCon tycon = evalDmd
793 | otherwise = Eval (Prod [ if the_arg_id == id then evalDmd else Abs
796 tycon = classTyCon clas
797 [data_con] = tyConDataCons tycon
798 tyvars = dataConUnivTyVars data_con
799 arg_tys = ASSERT( isVanillaDataCon data_con ) dataConRepArgTys data_con
800 the_arg_id = assoc "MkId.mkDictSelId" (map idName (classSelIds clas) `zip` arg_ids) name
802 pred = mkClassPred clas (mkTyVarTys tyvars)
803 (dict_id:arg_ids) = mkTemplateLocals (mkPredTy pred : arg_tys)
805 rhs = mkLams tyvars (Lam dict_id rhs_body)
806 rhs_body | isNewTyCon tycon = unwrapNewTypeBody tycon (map mkTyVarTy tyvars) (Var dict_id)
807 | otherwise = Case (Var dict_id) dict_id (idType the_arg_id)
808 [(DataAlt data_con, arg_ids, Var the_arg_id)]
810 wrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
811 -- The wrapper for the data constructor for a newtype looks like this:
812 -- newtype T a = MkT (a,Int)
813 -- MkT :: forall a. (a,Int) -> T a
814 -- MkT = /\a. \(x:(a,Int)). x `cast` sym (CoT a)
815 -- where CoT is the coercion TyCon assoicated with the newtype
817 -- The call (wrapNewTypeBody T [a] e) returns the
818 -- body of the wrapper, namely
819 -- e `cast` (CoT [a])
821 -- If a coercion constructor is prodivided in the newtype, then we use
822 -- it, otherwise the wrap/unwrap are both no-ops
824 -- If the we are dealing with a newtype instance, we have a second coercion
825 -- identifying the family instance with the constructor of the newtype
826 -- instance. This coercion is applied in any case (ie, composed with the
827 -- coercion constructor of the newtype or applied by itself).
829 wrapNewTypeBody tycon args result_expr
830 = wrapFamInstBody tycon args inner
833 | Just co_con <- newTyConCo_maybe tycon
834 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) result_expr
838 -- When unwrapping, we do *not* apply any family coercion, because this will
839 -- be done via a CoPat by the type checker. We have to do it this way as
840 -- computing the right type arguments for the coercion requires more than just
841 -- a spliting operation (cf, TcPat.tcConPat).
843 unwrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
844 unwrapNewTypeBody tycon args result_expr
845 | Just co_con <- newTyConCo_maybe tycon
846 = mkCoerce (mkTyConApp co_con args) result_expr
854 %************************************************************************
856 \subsection{Primitive operations
858 %************************************************************************
861 mkPrimOpId :: PrimOp -> Id
865 (tyvars,arg_tys,res_ty, arity, strict_sig) = primOpSig prim_op
866 ty = mkForAllTys tyvars (mkFunTys arg_tys res_ty)
867 name = mkWiredInName gHC_PRIM (primOpOcc prim_op)
868 (mkPrimOpIdUnique (primOpTag prim_op))
870 id = mkGlobalId (PrimOpId prim_op) name ty info
873 `setSpecInfo` mkSpecInfo (primOpRules prim_op name)
875 `setAllStrictnessInfo` Just strict_sig
877 -- For each ccall we manufacture a separate CCallOpId, giving it
878 -- a fresh unique, a type that is correct for this particular ccall,
879 -- and a CCall structure that gives the correct details about calling
882 -- The *name* of this Id is a local name whose OccName gives the full
883 -- details of the ccall, type and all. This means that the interface
884 -- file reader can reconstruct a suitable Id
886 mkFCallId :: Unique -> ForeignCall -> Type -> Id
887 mkFCallId uniq fcall ty
888 = ASSERT( isEmptyVarSet (tyVarsOfType ty) )
889 -- A CCallOpId should have no free type variables;
890 -- when doing substitutions won't substitute over it
891 mkGlobalId (FCallId fcall) name ty info
893 occ_str = showSDoc (braces (ppr fcall <+> ppr ty))
894 -- The "occurrence name" of a ccall is the full info about the
895 -- ccall; it is encoded, but may have embedded spaces etc!
897 name = mkFCallName uniq occ_str
901 `setAllStrictnessInfo` Just strict_sig
903 (_, tau) = tcSplitForAllTys ty
904 (arg_tys, _) = tcSplitFunTys tau
905 arity = length arg_tys
906 strict_sig = mkStrictSig (mkTopDmdType (replicate arity evalDmd) TopRes)
908 mkTickBoxOpId :: Unique
912 mkTickBoxOpId uniq mod ix = mkGlobalId (TickBoxOpId tickbox) name ty info
914 tickbox = TickBox mod ix
915 occ_str = showSDoc (braces (ppr tickbox))
916 name = mkTickBoxOpName uniq occ_str
918 ty = realWorldStatePrimTy
922 %************************************************************************
924 \subsection{DictFuns and default methods}
926 %************************************************************************
928 Important notes about dict funs and default methods
929 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
930 Dict funs and default methods are *not* ImplicitIds. Their definition
931 involves user-written code, so we can't figure out their strictness etc
932 based on fixed info, as we can for constructors and record selectors (say).
934 We build them as LocalIds, but with External Names. This ensures that
935 they are taken to account by free-variable finding and dependency
936 analysis (e.g. CoreFVs.exprFreeVars).
938 Why shouldn't they be bound as GlobalIds? Because, in particular, if
939 they are globals, the specialiser floats dict uses above their defns,
940 which prevents good simplifications happening. Also the strictness
941 analyser treats a occurrence of a GlobalId as imported and assumes it
942 contains strictness in its IdInfo, which isn't true if the thing is
943 bound in the same module as the occurrence.
945 It's OK for dfuns to be LocalIds, because we form the instance-env to
946 pass on to the next module (md_insts) in CoreTidy, afer tidying
947 and globalising the top-level Ids.
949 BUT make sure they are *exported* LocalIds (mkExportedLocalId) so
950 that they aren't discarded by the occurrence analyser.
953 mkDefaultMethodId dm_name ty = mkExportedLocalId dm_name ty
955 mkDictFunId :: Name -- Name to use for the dict fun;
962 mkDictFunId dfun_name inst_tyvars dfun_theta clas inst_tys
963 = mkExportedLocalId dfun_name dfun_ty
965 dfun_ty = mkSigmaTy inst_tyvars dfun_theta (mkDictTy clas inst_tys)
967 {- 1 dec 99: disable the Mark Jones optimisation for the sake
968 of compatibility with Hugs.
969 See `types/InstEnv' for a discussion related to this.
971 (class_tyvars, sc_theta, _, _) = classBigSig clas
972 not_const (clas, tys) = not (isEmptyVarSet (tyVarsOfTypes tys))
973 sc_theta' = substClasses (zipTopTvSubst class_tyvars inst_tys) sc_theta
974 dfun_theta = case inst_decl_theta of
975 [] -> [] -- If inst_decl_theta is empty, then we don't
976 -- want to have any dict arguments, so that we can
977 -- expose the constant methods.
979 other -> nub (inst_decl_theta ++ filter not_const sc_theta')
980 -- Otherwise we pass the superclass dictionaries to
981 -- the dictionary function; the Mark Jones optimisation.
983 -- NOTE the "nub". I got caught by this one:
984 -- class Monad m => MonadT t m where ...
985 -- instance Monad m => MonadT (EnvT env) m where ...
986 -- Here, the inst_decl_theta has (Monad m); but so
987 -- does the sc_theta'!
989 -- NOTE the "not_const". I got caught by this one too:
990 -- class Foo a => Baz a b where ...
991 -- instance Wob b => Baz T b where..
992 -- Now sc_theta' has Foo T
997 %************************************************************************
999 \subsection{Un-definable}
1001 %************************************************************************
1003 These Ids can't be defined in Haskell. They could be defined in
1004 unfoldings in the wired-in GHC.Prim interface file, but we'd have to
1005 ensure that they were definitely, definitely inlined, because there is
1006 no curried identifier for them. That's what mkCompulsoryUnfolding
1007 does. If we had a way to get a compulsory unfolding from an interface
1008 file, we could do that, but we don't right now.
1010 unsafeCoerce# isn't so much a PrimOp as a phantom identifier, that
1011 just gets expanded into a type coercion wherever it occurs. Hence we
1012 add it as a built-in Id with an unfolding here.
1014 The type variables we use here are "open" type variables: this means
1015 they can unify with both unlifted and lifted types. Hence we provide
1016 another gun with which to shoot yourself in the foot.
1019 mkWiredInIdName mod fs uniq id
1020 = mkWiredInName mod (mkOccNameFS varName fs) uniq (AnId id) UserSyntax
1022 unsafeCoerceName = mkWiredInIdName gHC_PRIM FSLIT("unsafeCoerce#") unsafeCoerceIdKey unsafeCoerceId
1023 nullAddrName = mkWiredInIdName gHC_PRIM FSLIT("nullAddr#") nullAddrIdKey nullAddrId
1024 seqName = mkWiredInIdName gHC_PRIM FSLIT("seq") seqIdKey seqId
1025 realWorldName = mkWiredInIdName gHC_PRIM FSLIT("realWorld#") realWorldPrimIdKey realWorldPrimId
1026 lazyIdName = mkWiredInIdName gHC_BASE FSLIT("lazy") lazyIdKey lazyId
1028 errorName = mkWiredInIdName gHC_ERR FSLIT("error") errorIdKey eRROR_ID
1029 recSelErrorName = mkWiredInIdName gHC_ERR FSLIT("recSelError") recSelErrorIdKey rEC_SEL_ERROR_ID
1030 runtimeErrorName = mkWiredInIdName gHC_ERR FSLIT("runtimeError") runtimeErrorIdKey rUNTIME_ERROR_ID
1031 irrefutPatErrorName = mkWiredInIdName gHC_ERR FSLIT("irrefutPatError") irrefutPatErrorIdKey iRREFUT_PAT_ERROR_ID
1032 recConErrorName = mkWiredInIdName gHC_ERR FSLIT("recConError") recConErrorIdKey rEC_CON_ERROR_ID
1033 patErrorName = mkWiredInIdName gHC_ERR FSLIT("patError") patErrorIdKey pAT_ERROR_ID
1034 noMethodBindingErrorName = mkWiredInIdName gHC_ERR FSLIT("noMethodBindingError")
1035 noMethodBindingErrorIdKey nO_METHOD_BINDING_ERROR_ID
1036 nonExhaustiveGuardsErrorName
1037 = mkWiredInIdName gHC_ERR FSLIT("nonExhaustiveGuardsError")
1038 nonExhaustiveGuardsErrorIdKey nON_EXHAUSTIVE_GUARDS_ERROR_ID
1042 -- unsafeCoerce# :: forall a b. a -> b
1044 = pcMiscPrelId unsafeCoerceName ty info
1046 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1049 ty = mkForAllTys [openAlphaTyVar,openBetaTyVar]
1050 (mkFunTy openAlphaTy openBetaTy)
1051 [x] = mkTemplateLocals [openAlphaTy]
1052 rhs = mkLams [openAlphaTyVar,openBetaTyVar,x] $
1053 Cast (Var x) (mkUnsafeCoercion openAlphaTy openBetaTy)
1055 -- nullAddr# :: Addr#
1056 -- The reason is is here is because we don't provide
1057 -- a way to write this literal in Haskell.
1059 = pcMiscPrelId nullAddrName addrPrimTy info
1061 info = noCafIdInfo `setUnfoldingInfo`
1062 mkCompulsoryUnfolding (Lit nullAddrLit)
1065 = pcMiscPrelId seqName ty info
1067 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1070 ty = mkForAllTys [alphaTyVar,openBetaTyVar]
1071 (mkFunTy alphaTy (mkFunTy openBetaTy openBetaTy))
1072 [x,y] = mkTemplateLocals [alphaTy, openBetaTy]
1073 rhs = mkLams [alphaTyVar,openBetaTyVar,x,y] (Case (Var x) x openBetaTy [(DEFAULT, [], Var y)])
1075 -- lazy :: forall a?. a? -> a? (i.e. works for unboxed types too)
1076 -- Used to lazify pseq: pseq a b = a `seq` lazy b
1078 -- Also, no strictness: by being a built-in Id, all the info about lazyId comes from here,
1079 -- not from GHC.Base.hi. This is important, because the strictness
1080 -- analyser will spot it as strict!
1082 -- Also no unfolding in lazyId: it gets "inlined" by a HACK in the worker/wrapper pass
1083 -- (see WorkWrap.wwExpr)
1084 -- We could use inline phases to do this, but that would be vulnerable to changes in
1085 -- phase numbering....we must inline precisely after strictness analysis.
1087 = pcMiscPrelId lazyIdName ty info
1090 ty = mkForAllTys [alphaTyVar] (mkFunTy alphaTy alphaTy)
1092 lazyIdUnfolding :: CoreExpr -- Used to expand 'lazyId' after strictness anal
1093 lazyIdUnfolding = mkLams [openAlphaTyVar,x] (Var x)
1095 [x] = mkTemplateLocals [openAlphaTy]
1098 @realWorld#@ used to be a magic literal, \tr{void#}. If things get
1099 nasty as-is, change it back to a literal (@Literal@).
1101 voidArgId is a Local Id used simply as an argument in functions
1102 where we just want an arg to avoid having a thunk of unlifted type.
1104 x = \ void :: State# RealWorld -> (# p, q #)
1106 This comes up in strictness analysis
1109 realWorldPrimId -- :: State# RealWorld
1110 = pcMiscPrelId realWorldName realWorldStatePrimTy
1111 (noCafIdInfo `setUnfoldingInfo` evaldUnfolding)
1112 -- The evaldUnfolding makes it look that realWorld# is evaluated
1113 -- which in turn makes Simplify.interestingArg return True,
1114 -- which in turn makes INLINE things applied to realWorld# likely
1117 voidArgId -- :: State# RealWorld
1118 = mkSysLocal FSLIT("void") voidArgIdKey realWorldStatePrimTy
1122 %************************************************************************
1124 \subsection[PrelVals-error-related]{@error@ and friends; @trace@}
1126 %************************************************************************
1128 GHC randomly injects these into the code.
1130 @patError@ is just a version of @error@ for pattern-matching
1131 failures. It knows various ``codes'' which expand to longer
1132 strings---this saves space!
1134 @absentErr@ is a thing we put in for ``absent'' arguments. They jolly
1135 well shouldn't be yanked on, but if one is, then you will get a
1136 friendly message from @absentErr@ (rather than a totally random
1139 @parError@ is a special version of @error@ which the compiler does
1140 not know to be a bottoming Id. It is used in the @_par_@ and @_seq_@
1141 templates, but we don't ever expect to generate code for it.
1145 :: Id -- Should be of type (forall a. Addr# -> a)
1146 -- where Addr# points to a UTF8 encoded string
1147 -> Type -- The type to instantiate 'a'
1148 -> String -- The string to print
1151 mkRuntimeErrorApp err_id res_ty err_msg
1152 = mkApps (Var err_id) [Type res_ty, err_string]
1154 err_string = Lit (mkStringLit err_msg)
1156 rEC_SEL_ERROR_ID = mkRuntimeErrorId recSelErrorName
1157 rUNTIME_ERROR_ID = mkRuntimeErrorId runtimeErrorName
1158 iRREFUT_PAT_ERROR_ID = mkRuntimeErrorId irrefutPatErrorName
1159 rEC_CON_ERROR_ID = mkRuntimeErrorId recConErrorName
1160 pAT_ERROR_ID = mkRuntimeErrorId patErrorName
1161 nO_METHOD_BINDING_ERROR_ID = mkRuntimeErrorId noMethodBindingErrorName
1162 nON_EXHAUSTIVE_GUARDS_ERROR_ID = mkRuntimeErrorId nonExhaustiveGuardsErrorName
1164 -- The runtime error Ids take a UTF8-encoded string as argument
1165 mkRuntimeErrorId name = pc_bottoming_Id name runtimeErrorTy
1166 runtimeErrorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTy addrPrimTy openAlphaTy)
1170 eRROR_ID = pc_bottoming_Id errorName errorTy
1173 errorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTys [mkListTy charTy] openAlphaTy)
1174 -- Notice the openAlphaTyVar. It says that "error" can be applied
1175 -- to unboxed as well as boxed types. This is OK because it never
1176 -- returns, so the return type is irrelevant.
1180 %************************************************************************
1182 \subsection{Utilities}
1184 %************************************************************************
1187 pcMiscPrelId :: Name -> Type -> IdInfo -> Id
1188 pcMiscPrelId name ty info
1189 = mkVanillaGlobal name ty info
1190 -- We lie and say the thing is imported; otherwise, we get into
1191 -- a mess with dependency analysis; e.g., core2stg may heave in
1192 -- random calls to GHCbase.unpackPS__. If GHCbase is the module
1193 -- being compiled, then it's just a matter of luck if the definition
1194 -- will be in "the right place" to be in scope.
1196 pc_bottoming_Id name ty
1197 = pcMiscPrelId name ty bottoming_info
1199 bottoming_info = vanillaIdInfo `setAllStrictnessInfo` Just strict_sig
1200 -- Do *not* mark them as NoCafRefs, because they can indeed have
1201 -- CAF refs. For example, pAT_ERROR_ID calls GHC.Err.untangle,
1202 -- which has some CAFs
1203 -- In due course we may arrange that these error-y things are
1204 -- regarded by the GC as permanently live, in which case we
1205 -- can give them NoCaf info. As it is, any function that calls
1206 -- any pc_bottoming_Id will itself have CafRefs, which bloats
1209 strict_sig = mkStrictSig (mkTopDmdType [evalDmd] BotRes)
1210 -- These "bottom" out, no matter what their arguments