2 % (c) The AQUA Project, Glasgow University, 1998
4 \section[StdIdInfo]{Standard unfoldings}
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,
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"
42 import BasicTypes ( Arity, StrictnessMark(..), isMarkedUnboxed, isMarkedStrict )
43 import Rules ( mkSpecInfo )
44 import TysPrim ( openAlphaTyVars, alphaTyVar, alphaTy,
45 realWorldStatePrimTy, addrPrimTy
47 import TysWiredIn ( charTy, mkListTy )
48 import PrelRules ( primOpRules )
49 import Type ( TyThing(..), mkForAllTy, tyVarsOfTypes,
50 newTyConInstRhs, mkTopTvSubst, substTyVar, substTy )
51 import TcGadt ( gadtRefine, refineType, emptyRefinement )
52 import HsBinds ( ExprCoFn(..), isIdCoercion )
53 import Coercion ( mkSymCoercion, mkUnsafeCoercion, isEqPred )
54 import TcType ( Type, ThetaType, mkDictTy, mkPredTys, mkPredTy,
55 mkTyConApp, mkTyVarTys, mkClassPred, isPredTy,
56 mkFunTys, mkFunTy, mkSigmaTy, tcSplitSigmaTy, tcEqType,
57 isUnLiftedType, mkForAllTys, mkTyVarTy, tyVarsOfType,
58 tcSplitFunTys, tcSplitForAllTys, dataConsStupidTheta
60 import CoreUtils ( exprType, dataConOrigInstPat, mkCoerce )
61 import CoreUnfold ( mkTopUnfolding, mkCompulsoryUnfolding )
62 import Literal ( nullAddrLit, mkStringLit )
63 import TyCon ( TyCon, isNewTyCon, tyConDataCons, FieldLabel,
64 tyConStupidTheta, isProductTyCon, isDataTyCon,
65 isRecursiveTyCon, isFamInstTyCon,
66 tyConFamInst_maybe, tyConFamilyCoercion_maybe,
68 import Class ( Class, classTyCon, classSelIds )
69 import Var ( Id, TyVar, Var, setIdType )
70 import VarSet ( isEmptyVarSet, subVarSet, varSetElems )
71 import Name ( mkFCallName, mkWiredInName, Name, BuiltInSyntax(..))
72 import OccName ( mkOccNameFS, varName )
73 import PrimOp ( PrimOp, primOpSig, primOpOcc, primOpTag )
74 import ForeignCall ( ForeignCall )
75 import DataCon ( DataCon, DataConIds(..), dataConTyCon,
77 dataConFieldLabels, dataConRepArity, dataConResTys,
78 dataConRepArgTys, dataConRepType, dataConFullSig,
79 dataConStrictMarks, dataConExStricts,
80 splitProductType, isVanillaDataCon, dataConFieldType,
83 import Id ( idType, mkGlobalId, mkVanillaGlobal, mkSysLocal,
84 mkTemplateLocals, mkTemplateLocalsNum, mkExportedLocalId,
85 mkTemplateLocal, idName
87 import IdInfo ( IdInfo, noCafIdInfo, setUnfoldingInfo,
88 setArityInfo, setSpecInfo, setCafInfo,
89 setAllStrictnessInfo, vanillaIdInfo,
90 GlobalIdDetails(..), CafInfo(..)
92 import NewDemand ( mkStrictSig, DmdResult(..),
93 mkTopDmdType, topDmd, evalDmd, lazyDmd, retCPR,
94 Demand(..), Demands(..) )
95 import DmdAnal ( dmdAnalTopRhs )
97 import Unique ( mkBuiltinUnique, mkPrimOpIdUnique )
98 import Maybe ( fromJust )
101 import Util ( dropList, isSingleton )
104 import ListSetOps ( assoc, minusList )
107 %************************************************************************
109 \subsection{Wired in Ids}
111 %************************************************************************
115 = [ -- These error-y things are wired in because we don't yet have
116 -- a way to express in an interface file that the result type variable
117 -- is 'open'; that is can be unified with an unboxed type
119 -- [The interface file format now carry such information, but there's
120 -- no way yet of expressing at the definition site for these
121 -- error-reporting functions that they have an 'open'
122 -- result type. -- sof 1/99]
124 eRROR_ID, -- This one isn't used anywhere else in the compiler
125 -- But we still need it in wiredInIds so that when GHC
126 -- compiles a program that mentions 'error' we don't
127 -- import its type from the interface file; we just get
128 -- the Id defined here. Which has an 'open-tyvar' type.
131 iRREFUT_PAT_ERROR_ID,
132 nON_EXHAUSTIVE_GUARDS_ERROR_ID,
133 nO_METHOD_BINDING_ERROR_ID,
140 -- These Ids are exported from GHC.Prim
142 = [ -- These can't be defined in Haskell, but they have
143 -- perfectly reasonable unfoldings in Core
151 %************************************************************************
153 \subsection{Data constructors}
155 %************************************************************************
157 The wrapper for a constructor is an ordinary top-level binding that evaluates
158 any strict args, unboxes any args that are going to be flattened, and calls
161 We're going to build a constructor that looks like:
163 data (Data a, C b) => T a b = T1 !a !Int b
166 \d1::Data a, d2::C b ->
167 \p q r -> case p of { p ->
169 Con T1 [a,b] [p,q,r]}}
173 * d2 is thrown away --- a context in a data decl is used to make sure
174 one *could* construct dictionaries at the site the constructor
175 is used, but the dictionary isn't actually used.
177 * We have to check that we can construct Data dictionaries for
178 the types a and Int. Once we've done that we can throw d1 away too.
180 * We use (case p of q -> ...) to evaluate p, rather than "seq" because
181 all that matters is that the arguments are evaluated. "seq" is
182 very careful to preserve evaluation order, which we don't need
185 You might think that we could simply give constructors some strictness
186 info, like PrimOps, and let CoreToStg do the let-to-case transformation.
187 But we don't do that because in the case of primops and functions strictness
188 is a *property* not a *requirement*. In the case of constructors we need to
189 do something active to evaluate the argument.
191 Making an explicit case expression allows the simplifier to eliminate
192 it in the (common) case where the constructor arg is already evaluated.
194 [Wrappers for data instance tycons]
195 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
196 In the case of data instances, the wrapper also applies the coercion turning
197 the representation type into the family instance type to cast the result of
198 the wrapper. For example, consider the declarations
200 data family Map k :: * -> *
201 data instance Map (a, b) v = MapPair (Map a (Pair b v))
203 The tycon to which the datacon MapPair belongs gets a unique internal name of
204 the form :R123Map, and we call it the representation tycon. In contrast, Map
205 is the family tycon (accessible via tyConFamInst_maybe). The wrapper and work
206 of MapPair get the types
208 $WMapPair :: forall a b v. Map a (Map a b v) -> Map (a, b) v
209 $wMapPair :: forall a b v. Map a (Map a b v) -> :R123Map a b v
211 which implies that the wrapper code will have to apply the coercion moving
212 between representation and family type. It is accessible via
213 tyConFamilyCoercion_maybe and has kind
215 Co123Map a b v :: {Map (a, b) v :=: :R123Map a b v}
217 This coercion is conditionally applied by wrapFamInstBody.
220 mkDataConIds :: Name -> Name -> DataCon -> DataConIds
221 mkDataConIds wrap_name wkr_name data_con
223 = DCIds Nothing nt_work_id -- Newtype, only has a worker
225 | any isMarkedStrict all_strict_marks -- Algebraic, needs wrapper
226 || not (null eq_spec)
227 || isFamInstTyCon tycon
228 = DCIds (Just alg_wrap_id) wrk_id
230 | otherwise -- Algebraic, no wrapper
231 = DCIds Nothing wrk_id
233 (univ_tvs, ex_tvs, eq_spec,
234 theta, orig_arg_tys) = dataConFullSig data_con
235 tycon = dataConTyCon data_con
237 ----------- Wrapper --------------
238 -- We used to include the stupid theta in the wrapper's args
239 -- but now we don't. Instead the type checker just injects these
240 -- extra constraints where necessary.
241 wrap_tvs = (univ_tvs `minusList` map fst eq_spec) ++ ex_tvs
242 subst = mkTopTvSubst eq_spec
243 dict_tys = mkPredTys theta
244 result_ty_args = map (substTyVar subst) univ_tvs
245 result_ty = case tyConFamInst_maybe tycon of
246 -- ordinary constructor
247 Nothing -> mkTyConApp tycon result_ty_args
248 -- family instance constructor
251 mkTyConApp familyTyCon (map (substTy subst) instTys)
252 wrap_ty = mkForAllTys wrap_tvs $ mkFunTys dict_tys $
253 mkFunTys orig_arg_tys $ result_ty
254 -- NB: watch out here if you allow user-written equality
255 -- constraints in data constructor signatures
257 ----------- Worker (algebraic data types only) --------------
258 -- The *worker* for the data constructor is the function that
259 -- takes the representation arguments and builds the constructor.
260 wrk_id = mkGlobalId (DataConWorkId data_con) wkr_name
261 (dataConRepType data_con) wkr_info
263 wkr_arity = dataConRepArity data_con
264 wkr_info = noCafIdInfo
265 `setArityInfo` wkr_arity
266 `setAllStrictnessInfo` Just wkr_sig
267 `setUnfoldingInfo` evaldUnfolding -- Record that it's evaluated,
270 wkr_sig = mkStrictSig (mkTopDmdType (replicate wkr_arity topDmd) cpr_info)
271 -- Notice that we do *not* say the worker is strict
272 -- even if the data constructor is declared strict
273 -- e.g. data T = MkT !(Int,Int)
274 -- Why? Because the *wrapper* is strict (and its unfolding has case
275 -- expresssions that do the evals) but the *worker* itself is not.
276 -- If we pretend it is strict then when we see
277 -- case x of y -> $wMkT y
278 -- the simplifier thinks that y is "sure to be evaluated" (because
279 -- $wMkT is strict) and drops the case. No, $wMkT is not strict.
281 -- When the simplifer sees a pattern
282 -- case e of MkT x -> ...
283 -- it uses the dataConRepStrictness of MkT to mark x as evaluated;
284 -- but that's fine... dataConRepStrictness comes from the data con
285 -- not from the worker Id.
287 cpr_info | isProductTyCon tycon &&
290 wkr_arity <= mAX_CPR_SIZE = retCPR
292 -- RetCPR is only true for products that are real data types;
293 -- that is, not unboxed tuples or [non-recursive] newtypes
295 ----------- Workers for newtypes --------------
296 nt_work_id = mkGlobalId (DataConWrapId data_con) wkr_name wrap_ty nt_work_info
297 nt_work_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
298 `setArityInfo` 1 -- Arity 1
299 `setUnfoldingInfo` newtype_unf
300 newtype_unf = ASSERT( isVanillaDataCon data_con &&
301 isSingleton orig_arg_tys )
302 -- No existentials on a newtype, but it can have a context
303 -- e.g. newtype Eq a => T a = MkT (...)
304 mkCompulsoryUnfolding $
305 mkLams wrap_tvs $ Lam id_arg1 $
306 wrapNewTypeBody tycon result_ty_args
309 id_arg1 = mkTemplateLocal 1 (head orig_arg_tys)
311 ----------- Wrappers for algebraic data types --------------
312 alg_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty alg_wrap_info
313 alg_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
314 `setArityInfo` alg_arity
315 -- It's important to specify the arity, so that partial
316 -- applications are treated as values
317 `setUnfoldingInfo` alg_unf
318 `setAllStrictnessInfo` Just wrap_sig
320 all_strict_marks = dataConExStricts data_con ++ dataConStrictMarks data_con
321 wrap_sig = mkStrictSig (mkTopDmdType arg_dmds cpr_info)
322 arg_dmds = map mk_dmd all_strict_marks
323 mk_dmd str | isMarkedStrict str = evalDmd
324 | otherwise = lazyDmd
325 -- The Cpr info can be important inside INLINE rhss, where the
326 -- wrapper constructor isn't inlined.
327 -- And the argument strictness can be important too; we
328 -- may not inline a contructor when it is partially applied.
330 -- data W = C !Int !Int !Int
331 -- ...(let w = C x in ...(w p q)...)...
332 -- we want to see that w is strict in its two arguments
334 alg_unf = mkTopUnfolding $ Note InlineMe $
336 mkLams dict_args $ mkLams id_args $
337 foldr mk_case con_app
338 (zip (dict_args ++ id_args) all_strict_marks)
341 con_app _ rep_ids = wrapFamInstBody tycon result_ty_args $
342 Var wrk_id `mkTyApps` result_ty_args
344 `mkTyApps` map snd eq_spec
345 `mkVarApps` reverse rep_ids
347 (dict_args,i2) = mkLocals 1 dict_tys
348 (id_args,i3) = mkLocals i2 orig_arg_tys
352 :: (Id, StrictnessMark) -- Arg, strictness
353 -> (Int -> [Id] -> CoreExpr) -- Body
354 -> Int -- Next rep arg id
355 -> [Id] -- Rep args so far, reversed
357 mk_case (arg,strict) body i rep_args
359 NotMarkedStrict -> body i (arg:rep_args)
361 | isUnLiftedType (idType arg) -> body i (arg:rep_args)
363 Case (Var arg) arg result_ty [(DEFAULT,[], body i (arg:rep_args))]
366 -> unboxProduct i (Var arg) (idType arg) the_body
368 the_body i con_args = body i (reverse con_args ++ rep_args)
370 mAX_CPR_SIZE :: Arity
372 -- We do not treat very big tuples as CPR-ish:
373 -- a) for a start we get into trouble because there aren't
374 -- "enough" unboxed tuple types (a tiresome restriction,
376 -- b) more importantly, big unboxed tuples get returned mainly
377 -- on the stack, and are often then allocated in the heap
378 -- by the caller. So doing CPR for them may in fact make
381 mkLocals i tys = (zipWith mkTemplateLocal [i..i+n-1] tys, i+n)
385 -- If the type constructor is a representation type of a data instance, wrap
386 -- the expression into a cast adjusting the expression type, which is an
387 -- instance of the representation type, to the corresponding instance of the
388 -- family instance type.
390 wrapFamInstBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
391 wrapFamInstBody tycon args result_expr
392 | Just co_con <- tyConFamilyCoercion_maybe tycon
393 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) result_expr
400 %************************************************************************
402 \subsection{Record selectors}
404 %************************************************************************
406 We're going to build a record selector unfolding that looks like this:
408 data T a b c = T1 { ..., op :: a, ...}
409 | T2 { ..., op :: a, ...}
412 sel = /\ a b c -> \ d -> case d of
417 Similarly for newtypes
419 newtype N a = MkN { unN :: a->a }
422 unN n = coerce (a->a) n
424 We need to take a little care if the field has a polymorphic type:
426 data R = R { f :: forall a. a->a }
430 f :: forall a. R -> a -> a
431 f = /\ a \ r = case r of
434 (not f :: R -> forall a. a->a, which gives the type inference mechanism
435 problems at call sites)
437 Similarly for (recursive) newtypes
439 newtype N = MkN { unN :: forall a. a->a }
441 unN :: forall b. N -> b -> b
442 unN = /\b -> \n:N -> (coerce (forall a. a->a) n)
445 Note [Naughty record selectors]
446 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
447 A "naughty" field is one for which we can't define a record
448 selector, because an existential type variable would escape. For example:
449 data T = forall a. MkT { x,y::a }
450 We obviously can't define
452 Nevertheless we *do* put a RecordSelId into the type environment
453 so that if the user tries to use 'x' as a selector we can bleat
454 helpfully, rather than saying unhelpfully that 'x' is not in scope.
455 Hence the sel_naughty flag, to identify record selectors that don't really exist.
457 In general, a field is naughty if its type mentions a type variable that
458 isn't in the result type of the constructor.
460 Note [GADT record selectors]
461 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
462 For GADTs, we require that all constructors with a common field 'f' have the same
463 result type (modulo alpha conversion). [Checked in TcTyClsDecls.checkValidTyCon]
466 T1 { f :: a } :: T [a]
467 T2 { f :: a, y :: b } :: T [a]
468 and now the selector takes that type as its argument:
469 f :: forall a. T [a] -> a
473 Note the forall'd tyvars of the selector are just the free tyvars
474 of the result type; there may be other tyvars in the constructor's
475 type (e.g. 'b' in T2).
479 -- Steps for handling "naughty" vs "non-naughty" selectors:
480 -- 1. Determine naughtiness by comparing field type vs result type
481 -- 2. Install naughty ones with selector_ty of type _|_ and fill in mzero for info
482 -- 3. If it's not naughty, do the normal plan.
484 mkRecordSelId :: TyCon -> FieldLabel -> Id
485 mkRecordSelId tycon field_label
486 -- Assumes that all fields with the same field label have the same type
487 | is_naughty = naughty_id
490 is_naughty = not (tyVarsOfType field_ty `subVarSet` res_tv_set)
491 sel_id_details = RecordSelId tycon field_label is_naughty
493 -- Escapist case here for naughty construcotrs
494 -- We give it no IdInfo, and a type of forall a.a (never looked at)
495 naughty_id = mkGlobalId sel_id_details field_label forall_a_a noCafIdInfo
496 forall_a_a = mkForAllTy alphaTyVar (mkTyVarTy alphaTyVar)
498 -- Normal case starts here
499 sel_id = mkGlobalId sel_id_details field_label selector_ty info
500 data_cons = tyConDataCons tycon
501 data_cons_w_field = filter has_field data_cons -- Can't be empty!
502 has_field con = field_label `elem` dataConFieldLabels con
504 con1 = head data_cons_w_field
505 res_tys = dataConResTys con1
506 res_tv_set = tyVarsOfTypes res_tys
507 res_tvs = varSetElems res_tv_set
508 data_ty = mkTyConApp tycon res_tys
509 field_ty = dataConFieldType con1 field_label
511 -- *Very* tiresomely, the selectors are (unnecessarily!) overloaded over
512 -- just the dictionaries in the types of the constructors that contain
513 -- the relevant field. [The Report says that pattern matching on a
514 -- constructor gives the same constraints as applying it.] Urgh.
516 -- However, not all data cons have all constraints (because of
517 -- BuildTyCl.mkDataConStupidTheta). So we need to find all the data cons
518 -- involved in the pattern match and take the union of their constraints.
519 stupid_dict_tys = mkPredTys (dataConsStupidTheta data_cons_w_field)
520 n_stupid_dicts = length stupid_dict_tys
522 (field_tyvars,pre_field_theta,field_tau) = tcSplitSigmaTy field_ty
524 field_theta = filter (not . isEqPred) pre_field_theta
525 field_dict_tys = mkPredTys field_theta
526 n_field_dict_tys = length field_dict_tys
527 -- If the field has a universally quantified type we have to
528 -- be a bit careful. Suppose we have
529 -- data R = R { op :: forall a. Foo a => a -> a }
530 -- Then we can't give op the type
531 -- op :: R -> forall a. Foo a => a -> a
532 -- because the typechecker doesn't understand foralls to the
533 -- right of an arrow. The "right" type to give it is
534 -- op :: forall a. Foo a => R -> a -> a
535 -- But then we must generate the right unfolding too:
536 -- op = /\a -> \dfoo -> \ r ->
539 -- Note that this is exactly the type we'd infer from a user defn
543 selector_ty = mkForAllTys res_tvs $ mkForAllTys field_tyvars $
544 mkFunTys stupid_dict_tys $ mkFunTys field_dict_tys $
545 mkFunTy data_ty field_tau
547 arity = 1 + n_stupid_dicts + n_field_dict_tys
549 (strict_sig, rhs_w_str) = dmdAnalTopRhs sel_rhs
550 -- Use the demand analyser to work out strictness.
551 -- With all this unpackery it's not easy!
554 `setCafInfo` caf_info
556 `setUnfoldingInfo` mkTopUnfolding rhs_w_str
557 `setAllStrictnessInfo` Just strict_sig
559 -- Allocate Ids. We do it a funny way round because field_dict_tys is
560 -- almost always empty. Also note that we use max_dict_tys
561 -- rather than n_dict_tys, because the latter gives an infinite loop:
562 -- n_dict tys depends on the_alts, which depens on arg_ids, which depends
563 -- on arity, which depends on n_dict tys. Sigh! Mega sigh!
564 stupid_dict_ids = mkTemplateLocalsNum 1 stupid_dict_tys
565 max_stupid_dicts = length (tyConStupidTheta tycon)
566 field_dict_base = max_stupid_dicts + 1
567 field_dict_ids = mkTemplateLocalsNum field_dict_base field_dict_tys
568 dict_id_base = field_dict_base + n_field_dict_tys
569 data_id = mkTemplateLocal dict_id_base data_ty
570 arg_base = dict_id_base + 1
572 the_alts :: [CoreAlt]
573 the_alts = map mk_alt data_cons_w_field -- Already sorted by data-con
574 no_default = length data_cons == length data_cons_w_field -- No default needed
576 default_alt | no_default = []
577 | otherwise = [(DEFAULT, [], error_expr)]
579 -- The default branch may have CAF refs, because it calls recSelError etc.
580 caf_info | no_default = NoCafRefs
581 | otherwise = MayHaveCafRefs
583 sel_rhs = mkLams res_tvs $ mkLams field_tyvars $
584 mkLams stupid_dict_ids $ mkLams field_dict_ids $
585 Lam data_id $ mk_result sel_body
587 -- NB: A newtype always has a vanilla DataCon; no existentials etc
588 -- res_tys will simply be the dataConUnivTyVars
589 sel_body | isNewTyCon tycon = unwrapNewTypeBody tycon res_tys (Var data_id)
590 | otherwise = Case (Var data_id) data_id field_ty (default_alt ++ the_alts)
592 mk_result poly_result = mkVarApps (mkVarApps poly_result field_tyvars) field_dict_ids
593 -- We pull the field lambdas to the top, so we need to
594 -- apply them in the body. For example:
595 -- data T = MkT { foo :: forall a. a->a }
597 -- foo :: forall a. T -> a -> a
598 -- foo = /\a. \t:T. case t of { MkT f -> f a }
601 = ASSERT2( res_ty `tcEqType` field_ty, ppr data_con $$ ppr res_ty $$ ppr field_ty )
602 mkReboxingAlt rebox_uniqs data_con (ex_tvs ++ co_tvs ++ arg_vs) rhs
604 -- get pattern binders with types appropriately instantiated
605 arg_uniqs = map mkBuiltinUnique [arg_base..]
606 (ex_tvs, co_tvs, arg_vs) = dataConOrigInstPat arg_uniqs data_con res_tys
608 rebox_base = arg_base + length ex_tvs + length co_tvs + length arg_vs
609 rebox_uniqs = map mkBuiltinUnique [rebox_base..]
611 -- data T :: *->* where T1 { fld :: Maybe b } -> T [b]
612 -- Hence T1 :: forall a b. (a=[b]) => b -> T a
613 -- fld :: forall b. T [b] -> Maybe b
614 -- fld = /\b.\(t:T[b]). case t of
615 -- T1 b' (c : [b]=[b']) (x:Maybe b')
616 -- -> x `cast` Maybe (sym (right c))
618 Succeeded refinement = gadtRefine emptyRefinement ex_tvs co_tvs
619 (co_fn, res_ty) = refineType refinement (idType the_arg_id)
620 -- Generate the refinement for b'=b,
621 -- and apply to (Maybe b'), to get (Maybe b)
624 ExprCoFn co -> Cast (Var the_arg_id) co
625 id_co -> ASSERT(isIdCoercion id_co) Var the_arg_id
627 field_vs = filter (not . isPredTy . idType) arg_vs
628 the_arg_id = assoc "mkRecordSelId:mk_alt" (field_lbls `zip` field_vs) field_label
629 field_lbls = dataConFieldLabels data_con
631 error_expr = mkRuntimeErrorApp rEC_SEL_ERROR_ID field_ty full_msg
632 full_msg = showSDoc (sep [text "No match in record selector", ppr sel_id])
634 -- unbox a product type...
635 -- we will recurse into newtypes, casting along the way, and unbox at the
636 -- first product data constructor we find. e.g.
638 -- data PairInt = PairInt Int Int
639 -- newtype S = MkS PairInt
642 -- If we have e = MkT (MkS (PairInt 0 1)) and some body expecting a list of
643 -- ids, we get (modulo int passing)
645 -- case (e `cast` CoT) `cast` CoS of
646 -- PairInt a b -> body [a,b]
648 -- The Ints passed around are just for creating fresh locals
649 unboxProduct :: Int -> CoreExpr -> Type -> (Int -> [Id] -> CoreExpr) -> CoreExpr
650 unboxProduct i arg arg_ty body
653 result = mkUnpackCase the_id arg con_args boxing_con rhs
654 (_tycon, _tycon_args, boxing_con, tys) = deepSplitProductType "unboxProduct" arg_ty
655 ([the_id], i') = mkLocals i [arg_ty]
656 (con_args, i'') = mkLocals i' tys
657 rhs = body i'' con_args
659 mkUnpackCase :: Id -> CoreExpr -> [Id] -> DataCon -> CoreExpr -> CoreExpr
660 -- (mkUnpackCase x e args Con body)
662 -- case (e `cast` ...) of bndr { Con args -> body }
664 -- the type of the bndr passed in is irrelevent
665 mkUnpackCase bndr arg unpk_args boxing_con body
666 = Case cast_arg (setIdType bndr bndr_ty) (exprType body) [(DataAlt boxing_con, unpk_args, body)]
668 (cast_arg, bndr_ty) = go (idType bndr) arg
670 | (tycon, tycon_args, _, _) <- splitProductType "mkUnpackCase" ty
671 , isNewTyCon tycon && not (isRecursiveTyCon tycon)
672 = go (newTyConInstRhs tycon tycon_args)
673 (unwrapNewTypeBody tycon tycon_args arg)
674 | otherwise = (arg, ty)
677 reboxProduct :: [Unique] -- uniques to create new local binders
678 -> Type -- type of product to box
679 -> ([Unique], -- remaining uniques
680 CoreExpr, -- boxed product
681 [Id]) -- Ids being boxed into product
684 (_tycon, _tycon_args, _pack_con, con_arg_tys) = deepSplitProductType "reboxProduct" ty
686 us' = dropList con_arg_tys us
688 arg_ids = zipWith (mkSysLocal FSLIT("rb")) us con_arg_tys
690 bind_rhs = mkProductBox arg_ids ty
693 (us', bind_rhs, arg_ids)
695 mkProductBox :: [Id] -> Type -> CoreExpr
696 mkProductBox arg_ids ty
699 (tycon, tycon_args, pack_con, _con_arg_tys) = splitProductType "mkProductBox" ty
702 | isNewTyCon tycon && not (isRecursiveTyCon tycon)
703 = wrap (mkProductBox arg_ids (newTyConInstRhs tycon tycon_args))
704 | otherwise = mkConApp pack_con (map Type tycon_args ++ map Var arg_ids)
706 wrap expr = wrapNewTypeBody tycon tycon_args expr
709 -- (mkReboxingAlt us con xs rhs) basically constructs the case
710 -- alternative (con, xs, rhs)
711 -- but it does the reboxing necessary to construct the *source*
712 -- arguments, xs, from the representation arguments ys.
714 -- data T = MkT !(Int,Int) Bool
716 -- mkReboxingAlt MkT [x,b] r
717 -- = (DataAlt MkT, [y::Int,z::Int,b], let x = (y,z) in r)
719 -- mkDataAlt should really be in DataCon, but it can't because
720 -- it manipulates CoreSyn.
723 :: [Unique] -- Uniques for the new Ids
725 -> [Var] -- Source-level args, including existential dicts
729 mkReboxingAlt us con args rhs
730 | not (any isMarkedUnboxed stricts)
731 = (DataAlt con, args, rhs)
735 (binds, args') = go args stricts us
737 (DataAlt con, args', mkLets binds rhs)
740 stricts = dataConExStricts con ++ dataConStrictMarks con
742 go [] _stricts _us = ([], [])
744 -- Type variable case
745 go (arg:args) stricts us
747 = let (binds, args') = go args stricts us
748 in (binds, arg:args')
750 -- Term variable case
751 go (arg:args) (str:stricts) us
752 | isMarkedUnboxed str
754 let (binds, unpacked_args') = go args stricts us'
755 (us', bind_rhs, unpacked_args) = reboxProduct us (idType arg)
757 (NonRec arg bind_rhs : binds, unpacked_args ++ unpacked_args')
759 = let (binds, args') = go args stricts us
760 in (binds, arg:args')
764 %************************************************************************
766 \subsection{Dictionary selectors}
768 %************************************************************************
770 Selecting a field for a dictionary. If there is just one field, then
771 there's nothing to do.
773 Dictionary selectors may get nested forall-types. Thus:
776 op :: forall b. Ord b => a -> b -> b
778 Then the top-level type for op is
780 op :: forall a. Foo a =>
784 This is unlike ordinary record selectors, which have all the for-alls
785 at the outside. When dealing with classes it's very convenient to
786 recover the original type signature from the class op selector.
789 mkDictSelId :: Name -> Class -> Id
790 mkDictSelId name clas
791 = mkGlobalId (ClassOpId clas) name sel_ty info
793 sel_ty = mkForAllTys tyvars (mkFunTy (idType dict_id) (idType the_arg_id))
794 -- We can't just say (exprType rhs), because that would give a type
796 -- for a single-op class (after all, the selector is the identity)
797 -- But it's type must expose the representation of the dictionary
798 -- to gat (say) C a -> (a -> a)
802 `setUnfoldingInfo` mkTopUnfolding rhs
803 `setAllStrictnessInfo` Just strict_sig
805 -- We no longer use 'must-inline' on record selectors. They'll
806 -- inline like crazy if they scrutinise a constructor
808 -- The strictness signature is of the form U(AAAVAAAA) -> T
809 -- where the V depends on which item we are selecting
810 -- It's worth giving one, so that absence info etc is generated
811 -- even if the selector isn't inlined
812 strict_sig = mkStrictSig (mkTopDmdType [arg_dmd] TopRes)
813 arg_dmd | isNewTyCon tycon = evalDmd
814 | otherwise = Eval (Prod [ if the_arg_id == id then evalDmd else Abs
817 tycon = classTyCon clas
818 [data_con] = tyConDataCons tycon
819 tyvars = dataConUnivTyVars data_con
820 arg_tys = ASSERT( isVanillaDataCon data_con ) dataConRepArgTys data_con
821 the_arg_id = assoc "MkId.mkDictSelId" (map idName (classSelIds clas) `zip` arg_ids) name
823 pred = mkClassPred clas (mkTyVarTys tyvars)
824 (dict_id:arg_ids) = mkTemplateLocals (mkPredTy pred : arg_tys)
826 rhs = mkLams tyvars (Lam dict_id rhs_body)
827 rhs_body | isNewTyCon tycon = unwrapNewTypeBody tycon (map mkTyVarTy tyvars) (Var dict_id)
828 | otherwise = Case (Var dict_id) dict_id (idType the_arg_id)
829 [(DataAlt data_con, arg_ids, Var the_arg_id)]
831 wrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
832 -- The wrapper for the data constructor for a newtype looks like this:
833 -- newtype T a = MkT (a,Int)
834 -- MkT :: forall a. (a,Int) -> T a
835 -- MkT = /\a. \(x:(a,Int)). x `cast` sym (CoT a)
836 -- where CoT is the coercion TyCon assoicated with the newtype
838 -- The call (wrapNewTypeBody T [a] e) returns the
839 -- body of the wrapper, namely
840 -- e `cast` (CoT [a])
842 -- If a coercion constructor is prodivided in the newtype, then we use
843 -- it, otherwise the wrap/unwrap are both no-ops
845 wrapNewTypeBody tycon args result_expr
846 | Just co_con <- newTyConCo tycon
847 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) result_expr
851 unwrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
852 unwrapNewTypeBody tycon args result_expr
853 | Just co_con <- newTyConCo tycon
854 = mkCoerce (mkTyConApp co_con args) result_expr
862 %************************************************************************
864 \subsection{Primitive operations
866 %************************************************************************
869 mkPrimOpId :: PrimOp -> Id
873 (tyvars,arg_tys,res_ty, arity, strict_sig) = primOpSig prim_op
874 ty = mkForAllTys tyvars (mkFunTys arg_tys res_ty)
875 name = mkWiredInName gHC_PRIM (primOpOcc prim_op)
876 (mkPrimOpIdUnique (primOpTag prim_op))
877 Nothing (AnId id) UserSyntax
878 id = mkGlobalId (PrimOpId prim_op) name ty info
881 `setSpecInfo` mkSpecInfo (primOpRules prim_op name)
883 `setAllStrictnessInfo` Just strict_sig
885 -- For each ccall we manufacture a separate CCallOpId, giving it
886 -- a fresh unique, a type that is correct for this particular ccall,
887 -- and a CCall structure that gives the correct details about calling
890 -- The *name* of this Id is a local name whose OccName gives the full
891 -- details of the ccall, type and all. This means that the interface
892 -- file reader can reconstruct a suitable Id
894 mkFCallId :: Unique -> ForeignCall -> Type -> Id
895 mkFCallId uniq fcall ty
896 = ASSERT( isEmptyVarSet (tyVarsOfType ty) )
897 -- A CCallOpId should have no free type variables;
898 -- when doing substitutions won't substitute over it
899 mkGlobalId (FCallId fcall) name ty info
901 occ_str = showSDoc (braces (ppr fcall <+> ppr ty))
902 -- The "occurrence name" of a ccall is the full info about the
903 -- ccall; it is encoded, but may have embedded spaces etc!
905 name = mkFCallName uniq occ_str
909 `setAllStrictnessInfo` Just strict_sig
911 (_, tau) = tcSplitForAllTys ty
912 (arg_tys, _) = tcSplitFunTys tau
913 arity = length arg_tys
914 strict_sig = mkStrictSig (mkTopDmdType (replicate arity evalDmd) TopRes)
918 %************************************************************************
920 \subsection{DictFuns and default methods}
922 %************************************************************************
924 Important notes about dict funs and default methods
925 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
926 Dict funs and default methods are *not* ImplicitIds. Their definition
927 involves user-written code, so we can't figure out their strictness etc
928 based on fixed info, as we can for constructors and record selectors (say).
930 We build them as LocalIds, but with External Names. This ensures that
931 they are taken to account by free-variable finding and dependency
932 analysis (e.g. CoreFVs.exprFreeVars).
934 Why shouldn't they be bound as GlobalIds? Because, in particular, if
935 they are globals, the specialiser floats dict uses above their defns,
936 which prevents good simplifications happening. Also the strictness
937 analyser treats a occurrence of a GlobalId as imported and assumes it
938 contains strictness in its IdInfo, which isn't true if the thing is
939 bound in the same module as the occurrence.
941 It's OK for dfuns to be LocalIds, because we form the instance-env to
942 pass on to the next module (md_insts) in CoreTidy, afer tidying
943 and globalising the top-level Ids.
945 BUT make sure they are *exported* LocalIds (mkExportedLocalId) so
946 that they aren't discarded by the occurrence analyser.
949 mkDefaultMethodId dm_name ty = mkExportedLocalId dm_name ty
951 mkDictFunId :: Name -- Name to use for the dict fun;
958 mkDictFunId dfun_name inst_tyvars dfun_theta clas inst_tys
959 = mkExportedLocalId dfun_name dfun_ty
961 dfun_ty = mkSigmaTy inst_tyvars dfun_theta (mkDictTy clas inst_tys)
963 {- 1 dec 99: disable the Mark Jones optimisation for the sake
964 of compatibility with Hugs.
965 See `types/InstEnv' for a discussion related to this.
967 (class_tyvars, sc_theta, _, _) = classBigSig clas
968 not_const (clas, tys) = not (isEmptyVarSet (tyVarsOfTypes tys))
969 sc_theta' = substClasses (zipTopTvSubst class_tyvars inst_tys) sc_theta
970 dfun_theta = case inst_decl_theta of
971 [] -> [] -- If inst_decl_theta is empty, then we don't
972 -- want to have any dict arguments, so that we can
973 -- expose the constant methods.
975 other -> nub (inst_decl_theta ++ filter not_const sc_theta')
976 -- Otherwise we pass the superclass dictionaries to
977 -- the dictionary function; the Mark Jones optimisation.
979 -- NOTE the "nub". I got caught by this one:
980 -- class Monad m => MonadT t m where ...
981 -- instance Monad m => MonadT (EnvT env) m where ...
982 -- Here, the inst_decl_theta has (Monad m); but so
983 -- does the sc_theta'!
985 -- NOTE the "not_const". I got caught by this one too:
986 -- class Foo a => Baz a b where ...
987 -- instance Wob b => Baz T b where..
988 -- Now sc_theta' has Foo T
993 %************************************************************************
995 \subsection{Un-definable}
997 %************************************************************************
999 These Ids can't be defined in Haskell. They could be defined in
1000 unfoldings in the wired-in GHC.Prim interface file, but we'd have to
1001 ensure that they were definitely, definitely inlined, because there is
1002 no curried identifier for them. That's what mkCompulsoryUnfolding
1003 does. If we had a way to get a compulsory unfolding from an interface
1004 file, we could do that, but we don't right now.
1006 unsafeCoerce# isn't so much a PrimOp as a phantom identifier, that
1007 just gets expanded into a type coercion wherever it occurs. Hence we
1008 add it as a built-in Id with an unfolding here.
1010 The type variables we use here are "open" type variables: this means
1011 they can unify with both unlifted and lifted types. Hence we provide
1012 another gun with which to shoot yourself in the foot.
1015 mkWiredInIdName mod fs uniq id
1016 = mkWiredInName mod (mkOccNameFS varName fs) uniq Nothing (AnId id) UserSyntax
1018 unsafeCoerceName = mkWiredInIdName gHC_PRIM FSLIT("unsafeCoerce#") unsafeCoerceIdKey unsafeCoerceId
1019 nullAddrName = mkWiredInIdName gHC_PRIM FSLIT("nullAddr#") nullAddrIdKey nullAddrId
1020 seqName = mkWiredInIdName gHC_PRIM FSLIT("seq") seqIdKey seqId
1021 realWorldName = mkWiredInIdName gHC_PRIM FSLIT("realWorld#") realWorldPrimIdKey realWorldPrimId
1022 lazyIdName = mkWiredInIdName gHC_BASE FSLIT("lazy") lazyIdKey lazyId
1024 errorName = mkWiredInIdName gHC_ERR FSLIT("error") errorIdKey eRROR_ID
1025 recSelErrorName = mkWiredInIdName gHC_ERR FSLIT("recSelError") recSelErrorIdKey rEC_SEL_ERROR_ID
1026 runtimeErrorName = mkWiredInIdName gHC_ERR FSLIT("runtimeError") runtimeErrorIdKey rUNTIME_ERROR_ID
1027 irrefutPatErrorName = mkWiredInIdName gHC_ERR FSLIT("irrefutPatError") irrefutPatErrorIdKey iRREFUT_PAT_ERROR_ID
1028 recConErrorName = mkWiredInIdName gHC_ERR FSLIT("recConError") recConErrorIdKey rEC_CON_ERROR_ID
1029 patErrorName = mkWiredInIdName gHC_ERR FSLIT("patError") patErrorIdKey pAT_ERROR_ID
1030 noMethodBindingErrorName = mkWiredInIdName gHC_ERR FSLIT("noMethodBindingError")
1031 noMethodBindingErrorIdKey nO_METHOD_BINDING_ERROR_ID
1032 nonExhaustiveGuardsErrorName
1033 = mkWiredInIdName gHC_ERR FSLIT("nonExhaustiveGuardsError")
1034 nonExhaustiveGuardsErrorIdKey nON_EXHAUSTIVE_GUARDS_ERROR_ID
1038 -- unsafeCoerce# :: forall a b. a -> b
1040 = pcMiscPrelId unsafeCoerceName ty info
1042 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1045 ty = mkForAllTys [openAlphaTyVar,openBetaTyVar]
1046 (mkFunTy openAlphaTy openBetaTy)
1047 [x] = mkTemplateLocals [openAlphaTy]
1048 rhs = mkLams [openAlphaTyVar,openBetaTyVar,x] $
1049 -- Note (Coerce openBetaTy openAlphaTy) (Var x)
1050 Cast (Var x) (mkUnsafeCoercion openAlphaTy openBetaTy)
1052 -- nullAddr# :: Addr#
1053 -- The reason is is here is because we don't provide
1054 -- a way to write this literal in Haskell.
1056 = pcMiscPrelId nullAddrName addrPrimTy info
1058 info = noCafIdInfo `setUnfoldingInfo`
1059 mkCompulsoryUnfolding (Lit nullAddrLit)
1062 = pcMiscPrelId seqName ty info
1064 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1067 ty = mkForAllTys [alphaTyVar,openBetaTyVar]
1068 (mkFunTy alphaTy (mkFunTy openBetaTy openBetaTy))
1069 [x,y] = mkTemplateLocals [alphaTy, openBetaTy]
1070 rhs = mkLams [alphaTyVar,openBetaTyVar,x,y] (Case (Var x) x openBetaTy [(DEFAULT, [], Var y)])
1072 -- lazy :: forall a?. a? -> a? (i.e. works for unboxed types too)
1073 -- Used to lazify pseq: pseq a b = a `seq` lazy b
1075 -- Also, no strictness: by being a built-in Id, all the info about lazyId comes from here,
1076 -- not from GHC.Base.hi. This is important, because the strictness
1077 -- analyser will spot it as strict!
1079 -- Also no unfolding in lazyId: it gets "inlined" by a HACK in the worker/wrapper pass
1080 -- (see WorkWrap.wwExpr)
1081 -- We could use inline phases to do this, but that would be vulnerable to changes in
1082 -- phase numbering....we must inline precisely after strictness analysis.
1084 = pcMiscPrelId lazyIdName ty info
1087 ty = mkForAllTys [alphaTyVar] (mkFunTy alphaTy alphaTy)
1089 lazyIdUnfolding :: CoreExpr -- Used to expand 'lazyId' after strictness anal
1090 lazyIdUnfolding = mkLams [openAlphaTyVar,x] (Var x)
1092 [x] = mkTemplateLocals [openAlphaTy]
1095 @realWorld#@ used to be a magic literal, \tr{void#}. If things get
1096 nasty as-is, change it back to a literal (@Literal@).
1098 voidArgId is a Local Id used simply as an argument in functions
1099 where we just want an arg to avoid having a thunk of unlifted type.
1101 x = \ void :: State# RealWorld -> (# p, q #)
1103 This comes up in strictness analysis
1106 realWorldPrimId -- :: State# RealWorld
1107 = pcMiscPrelId realWorldName realWorldStatePrimTy
1108 (noCafIdInfo `setUnfoldingInfo` evaldUnfolding)
1109 -- The evaldUnfolding makes it look that realWorld# is evaluated
1110 -- which in turn makes Simplify.interestingArg return True,
1111 -- which in turn makes INLINE things applied to realWorld# likely
1114 voidArgId -- :: State# RealWorld
1115 = mkSysLocal FSLIT("void") voidArgIdKey realWorldStatePrimTy
1119 %************************************************************************
1121 \subsection[PrelVals-error-related]{@error@ and friends; @trace@}
1123 %************************************************************************
1125 GHC randomly injects these into the code.
1127 @patError@ is just a version of @error@ for pattern-matching
1128 failures. It knows various ``codes'' which expand to longer
1129 strings---this saves space!
1131 @absentErr@ is a thing we put in for ``absent'' arguments. They jolly
1132 well shouldn't be yanked on, but if one is, then you will get a
1133 friendly message from @absentErr@ (rather than a totally random
1136 @parError@ is a special version of @error@ which the compiler does
1137 not know to be a bottoming Id. It is used in the @_par_@ and @_seq_@
1138 templates, but we don't ever expect to generate code for it.
1142 :: Id -- Should be of type (forall a. Addr# -> a)
1143 -- where Addr# points to a UTF8 encoded string
1144 -> Type -- The type to instantiate 'a'
1145 -> String -- The string to print
1148 mkRuntimeErrorApp err_id res_ty err_msg
1149 = mkApps (Var err_id) [Type res_ty, err_string]
1151 err_string = Lit (mkStringLit err_msg)
1153 rEC_SEL_ERROR_ID = mkRuntimeErrorId recSelErrorName
1154 rUNTIME_ERROR_ID = mkRuntimeErrorId runtimeErrorName
1155 iRREFUT_PAT_ERROR_ID = mkRuntimeErrorId irrefutPatErrorName
1156 rEC_CON_ERROR_ID = mkRuntimeErrorId recConErrorName
1157 pAT_ERROR_ID = mkRuntimeErrorId patErrorName
1158 nO_METHOD_BINDING_ERROR_ID = mkRuntimeErrorId noMethodBindingErrorName
1159 nON_EXHAUSTIVE_GUARDS_ERROR_ID = mkRuntimeErrorId nonExhaustiveGuardsErrorName
1161 -- The runtime error Ids take a UTF8-encoded string as argument
1162 mkRuntimeErrorId name = pc_bottoming_Id name runtimeErrorTy
1163 runtimeErrorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTy addrPrimTy openAlphaTy)
1167 eRROR_ID = pc_bottoming_Id errorName errorTy
1170 errorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTys [mkListTy charTy] openAlphaTy)
1171 -- Notice the openAlphaTyVar. It says that "error" can be applied
1172 -- to unboxed as well as boxed types. This is OK because it never
1173 -- returns, so the return type is irrelevant.
1177 %************************************************************************
1179 \subsection{Utilities}
1181 %************************************************************************
1184 pcMiscPrelId :: Name -> Type -> IdInfo -> Id
1185 pcMiscPrelId name ty info
1186 = mkVanillaGlobal name ty info
1187 -- We lie and say the thing is imported; otherwise, we get into
1188 -- a mess with dependency analysis; e.g., core2stg may heave in
1189 -- random calls to GHCbase.unpackPS__. If GHCbase is the module
1190 -- being compiled, then it's just a matter of luck if the definition
1191 -- will be in "the right place" to be in scope.
1193 pc_bottoming_Id name ty
1194 = pcMiscPrelId name ty bottoming_info
1196 bottoming_info = vanillaIdInfo `setAllStrictnessInfo` Just strict_sig
1197 -- Do *not* mark them as NoCafRefs, because they can indeed have
1198 -- CAF refs. For example, pAT_ERROR_ID calls GHC.Err.untangle,
1199 -- which has some CAFs
1200 -- In due course we may arrange that these error-y things are
1201 -- regarded by the GC as permanently live, in which case we
1202 -- can give them NoCaf info. As it is, any function that calls
1203 -- any pc_bottoming_Id will itself have CafRefs, which bloats
1206 strict_sig = mkStrictSig (mkTopDmdType [evalDmd] BotRes)
1207 -- These "bottom" out, no matter what their arguments
1209 (openAlphaTyVar:openBetaTyVar:_) = openAlphaTyVars
1210 openAlphaTy = mkTyVarTy openAlphaTyVar
1211 openBetaTy = mkTyVarTy openBetaTyVar