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, newTyConCo )
67 import Class ( Class, classTyCon, classSelIds )
68 import Var ( Id, TyVar, Var, setIdType )
69 import VarSet ( isEmptyVarSet, subVarSet, varSetElems )
70 import Name ( mkFCallName, mkWiredInName, Name, BuiltInSyntax(..))
71 import OccName ( mkOccNameFS, varName )
72 import PrimOp ( PrimOp, primOpSig, primOpOcc, primOpTag )
73 import ForeignCall ( ForeignCall )
74 import DataCon ( DataCon, DataConIds(..), dataConTyCon,
76 dataConFieldLabels, dataConRepArity, dataConResTys,
77 dataConRepArgTys, dataConRepType, dataConFullSig,
78 dataConStrictMarks, dataConExStricts,
79 splitProductType, isVanillaDataCon, dataConFieldType,
82 import Id ( idType, mkGlobalId, mkVanillaGlobal, mkSysLocal,
83 mkTemplateLocals, mkTemplateLocalsNum, mkExportedLocalId,
84 mkTemplateLocal, idName
86 import IdInfo ( IdInfo, noCafIdInfo, setUnfoldingInfo,
87 setArityInfo, setSpecInfo, setCafInfo,
88 setAllStrictnessInfo, vanillaIdInfo,
89 GlobalIdDetails(..), CafInfo(..)
91 import NewDemand ( mkStrictSig, DmdResult(..),
92 mkTopDmdType, topDmd, evalDmd, lazyDmd, retCPR,
93 Demand(..), Demands(..) )
94 import DmdAnal ( dmdAnalTopRhs )
96 import Unique ( mkBuiltinUnique, mkPrimOpIdUnique )
97 import Maybe ( fromJust )
100 import Util ( dropList, isSingleton )
103 import ListSetOps ( assoc, minusList )
106 %************************************************************************
108 \subsection{Wired in Ids}
110 %************************************************************************
114 = [ -- These error-y things are wired in because we don't yet have
115 -- a way to express in an interface file that the result type variable
116 -- is 'open'; that is can be unified with an unboxed type
118 -- [The interface file format now carry such information, but there's
119 -- no way yet of expressing at the definition site for these
120 -- error-reporting functions that they have an 'open'
121 -- result type. -- sof 1/99]
123 eRROR_ID, -- This one isn't used anywhere else in the compiler
124 -- But we still need it in wiredInIds so that when GHC
125 -- compiles a program that mentions 'error' we don't
126 -- import its type from the interface file; we just get
127 -- the Id defined here. Which has an 'open-tyvar' type.
130 iRREFUT_PAT_ERROR_ID,
131 nON_EXHAUSTIVE_GUARDS_ERROR_ID,
132 nO_METHOD_BINDING_ERROR_ID,
139 -- These Ids are exported from GHC.Prim
141 = [ -- These can't be defined in Haskell, but they have
142 -- perfectly reasonable unfoldings in Core
150 %************************************************************************
152 \subsection{Data constructors}
154 %************************************************************************
156 The wrapper for a constructor is an ordinary top-level binding that evaluates
157 any strict args, unboxes any args that are going to be flattened, and calls
160 We're going to build a constructor that looks like:
162 data (Data a, C b) => T a b = T1 !a !Int b
165 \d1::Data a, d2::C b ->
166 \p q r -> case p of { p ->
168 Con T1 [a,b] [p,q,r]}}
172 * d2 is thrown away --- a context in a data decl is used to make sure
173 one *could* construct dictionaries at the site the constructor
174 is used, but the dictionary isn't actually used.
176 * We have to check that we can construct Data dictionaries for
177 the types a and Int. Once we've done that we can throw d1 away too.
179 * We use (case p of q -> ...) to evaluate p, rather than "seq" because
180 all that matters is that the arguments are evaluated. "seq" is
181 very careful to preserve evaluation order, which we don't need
184 You might think that we could simply give constructors some strictness
185 info, like PrimOps, and let CoreToStg do the let-to-case transformation.
186 But we don't do that because in the case of primops and functions strictness
187 is a *property* not a *requirement*. In the case of constructors we need to
188 do something active to evaluate the argument.
190 Making an explicit case expression allows the simplifier to eliminate
191 it in the (common) case where the constructor arg is already evaluated.
193 In the case of data instances, the wrapper also applies the coercion turning
194 the representation type into the family instance type to cast the result of
198 mkDataConIds :: Name -> Name -> DataCon -> DataConIds
199 mkDataConIds wrap_name wkr_name data_con
201 = DCIds Nothing nt_work_id -- Newtype, only has a worker
203 | any isMarkedStrict all_strict_marks -- Algebraic, needs wrapper
204 || not (null eq_spec)
205 || isFamInstTyCon tycon
206 = DCIds (Just alg_wrap_id) wrk_id
208 | otherwise -- Algebraic, no wrapper
209 = DCIds Nothing wrk_id
211 (univ_tvs, ex_tvs, eq_spec,
212 theta, orig_arg_tys) = dataConFullSig data_con
213 tycon = dataConTyCon data_con
215 ----------- Wrapper --------------
216 -- We used to include the stupid theta in the wrapper's args
217 -- but now we don't. Instead the type checker just injects these
218 -- extra constraints where necessary.
219 wrap_tvs = (univ_tvs `minusList` map fst eq_spec) ++ ex_tvs
220 subst = mkTopTvSubst eq_spec
221 dict_tys = mkPredTys theta
222 result_ty_args = map (substTyVar subst) univ_tvs
223 result_ty = case tyConFamInst_maybe tycon of
224 -- ordinary constructor
225 Nothing -> mkTyConApp tycon result_ty_args
226 -- family instance constructor
229 mkTyConApp familyTyCon (map (substTy subst) instTys)
230 wrap_ty = mkForAllTys wrap_tvs $ mkFunTys dict_tys $
231 mkFunTys orig_arg_tys $ result_ty
232 -- NB: watch out here if you allow user-written equality
233 -- constraints in data constructor signatures
235 ----------- Worker (algebraic data types only) --------------
236 -- The *worker* for the data constructor is the function that
237 -- takes the representation arguments and builds the constructor.
238 wrk_id = mkGlobalId (DataConWorkId data_con) wkr_name
239 (dataConRepType data_con) wkr_info
241 wkr_arity = dataConRepArity data_con
242 wkr_info = noCafIdInfo
243 `setArityInfo` wkr_arity
244 `setAllStrictnessInfo` Just wkr_sig
245 `setUnfoldingInfo` evaldUnfolding -- Record that it's evaluated,
248 wkr_sig = mkStrictSig (mkTopDmdType (replicate wkr_arity topDmd) cpr_info)
249 -- Notice that we do *not* say the worker is strict
250 -- even if the data constructor is declared strict
251 -- e.g. data T = MkT !(Int,Int)
252 -- Why? Because the *wrapper* is strict (and its unfolding has case
253 -- expresssions that do the evals) but the *worker* itself is not.
254 -- If we pretend it is strict then when we see
255 -- case x of y -> $wMkT y
256 -- the simplifier thinks that y is "sure to be evaluated" (because
257 -- $wMkT is strict) and drops the case. No, $wMkT is not strict.
259 -- When the simplifer sees a pattern
260 -- case e of MkT x -> ...
261 -- it uses the dataConRepStrictness of MkT to mark x as evaluated;
262 -- but that's fine... dataConRepStrictness comes from the data con
263 -- not from the worker Id.
265 cpr_info | isProductTyCon tycon &&
268 wkr_arity <= mAX_CPR_SIZE = retCPR
270 -- RetCPR is only true for products that are real data types;
271 -- that is, not unboxed tuples or [non-recursive] newtypes
273 ----------- Workers for newtypes --------------
274 nt_work_id = mkGlobalId (DataConWrapId data_con) wkr_name wrap_ty nt_work_info
275 nt_work_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
276 `setArityInfo` 1 -- Arity 1
277 `setUnfoldingInfo` newtype_unf
278 newtype_unf = ASSERT( isVanillaDataCon data_con &&
279 isSingleton orig_arg_tys )
280 -- No existentials on a newtype, but it can have a context
281 -- e.g. newtype Eq a => T a = MkT (...)
282 mkCompulsoryUnfolding $
283 mkLams wrap_tvs $ Lam id_arg1 $
284 wrapNewTypeBody tycon result_ty_args
287 id_arg1 = mkTemplateLocal 1 (head orig_arg_tys)
289 ----------- Wrappers for algebraic data types --------------
290 alg_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty alg_wrap_info
291 alg_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
292 `setArityInfo` alg_arity
293 -- It's important to specify the arity, so that partial
294 -- applications are treated as values
295 `setUnfoldingInfo` alg_unf
296 `setAllStrictnessInfo` Just wrap_sig
298 all_strict_marks = dataConExStricts data_con ++ dataConStrictMarks data_con
299 wrap_sig = mkStrictSig (mkTopDmdType arg_dmds cpr_info)
300 arg_dmds = map mk_dmd all_strict_marks
301 mk_dmd str | isMarkedStrict str = evalDmd
302 | otherwise = lazyDmd
303 -- The Cpr info can be important inside INLINE rhss, where the
304 -- wrapper constructor isn't inlined.
305 -- And the argument strictness can be important too; we
306 -- may not inline a contructor when it is partially applied.
308 -- data W = C !Int !Int !Int
309 -- ...(let w = C x in ...(w p q)...)...
310 -- we want to see that w is strict in its two arguments
312 alg_unf = mkTopUnfolding $ Note InlineMe $
314 mkLams dict_args $ mkLams id_args $
315 foldr mk_case con_app
316 (zip (dict_args ++ id_args) all_strict_marks)
319 con_app _ rep_ids = wrapFamInstBody tycon result_ty_args $
320 Var wrk_id `mkTyApps` result_ty_args
322 `mkTyApps` map snd eq_spec
323 `mkVarApps` reverse rep_ids
325 (dict_args,i2) = mkLocals 1 dict_tys
326 (id_args,i3) = mkLocals i2 orig_arg_tys
330 :: (Id, StrictnessMark) -- Arg, strictness
331 -> (Int -> [Id] -> CoreExpr) -- Body
332 -> Int -- Next rep arg id
333 -> [Id] -- Rep args so far, reversed
335 mk_case (arg,strict) body i rep_args
337 NotMarkedStrict -> body i (arg:rep_args)
339 | isUnLiftedType (idType arg) -> body i (arg:rep_args)
341 Case (Var arg) arg result_ty [(DEFAULT,[], body i (arg:rep_args))]
344 -> unboxProduct i (Var arg) (idType arg) the_body
346 the_body i con_args = body i (reverse con_args ++ rep_args)
348 mAX_CPR_SIZE :: Arity
350 -- We do not treat very big tuples as CPR-ish:
351 -- a) for a start we get into trouble because there aren't
352 -- "enough" unboxed tuple types (a tiresome restriction,
354 -- b) more importantly, big unboxed tuples get returned mainly
355 -- on the stack, and are often then allocated in the heap
356 -- by the caller. So doing CPR for them may in fact make
359 mkLocals i tys = (zipWith mkTemplateLocal [i..i+n-1] tys, i+n)
363 -- If the type constructor is a representation type of a data instance, wrap
364 -- the expression into a cast adjusting the expression type, which is an
365 -- instance of the representation type, to the corresponding instance of the
366 -- family instance type.
368 wrapFamInstBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
369 wrapFamInstBody tycon args result_expr
370 | Just (co_con, _) <- tyConFamInst_maybe tycon
371 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) result_expr
378 %************************************************************************
380 \subsection{Record selectors}
382 %************************************************************************
384 We're going to build a record selector unfolding that looks like this:
386 data T a b c = T1 { ..., op :: a, ...}
387 | T2 { ..., op :: a, ...}
390 sel = /\ a b c -> \ d -> case d of
395 Similarly for newtypes
397 newtype N a = MkN { unN :: a->a }
400 unN n = coerce (a->a) n
402 We need to take a little care if the field has a polymorphic type:
404 data R = R { f :: forall a. a->a }
408 f :: forall a. R -> a -> a
409 f = /\ a \ r = case r of
412 (not f :: R -> forall a. a->a, which gives the type inference mechanism
413 problems at call sites)
415 Similarly for (recursive) newtypes
417 newtype N = MkN { unN :: forall a. a->a }
419 unN :: forall b. N -> b -> b
420 unN = /\b -> \n:N -> (coerce (forall a. a->a) n)
423 Note [Naughty record selectors]
424 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
425 A "naughty" field is one for which we can't define a record
426 selector, because an existential type variable would escape. For example:
427 data T = forall a. MkT { x,y::a }
428 We obviously can't define
430 Nevertheless we *do* put a RecordSelId into the type environment
431 so that if the user tries to use 'x' as a selector we can bleat
432 helpfully, rather than saying unhelpfully that 'x' is not in scope.
433 Hence the sel_naughty flag, to identify record selectors that don't really exist.
435 In general, a field is naughty if its type mentions a type variable that
436 isn't in the result type of the constructor.
438 Note [GADT record selectors]
439 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
440 For GADTs, we require that all constructors with a common field 'f' have the same
441 result type (modulo alpha conversion). [Checked in TcTyClsDecls.checkValidTyCon]
444 T1 { f :: a } :: T [a]
445 T2 { f :: a, y :: b } :: T [a]
446 and now the selector takes that type as its argument:
447 f :: forall a. T [a] -> a
451 Note the forall'd tyvars of the selector are just the free tyvars
452 of the result type; there may be other tyvars in the constructor's
453 type (e.g. 'b' in T2).
457 -- Steps for handling "naughty" vs "non-naughty" selectors:
458 -- 1. Determine naughtiness by comparing field type vs result type
459 -- 2. Install naughty ones with selector_ty of type _|_ and fill in mzero for info
460 -- 3. If it's not naughty, do the normal plan.
462 mkRecordSelId :: TyCon -> FieldLabel -> Id
463 mkRecordSelId tycon field_label
464 -- Assumes that all fields with the same field label have the same type
465 | is_naughty = naughty_id
468 is_naughty = not (tyVarsOfType field_ty `subVarSet` res_tv_set)
469 sel_id_details = RecordSelId tycon field_label is_naughty
471 -- Escapist case here for naughty construcotrs
472 -- We give it no IdInfo, and a type of forall a.a (never looked at)
473 naughty_id = mkGlobalId sel_id_details field_label forall_a_a noCafIdInfo
474 forall_a_a = mkForAllTy alphaTyVar (mkTyVarTy alphaTyVar)
476 -- Normal case starts here
477 sel_id = mkGlobalId sel_id_details field_label selector_ty info
478 data_cons = tyConDataCons tycon
479 data_cons_w_field = filter has_field data_cons -- Can't be empty!
480 has_field con = field_label `elem` dataConFieldLabels con
482 con1 = head data_cons_w_field
483 res_tys = dataConResTys con1
484 res_tv_set = tyVarsOfTypes res_tys
485 res_tvs = varSetElems res_tv_set
486 data_ty = mkTyConApp tycon res_tys
487 field_ty = dataConFieldType con1 field_label
489 -- *Very* tiresomely, the selectors are (unnecessarily!) overloaded over
490 -- just the dictionaries in the types of the constructors that contain
491 -- the relevant field. [The Report says that pattern matching on a
492 -- constructor gives the same constraints as applying it.] Urgh.
494 -- However, not all data cons have all constraints (because of
495 -- BuildTyCl.mkDataConStupidTheta). So we need to find all the data cons
496 -- involved in the pattern match and take the union of their constraints.
497 stupid_dict_tys = mkPredTys (dataConsStupidTheta data_cons_w_field)
498 n_stupid_dicts = length stupid_dict_tys
500 (field_tyvars,pre_field_theta,field_tau) = tcSplitSigmaTy field_ty
502 field_theta = filter (not . isEqPred) pre_field_theta
503 field_dict_tys = mkPredTys field_theta
504 n_field_dict_tys = length field_dict_tys
505 -- If the field has a universally quantified type we have to
506 -- be a bit careful. Suppose we have
507 -- data R = R { op :: forall a. Foo a => a -> a }
508 -- Then we can't give op the type
509 -- op :: R -> forall a. Foo a => a -> a
510 -- because the typechecker doesn't understand foralls to the
511 -- right of an arrow. The "right" type to give it is
512 -- op :: forall a. Foo a => R -> a -> a
513 -- But then we must generate the right unfolding too:
514 -- op = /\a -> \dfoo -> \ r ->
517 -- Note that this is exactly the type we'd infer from a user defn
521 selector_ty = mkForAllTys res_tvs $ mkForAllTys field_tyvars $
522 mkFunTys stupid_dict_tys $ mkFunTys field_dict_tys $
523 mkFunTy data_ty field_tau
525 arity = 1 + n_stupid_dicts + n_field_dict_tys
527 (strict_sig, rhs_w_str) = dmdAnalTopRhs sel_rhs
528 -- Use the demand analyser to work out strictness.
529 -- With all this unpackery it's not easy!
532 `setCafInfo` caf_info
534 `setUnfoldingInfo` mkTopUnfolding rhs_w_str
535 `setAllStrictnessInfo` Just strict_sig
537 -- Allocate Ids. We do it a funny way round because field_dict_tys is
538 -- almost always empty. Also note that we use max_dict_tys
539 -- rather than n_dict_tys, because the latter gives an infinite loop:
540 -- n_dict tys depends on the_alts, which depens on arg_ids, which depends
541 -- on arity, which depends on n_dict tys. Sigh! Mega sigh!
542 stupid_dict_ids = mkTemplateLocalsNum 1 stupid_dict_tys
543 max_stupid_dicts = length (tyConStupidTheta tycon)
544 field_dict_base = max_stupid_dicts + 1
545 field_dict_ids = mkTemplateLocalsNum field_dict_base field_dict_tys
546 dict_id_base = field_dict_base + n_field_dict_tys
547 data_id = mkTemplateLocal dict_id_base data_ty
548 arg_base = dict_id_base + 1
550 the_alts :: [CoreAlt]
551 the_alts = map mk_alt data_cons_w_field -- Already sorted by data-con
552 no_default = length data_cons == length data_cons_w_field -- No default needed
554 default_alt | no_default = []
555 | otherwise = [(DEFAULT, [], error_expr)]
557 -- The default branch may have CAF refs, because it calls recSelError etc.
558 caf_info | no_default = NoCafRefs
559 | otherwise = MayHaveCafRefs
561 sel_rhs = mkLams res_tvs $ mkLams field_tyvars $
562 mkLams stupid_dict_ids $ mkLams field_dict_ids $
563 Lam data_id $ mk_result sel_body
565 -- NB: A newtype always has a vanilla DataCon; no existentials etc
566 -- res_tys will simply be the dataConUnivTyVars
567 sel_body | isNewTyCon tycon = unwrapNewTypeBody tycon res_tys (Var data_id)
568 | otherwise = Case (Var data_id) data_id field_ty (default_alt ++ the_alts)
570 mk_result poly_result = mkVarApps (mkVarApps poly_result field_tyvars) field_dict_ids
571 -- We pull the field lambdas to the top, so we need to
572 -- apply them in the body. For example:
573 -- data T = MkT { foo :: forall a. a->a }
575 -- foo :: forall a. T -> a -> a
576 -- foo = /\a. \t:T. case t of { MkT f -> f a }
579 = ASSERT2( res_ty `tcEqType` field_ty, ppr data_con $$ ppr res_ty $$ ppr field_ty )
580 mkReboxingAlt rebox_uniqs data_con (ex_tvs ++ co_tvs ++ arg_vs) rhs
582 -- get pattern binders with types appropriately instantiated
583 arg_uniqs = map mkBuiltinUnique [arg_base..]
584 (ex_tvs, co_tvs, arg_vs) = dataConOrigInstPat arg_uniqs data_con res_tys
586 rebox_base = arg_base + length ex_tvs + length co_tvs + length arg_vs
587 rebox_uniqs = map mkBuiltinUnique [rebox_base..]
589 -- data T :: *->* where T1 { fld :: Maybe b } -> T [b]
590 -- Hence T1 :: forall a b. (a=[b]) => b -> T a
591 -- fld :: forall b. T [b] -> Maybe b
592 -- fld = /\b.\(t:T[b]). case t of
593 -- T1 b' (c : [b]=[b']) (x:Maybe b')
594 -- -> x `cast` Maybe (sym (right c))
596 Succeeded refinement = gadtRefine emptyRefinement ex_tvs co_tvs
597 (co_fn, res_ty) = refineType refinement (idType the_arg_id)
598 -- Generate the refinement for b'=b,
599 -- and apply to (Maybe b'), to get (Maybe b)
602 ExprCoFn co -> Cast (Var the_arg_id) co
603 id_co -> ASSERT(isIdCoercion id_co) Var the_arg_id
605 field_vs = filter (not . isPredTy . idType) arg_vs
606 the_arg_id = assoc "mkRecordSelId:mk_alt" (field_lbls `zip` field_vs) field_label
607 field_lbls = dataConFieldLabels data_con
609 error_expr = mkRuntimeErrorApp rEC_SEL_ERROR_ID field_ty full_msg
610 full_msg = showSDoc (sep [text "No match in record selector", ppr sel_id])
612 -- unbox a product type...
613 -- we will recurse into newtypes, casting along the way, and unbox at the
614 -- first product data constructor we find. e.g.
616 -- data PairInt = PairInt Int Int
617 -- newtype S = MkS PairInt
620 -- If we have e = MkT (MkS (PairInt 0 1)) and some body expecting a list of
621 -- ids, we get (modulo int passing)
623 -- case (e `cast` CoT) `cast` CoS of
624 -- PairInt a b -> body [a,b]
626 -- The Ints passed around are just for creating fresh locals
627 unboxProduct :: Int -> CoreExpr -> Type -> (Int -> [Id] -> CoreExpr) -> CoreExpr
628 unboxProduct i arg arg_ty body
631 result = mkUnpackCase the_id arg con_args boxing_con rhs
632 (_tycon, _tycon_args, boxing_con, tys) = deepSplitProductType "unboxProduct" arg_ty
633 ([the_id], i') = mkLocals i [arg_ty]
634 (con_args, i'') = mkLocals i' tys
635 rhs = body i'' con_args
637 mkUnpackCase :: Id -> CoreExpr -> [Id] -> DataCon -> CoreExpr -> CoreExpr
638 -- (mkUnpackCase x e args Con body)
640 -- case (e `cast` ...) of bndr { Con args -> body }
642 -- the type of the bndr passed in is irrelevent
643 mkUnpackCase bndr arg unpk_args boxing_con body
644 = Case cast_arg (setIdType bndr bndr_ty) (exprType body) [(DataAlt boxing_con, unpk_args, body)]
646 (cast_arg, bndr_ty) = go (idType bndr) arg
648 | (tycon, tycon_args, _, _) <- splitProductType "mkUnpackCase" ty
649 , isNewTyCon tycon && not (isRecursiveTyCon tycon)
650 = go (newTyConInstRhs tycon tycon_args)
651 (unwrapNewTypeBody tycon tycon_args arg)
652 | otherwise = (arg, ty)
655 reboxProduct :: [Unique] -- uniques to create new local binders
656 -> Type -- type of product to box
657 -> ([Unique], -- remaining uniques
658 CoreExpr, -- boxed product
659 [Id]) -- Ids being boxed into product
662 (_tycon, _tycon_args, _pack_con, con_arg_tys) = deepSplitProductType "reboxProduct" ty
664 us' = dropList con_arg_tys us
666 arg_ids = zipWith (mkSysLocal FSLIT("rb")) us con_arg_tys
668 bind_rhs = mkProductBox arg_ids ty
671 (us', bind_rhs, arg_ids)
673 mkProductBox :: [Id] -> Type -> CoreExpr
674 mkProductBox arg_ids ty
677 (tycon, tycon_args, pack_con, _con_arg_tys) = splitProductType "mkProductBox" ty
680 | isNewTyCon tycon && not (isRecursiveTyCon tycon)
681 = wrap (mkProductBox arg_ids (newTyConInstRhs tycon tycon_args))
682 | otherwise = mkConApp pack_con (map Type tycon_args ++ map Var arg_ids)
684 wrap expr = wrapNewTypeBody tycon tycon_args expr
687 -- (mkReboxingAlt us con xs rhs) basically constructs the case
688 -- alternative (con, xs, rhs)
689 -- but it does the reboxing necessary to construct the *source*
690 -- arguments, xs, from the representation arguments ys.
692 -- data T = MkT !(Int,Int) Bool
694 -- mkReboxingAlt MkT [x,b] r
695 -- = (DataAlt MkT, [y::Int,z::Int,b], let x = (y,z) in r)
697 -- mkDataAlt should really be in DataCon, but it can't because
698 -- it manipulates CoreSyn.
701 :: [Unique] -- Uniques for the new Ids
703 -> [Var] -- Source-level args, including existential dicts
707 mkReboxingAlt us con args rhs
708 | not (any isMarkedUnboxed stricts)
709 = (DataAlt con, args, rhs)
713 (binds, args') = go args stricts us
715 (DataAlt con, args', mkLets binds rhs)
718 stricts = dataConExStricts con ++ dataConStrictMarks con
720 go [] _stricts _us = ([], [])
722 -- Type variable case
723 go (arg:args) stricts us
725 = let (binds, args') = go args stricts us
726 in (binds, arg:args')
728 -- Term variable case
729 go (arg:args) (str:stricts) us
730 | isMarkedUnboxed str
732 let (binds, unpacked_args') = go args stricts us'
733 (us', bind_rhs, unpacked_args) = reboxProduct us (idType arg)
735 (NonRec arg bind_rhs : binds, unpacked_args ++ unpacked_args')
737 = let (binds, args') = go args stricts us
738 in (binds, arg:args')
742 %************************************************************************
744 \subsection{Dictionary selectors}
746 %************************************************************************
748 Selecting a field for a dictionary. If there is just one field, then
749 there's nothing to do.
751 Dictionary selectors may get nested forall-types. Thus:
754 op :: forall b. Ord b => a -> b -> b
756 Then the top-level type for op is
758 op :: forall a. Foo a =>
762 This is unlike ordinary record selectors, which have all the for-alls
763 at the outside. When dealing with classes it's very convenient to
764 recover the original type signature from the class op selector.
767 mkDictSelId :: Name -> Class -> Id
768 mkDictSelId name clas
769 = mkGlobalId (ClassOpId clas) name sel_ty info
771 sel_ty = mkForAllTys tyvars (mkFunTy (idType dict_id) (idType the_arg_id))
772 -- We can't just say (exprType rhs), because that would give a type
774 -- for a single-op class (after all, the selector is the identity)
775 -- But it's type must expose the representation of the dictionary
776 -- to gat (say) C a -> (a -> a)
780 `setUnfoldingInfo` mkTopUnfolding rhs
781 `setAllStrictnessInfo` Just strict_sig
783 -- We no longer use 'must-inline' on record selectors. They'll
784 -- inline like crazy if they scrutinise a constructor
786 -- The strictness signature is of the form U(AAAVAAAA) -> T
787 -- where the V depends on which item we are selecting
788 -- It's worth giving one, so that absence info etc is generated
789 -- even if the selector isn't inlined
790 strict_sig = mkStrictSig (mkTopDmdType [arg_dmd] TopRes)
791 arg_dmd | isNewTyCon tycon = evalDmd
792 | otherwise = Eval (Prod [ if the_arg_id == id then evalDmd else Abs
795 tycon = classTyCon clas
796 [data_con] = tyConDataCons tycon
797 tyvars = dataConUnivTyVars data_con
798 arg_tys = ASSERT( isVanillaDataCon data_con ) dataConRepArgTys data_con
799 the_arg_id = assoc "MkId.mkDictSelId" (map idName (classSelIds clas) `zip` arg_ids) name
801 pred = mkClassPred clas (mkTyVarTys tyvars)
802 (dict_id:arg_ids) = mkTemplateLocals (mkPredTy pred : arg_tys)
804 rhs = mkLams tyvars (Lam dict_id rhs_body)
805 rhs_body | isNewTyCon tycon = unwrapNewTypeBody tycon (map mkTyVarTy tyvars) (Var dict_id)
806 | otherwise = Case (Var dict_id) dict_id (idType the_arg_id)
807 [(DataAlt data_con, arg_ids, Var the_arg_id)]
809 wrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
810 -- The wrapper for the data constructor for a newtype looks like this:
811 -- newtype T a = MkT (a,Int)
812 -- MkT :: forall a. (a,Int) -> T a
813 -- MkT = /\a. \(x:(a,Int)). x `cast` sym (CoT a)
814 -- where CoT is the coercion TyCon assoicated with the newtype
816 -- The call (wrapNewTypeBody T [a] e) returns the
817 -- body of the wrapper, namely
818 -- e `cast` (CoT [a])
820 -- If a coercion constructor is prodivided in the newtype, then we use
821 -- it, otherwise the wrap/unwrap are both no-ops
823 wrapNewTypeBody tycon args result_expr
824 | Just co_con <- newTyConCo tycon
825 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) result_expr
829 unwrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
830 unwrapNewTypeBody tycon args result_expr
831 | Just co_con <- newTyConCo tycon
832 = mkCoerce (mkTyConApp co_con args) result_expr
840 %************************************************************************
842 \subsection{Primitive operations
844 %************************************************************************
847 mkPrimOpId :: PrimOp -> Id
851 (tyvars,arg_tys,res_ty, arity, strict_sig) = primOpSig prim_op
852 ty = mkForAllTys tyvars (mkFunTys arg_tys res_ty)
853 name = mkWiredInName gHC_PRIM (primOpOcc prim_op)
854 (mkPrimOpIdUnique (primOpTag prim_op))
855 Nothing (AnId id) UserSyntax
856 id = mkGlobalId (PrimOpId prim_op) name ty info
859 `setSpecInfo` mkSpecInfo (primOpRules prim_op name)
861 `setAllStrictnessInfo` Just strict_sig
863 -- For each ccall we manufacture a separate CCallOpId, giving it
864 -- a fresh unique, a type that is correct for this particular ccall,
865 -- and a CCall structure that gives the correct details about calling
868 -- The *name* of this Id is a local name whose OccName gives the full
869 -- details of the ccall, type and all. This means that the interface
870 -- file reader can reconstruct a suitable Id
872 mkFCallId :: Unique -> ForeignCall -> Type -> Id
873 mkFCallId uniq fcall ty
874 = ASSERT( isEmptyVarSet (tyVarsOfType ty) )
875 -- A CCallOpId should have no free type variables;
876 -- when doing substitutions won't substitute over it
877 mkGlobalId (FCallId fcall) name ty info
879 occ_str = showSDoc (braces (ppr fcall <+> ppr ty))
880 -- The "occurrence name" of a ccall is the full info about the
881 -- ccall; it is encoded, but may have embedded spaces etc!
883 name = mkFCallName uniq occ_str
887 `setAllStrictnessInfo` Just strict_sig
889 (_, tau) = tcSplitForAllTys ty
890 (arg_tys, _) = tcSplitFunTys tau
891 arity = length arg_tys
892 strict_sig = mkStrictSig (mkTopDmdType (replicate arity evalDmd) TopRes)
896 %************************************************************************
898 \subsection{DictFuns and default methods}
900 %************************************************************************
902 Important notes about dict funs and default methods
903 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
904 Dict funs and default methods are *not* ImplicitIds. Their definition
905 involves user-written code, so we can't figure out their strictness etc
906 based on fixed info, as we can for constructors and record selectors (say).
908 We build them as LocalIds, but with External Names. This ensures that
909 they are taken to account by free-variable finding and dependency
910 analysis (e.g. CoreFVs.exprFreeVars).
912 Why shouldn't they be bound as GlobalIds? Because, in particular, if
913 they are globals, the specialiser floats dict uses above their defns,
914 which prevents good simplifications happening. Also the strictness
915 analyser treats a occurrence of a GlobalId as imported and assumes it
916 contains strictness in its IdInfo, which isn't true if the thing is
917 bound in the same module as the occurrence.
919 It's OK for dfuns to be LocalIds, because we form the instance-env to
920 pass on to the next module (md_insts) in CoreTidy, afer tidying
921 and globalising the top-level Ids.
923 BUT make sure they are *exported* LocalIds (mkExportedLocalId) so
924 that they aren't discarded by the occurrence analyser.
927 mkDefaultMethodId dm_name ty = mkExportedLocalId dm_name ty
929 mkDictFunId :: Name -- Name to use for the dict fun;
936 mkDictFunId dfun_name inst_tyvars dfun_theta clas inst_tys
937 = mkExportedLocalId dfun_name dfun_ty
939 dfun_ty = mkSigmaTy inst_tyvars dfun_theta (mkDictTy clas inst_tys)
941 {- 1 dec 99: disable the Mark Jones optimisation for the sake
942 of compatibility with Hugs.
943 See `types/InstEnv' for a discussion related to this.
945 (class_tyvars, sc_theta, _, _) = classBigSig clas
946 not_const (clas, tys) = not (isEmptyVarSet (tyVarsOfTypes tys))
947 sc_theta' = substClasses (zipTopTvSubst class_tyvars inst_tys) sc_theta
948 dfun_theta = case inst_decl_theta of
949 [] -> [] -- If inst_decl_theta is empty, then we don't
950 -- want to have any dict arguments, so that we can
951 -- expose the constant methods.
953 other -> nub (inst_decl_theta ++ filter not_const sc_theta')
954 -- Otherwise we pass the superclass dictionaries to
955 -- the dictionary function; the Mark Jones optimisation.
957 -- NOTE the "nub". I got caught by this one:
958 -- class Monad m => MonadT t m where ...
959 -- instance Monad m => MonadT (EnvT env) m where ...
960 -- Here, the inst_decl_theta has (Monad m); but so
961 -- does the sc_theta'!
963 -- NOTE the "not_const". I got caught by this one too:
964 -- class Foo a => Baz a b where ...
965 -- instance Wob b => Baz T b where..
966 -- Now sc_theta' has Foo T
971 %************************************************************************
973 \subsection{Un-definable}
975 %************************************************************************
977 These Ids can't be defined in Haskell. They could be defined in
978 unfoldings in the wired-in GHC.Prim interface file, but we'd have to
979 ensure that they were definitely, definitely inlined, because there is
980 no curried identifier for them. That's what mkCompulsoryUnfolding
981 does. If we had a way to get a compulsory unfolding from an interface
982 file, we could do that, but we don't right now.
984 unsafeCoerce# isn't so much a PrimOp as a phantom identifier, that
985 just gets expanded into a type coercion wherever it occurs. Hence we
986 add it as a built-in Id with an unfolding here.
988 The type variables we use here are "open" type variables: this means
989 they can unify with both unlifted and lifted types. Hence we provide
990 another gun with which to shoot yourself in the foot.
993 mkWiredInIdName mod fs uniq id
994 = mkWiredInName mod (mkOccNameFS varName fs) uniq Nothing (AnId id) UserSyntax
996 unsafeCoerceName = mkWiredInIdName gHC_PRIM FSLIT("unsafeCoerce#") unsafeCoerceIdKey unsafeCoerceId
997 nullAddrName = mkWiredInIdName gHC_PRIM FSLIT("nullAddr#") nullAddrIdKey nullAddrId
998 seqName = mkWiredInIdName gHC_PRIM FSLIT("seq") seqIdKey seqId
999 realWorldName = mkWiredInIdName gHC_PRIM FSLIT("realWorld#") realWorldPrimIdKey realWorldPrimId
1000 lazyIdName = mkWiredInIdName gHC_BASE FSLIT("lazy") lazyIdKey lazyId
1002 errorName = mkWiredInIdName gHC_ERR FSLIT("error") errorIdKey eRROR_ID
1003 recSelErrorName = mkWiredInIdName gHC_ERR FSLIT("recSelError") recSelErrorIdKey rEC_SEL_ERROR_ID
1004 runtimeErrorName = mkWiredInIdName gHC_ERR FSLIT("runtimeError") runtimeErrorIdKey rUNTIME_ERROR_ID
1005 irrefutPatErrorName = mkWiredInIdName gHC_ERR FSLIT("irrefutPatError") irrefutPatErrorIdKey iRREFUT_PAT_ERROR_ID
1006 recConErrorName = mkWiredInIdName gHC_ERR FSLIT("recConError") recConErrorIdKey rEC_CON_ERROR_ID
1007 patErrorName = mkWiredInIdName gHC_ERR FSLIT("patError") patErrorIdKey pAT_ERROR_ID
1008 noMethodBindingErrorName = mkWiredInIdName gHC_ERR FSLIT("noMethodBindingError")
1009 noMethodBindingErrorIdKey nO_METHOD_BINDING_ERROR_ID
1010 nonExhaustiveGuardsErrorName
1011 = mkWiredInIdName gHC_ERR FSLIT("nonExhaustiveGuardsError")
1012 nonExhaustiveGuardsErrorIdKey nON_EXHAUSTIVE_GUARDS_ERROR_ID
1016 -- unsafeCoerce# :: forall a b. a -> b
1018 = pcMiscPrelId unsafeCoerceName ty info
1020 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1023 ty = mkForAllTys [openAlphaTyVar,openBetaTyVar]
1024 (mkFunTy openAlphaTy openBetaTy)
1025 [x] = mkTemplateLocals [openAlphaTy]
1026 rhs = mkLams [openAlphaTyVar,openBetaTyVar,x] $
1027 -- Note (Coerce openBetaTy openAlphaTy) (Var x)
1028 Cast (Var x) (mkUnsafeCoercion openAlphaTy openBetaTy)
1030 -- nullAddr# :: Addr#
1031 -- The reason is is here is because we don't provide
1032 -- a way to write this literal in Haskell.
1034 = pcMiscPrelId nullAddrName addrPrimTy info
1036 info = noCafIdInfo `setUnfoldingInfo`
1037 mkCompulsoryUnfolding (Lit nullAddrLit)
1040 = pcMiscPrelId seqName ty info
1042 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1045 ty = mkForAllTys [alphaTyVar,openBetaTyVar]
1046 (mkFunTy alphaTy (mkFunTy openBetaTy openBetaTy))
1047 [x,y] = mkTemplateLocals [alphaTy, openBetaTy]
1048 rhs = mkLams [alphaTyVar,openBetaTyVar,x,y] (Case (Var x) x openBetaTy [(DEFAULT, [], Var y)])
1050 -- lazy :: forall a?. a? -> a? (i.e. works for unboxed types too)
1051 -- Used to lazify pseq: pseq a b = a `seq` lazy b
1053 -- Also, no strictness: by being a built-in Id, all the info about lazyId comes from here,
1054 -- not from GHC.Base.hi. This is important, because the strictness
1055 -- analyser will spot it as strict!
1057 -- Also no unfolding in lazyId: it gets "inlined" by a HACK in the worker/wrapper pass
1058 -- (see WorkWrap.wwExpr)
1059 -- We could use inline phases to do this, but that would be vulnerable to changes in
1060 -- phase numbering....we must inline precisely after strictness analysis.
1062 = pcMiscPrelId lazyIdName ty info
1065 ty = mkForAllTys [alphaTyVar] (mkFunTy alphaTy alphaTy)
1067 lazyIdUnfolding :: CoreExpr -- Used to expand 'lazyId' after strictness anal
1068 lazyIdUnfolding = mkLams [openAlphaTyVar,x] (Var x)
1070 [x] = mkTemplateLocals [openAlphaTy]
1073 @realWorld#@ used to be a magic literal, \tr{void#}. If things get
1074 nasty as-is, change it back to a literal (@Literal@).
1076 voidArgId is a Local Id used simply as an argument in functions
1077 where we just want an arg to avoid having a thunk of unlifted type.
1079 x = \ void :: State# RealWorld -> (# p, q #)
1081 This comes up in strictness analysis
1084 realWorldPrimId -- :: State# RealWorld
1085 = pcMiscPrelId realWorldName realWorldStatePrimTy
1086 (noCafIdInfo `setUnfoldingInfo` evaldUnfolding)
1087 -- The evaldUnfolding makes it look that realWorld# is evaluated
1088 -- which in turn makes Simplify.interestingArg return True,
1089 -- which in turn makes INLINE things applied to realWorld# likely
1092 voidArgId -- :: State# RealWorld
1093 = mkSysLocal FSLIT("void") voidArgIdKey realWorldStatePrimTy
1097 %************************************************************************
1099 \subsection[PrelVals-error-related]{@error@ and friends; @trace@}
1101 %************************************************************************
1103 GHC randomly injects these into the code.
1105 @patError@ is just a version of @error@ for pattern-matching
1106 failures. It knows various ``codes'' which expand to longer
1107 strings---this saves space!
1109 @absentErr@ is a thing we put in for ``absent'' arguments. They jolly
1110 well shouldn't be yanked on, but if one is, then you will get a
1111 friendly message from @absentErr@ (rather than a totally random
1114 @parError@ is a special version of @error@ which the compiler does
1115 not know to be a bottoming Id. It is used in the @_par_@ and @_seq_@
1116 templates, but we don't ever expect to generate code for it.
1120 :: Id -- Should be of type (forall a. Addr# -> a)
1121 -- where Addr# points to a UTF8 encoded string
1122 -> Type -- The type to instantiate 'a'
1123 -> String -- The string to print
1126 mkRuntimeErrorApp err_id res_ty err_msg
1127 = mkApps (Var err_id) [Type res_ty, err_string]
1129 err_string = Lit (mkStringLit err_msg)
1131 rEC_SEL_ERROR_ID = mkRuntimeErrorId recSelErrorName
1132 rUNTIME_ERROR_ID = mkRuntimeErrorId runtimeErrorName
1133 iRREFUT_PAT_ERROR_ID = mkRuntimeErrorId irrefutPatErrorName
1134 rEC_CON_ERROR_ID = mkRuntimeErrorId recConErrorName
1135 pAT_ERROR_ID = mkRuntimeErrorId patErrorName
1136 nO_METHOD_BINDING_ERROR_ID = mkRuntimeErrorId noMethodBindingErrorName
1137 nON_EXHAUSTIVE_GUARDS_ERROR_ID = mkRuntimeErrorId nonExhaustiveGuardsErrorName
1139 -- The runtime error Ids take a UTF8-encoded string as argument
1140 mkRuntimeErrorId name = pc_bottoming_Id name runtimeErrorTy
1141 runtimeErrorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTy addrPrimTy openAlphaTy)
1145 eRROR_ID = pc_bottoming_Id errorName errorTy
1148 errorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTys [mkListTy charTy] openAlphaTy)
1149 -- Notice the openAlphaTyVar. It says that "error" can be applied
1150 -- to unboxed as well as boxed types. This is OK because it never
1151 -- returns, so the return type is irrelevant.
1155 %************************************************************************
1157 \subsection{Utilities}
1159 %************************************************************************
1162 pcMiscPrelId :: Name -> Type -> IdInfo -> Id
1163 pcMiscPrelId name ty info
1164 = mkVanillaGlobal name ty info
1165 -- We lie and say the thing is imported; otherwise, we get into
1166 -- a mess with dependency analysis; e.g., core2stg may heave in
1167 -- random calls to GHCbase.unpackPS__. If GHCbase is the module
1168 -- being compiled, then it's just a matter of luck if the definition
1169 -- will be in "the right place" to be in scope.
1171 pc_bottoming_Id name ty
1172 = pcMiscPrelId name ty bottoming_info
1174 bottoming_info = vanillaIdInfo `setAllStrictnessInfo` Just strict_sig
1175 -- Do *not* mark them as NoCafRefs, because they can indeed have
1176 -- CAF refs. For example, pAT_ERROR_ID calls GHC.Err.untangle,
1177 -- which has some CAFs
1178 -- In due course we may arrange that these error-y things are
1179 -- regarded by the GC as permanently live, in which case we
1180 -- can give them NoCaf info. As it is, any function that calls
1181 -- any pc_bottoming_Id will itself have CafRefs, which bloats
1184 strict_sig = mkStrictSig (mkTopDmdType [evalDmd] BotRes)
1185 -- These "bottom" out, no matter what their arguments
1187 (openAlphaTyVar:openBetaTyVar:_) = openAlphaTyVars
1188 openAlphaTy = mkTyVarTy openAlphaTyVar
1189 openBetaTy = mkTyVarTy openBetaTyVar