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, newTyConInstRhs, coreEqType,
51 mkTopTvSubst, substTyVar )
52 import TcGadt ( gadtRefine, refineType, emptyRefinement )
53 import HsBinds ( ExprCoFn(..), isIdCoercion )
54 import Coercion ( mkSymCoercion, mkUnsafeCoercion,
55 splitNewTypeRepCo_maybe, isEqPred )
56 import TcType ( Type, ThetaType, mkDictTy, mkPredTys, mkPredTy,
57 mkTyConApp, mkTyVarTys, mkClassPred,
58 mkFunTys, mkFunTy, mkSigmaTy, tcSplitSigmaTy,
59 isUnLiftedType, mkForAllTys, mkTyVarTy, tyVarsOfType,
60 tcSplitFunTys, tcSplitForAllTys, dataConsStupidTheta
62 import CoreUtils ( exprType, dataConInstPat )
63 import CoreUnfold ( mkTopUnfolding, mkCompulsoryUnfolding )
64 import Literal ( nullAddrLit, mkStringLit )
65 import TyCon ( TyCon, isNewTyCon, tyConDataCons, FieldLabel,
66 tyConStupidTheta, isProductTyCon, isDataTyCon, isRecursiveTyCon,
67 newTyConCo, tyConArity )
68 import Class ( Class, classTyCon, classSelIds )
69 import Var ( Id, TyVar, Var, setIdType, mkCoVar, mkWildCoVar )
70 import VarSet ( isEmptyVarSet, subVarSet, varSetElems )
71 import Name ( mkFCallName, mkWiredInName, Name, BuiltInSyntax(..),
73 import OccName ( mkOccNameFS, varName )
74 import PrimOp ( PrimOp, primOpSig, primOpOcc, primOpTag )
75 import ForeignCall ( ForeignCall )
76 import DataCon ( DataCon, DataConIds(..), dataConTyCon, dataConUnivTyVars,
77 dataConFieldLabels, dataConRepArity, dataConResTys,
78 dataConRepArgTys, dataConRepType, dataConFullSig,
79 dataConSig, dataConStrictMarks, dataConExStricts,
80 splitProductType, isVanillaDataCon, dataConFieldType,
81 dataConInstOrigArgTys, deepSplitProductType
83 import Id ( idType, mkGlobalId, mkVanillaGlobal, mkSysLocal,
84 mkTemplateLocals, mkTemplateLocalsNum, mkExportedLocalId,
85 mkTemplateLocal, idName, mkWildId
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 )
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.
195 mkDataConIds :: Name -> Name -> DataCon -> DataConIds
196 mkDataConIds wrap_name wkr_name data_con
200 | any isMarkedStrict all_strict_marks -- Algebraic, needs wrapper
201 || not (null eq_spec)
202 = AlgDC (Just alg_wrap_id) wrk_id
204 | otherwise -- Algebraic, no wrapper
205 = AlgDC Nothing wrk_id
207 (univ_tvs, ex_tvs, eq_spec, theta, orig_arg_tys) = dataConFullSig data_con
208 tycon = dataConTyCon data_con
210 ----------- Wrapper --------------
211 -- We used to include the stupid theta in the wrapper's args
212 -- but now we don't. Instead the type checker just injects these
213 -- extra constraints where necessary.
214 wrap_tvs = (univ_tvs `minusList` map fst eq_spec) ++ ex_tvs
215 subst = mkTopTvSubst eq_spec
216 dict_tys = mkPredTys theta
217 result_ty_args = map (substTyVar subst) univ_tvs
218 result_ty = mkTyConApp tycon result_ty_args
219 wrap_ty = mkForAllTys wrap_tvs $ mkFunTys dict_tys $
220 mkFunTys orig_arg_tys $ result_ty
221 -- NB: watch out here if you allow user-written equality
222 -- constraints in data constructor signatures
224 ----------- Worker (algebraic data types only) --------------
225 -- The *worker* for the data constructor is the function that
226 -- takes the representation arguments and builds the constructor.
227 wrk_id = mkGlobalId (DataConWorkId data_con) wkr_name
228 (dataConRepType data_con) wkr_info
230 wkr_arity = dataConRepArity data_con
231 wkr_info = noCafIdInfo
232 `setArityInfo` wkr_arity
233 `setAllStrictnessInfo` Just wkr_sig
234 `setUnfoldingInfo` evaldUnfolding -- Record that it's evaluated,
237 wkr_sig = mkStrictSig (mkTopDmdType (replicate wkr_arity topDmd) cpr_info)
238 -- Notice that we do *not* say the worker is strict
239 -- even if the data constructor is declared strict
240 -- e.g. data T = MkT !(Int,Int)
241 -- Why? Because the *wrapper* is strict (and its unfolding has case
242 -- expresssions that do the evals) but the *worker* itself is not.
243 -- If we pretend it is strict then when we see
244 -- case x of y -> $wMkT y
245 -- the simplifier thinks that y is "sure to be evaluated" (because
246 -- $wMkT is strict) and drops the case. No, $wMkT is not strict.
248 -- When the simplifer sees a pattern
249 -- case e of MkT x -> ...
250 -- it uses the dataConRepStrictness of MkT to mark x as evaluated;
251 -- but that's fine... dataConRepStrictness comes from the data con
252 -- not from the worker Id.
254 cpr_info | isProductTyCon tycon &&
257 wkr_arity <= mAX_CPR_SIZE = retCPR
259 -- RetCPR is only true for products that are real data types;
260 -- that is, not unboxed tuples or [non-recursive] newtypes
262 ----------- Wrappers for newtypes --------------
263 nt_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty nt_wrap_info
264 nt_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
265 `setArityInfo` 1 -- Arity 1
266 `setUnfoldingInfo` newtype_unf
267 newtype_unf = ASSERT( isVanillaDataCon data_con &&
268 isSingleton orig_arg_tys )
269 -- No existentials on a newtype, but it can have a context
270 -- e.g. newtype Eq a => T a = MkT (...)
271 mkCompulsoryUnfolding $
272 mkLams wrap_tvs $ Lam id_arg1 $
273 wrapNewTypeBody tycon result_ty_args
276 id_arg1 = mkTemplateLocal 1 (head orig_arg_tys)
278 ----------- Wrappers for algebraic data types --------------
279 alg_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty alg_wrap_info
280 alg_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
281 `setArityInfo` alg_arity
282 -- It's important to specify the arity, so that partial
283 -- applications are treated as values
284 `setUnfoldingInfo` alg_unf
285 `setAllStrictnessInfo` Just wrap_sig
287 all_strict_marks = dataConExStricts data_con ++ dataConStrictMarks data_con
288 wrap_sig = mkStrictSig (mkTopDmdType arg_dmds cpr_info)
289 arg_dmds = map mk_dmd all_strict_marks
290 mk_dmd str | isMarkedStrict str = evalDmd
291 | otherwise = lazyDmd
292 -- The Cpr info can be important inside INLINE rhss, where the
293 -- wrapper constructor isn't inlined.
294 -- And the argument strictness can be important too; we
295 -- may not inline a contructor when it is partially applied.
297 -- data W = C !Int !Int !Int
298 -- ...(let w = C x in ...(w p q)...)...
299 -- we want to see that w is strict in its two arguments
301 alg_unf = mkTopUnfolding $ Note InlineMe $
303 mkLams dict_args $ mkLams id_args $
304 foldr mk_case con_app
305 (zip (dict_args ++ id_args) all_strict_marks)
308 con_app i rep_ids = Var wrk_id `mkTyApps` result_ty_args
310 `mkTyApps` map snd eq_spec
311 `mkVarApps` reverse rep_ids
313 (dict_args,i2) = mkLocals 1 dict_tys
314 (id_args,i3) = mkLocals i2 orig_arg_tys
318 :: (Id, StrictnessMark) -- Arg, strictness
319 -> (Int -> [Id] -> CoreExpr) -- Body
320 -> Int -- Next rep arg id
321 -> [Id] -- Rep args so far, reversed
323 mk_case (arg,strict) body i rep_args
325 NotMarkedStrict -> body i (arg:rep_args)
327 | isUnLiftedType (idType arg) -> body i (arg:rep_args)
329 Case (Var arg) arg result_ty [(DEFAULT,[], body i (arg:rep_args))]
332 -> unboxProduct i (Var arg) (idType arg) the_body result_ty
334 the_body i con_args = body i (reverse con_args ++ rep_args)
336 mAX_CPR_SIZE :: Arity
338 -- We do not treat very big tuples as CPR-ish:
339 -- a) for a start we get into trouble because there aren't
340 -- "enough" unboxed tuple types (a tiresome restriction,
342 -- b) more importantly, big unboxed tuples get returned mainly
343 -- on the stack, and are often then allocated in the heap
344 -- by the caller. So doing CPR for them may in fact make
347 mkLocals i tys = (zipWith mkTemplateLocal [i..i+n-1] tys, i+n)
353 %************************************************************************
355 \subsection{Record selectors}
357 %************************************************************************
359 We're going to build a record selector unfolding that looks like this:
361 data T a b c = T1 { ..., op :: a, ...}
362 | T2 { ..., op :: a, ...}
365 sel = /\ a b c -> \ d -> case d of
370 Similarly for newtypes
372 newtype N a = MkN { unN :: a->a }
375 unN n = coerce (a->a) n
377 We need to take a little care if the field has a polymorphic type:
379 data R = R { f :: forall a. a->a }
383 f :: forall a. R -> a -> a
384 f = /\ a \ r = case r of
387 (not f :: R -> forall a. a->a, which gives the type inference mechanism
388 problems at call sites)
390 Similarly for (recursive) newtypes
392 newtype N = MkN { unN :: forall a. a->a }
394 unN :: forall b. N -> b -> b
395 unN = /\b -> \n:N -> (coerce (forall a. a->a) n)
398 Note [Naughty record selectors]
399 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
400 A "naughty" field is one for which we can't define a record
401 selector, because an existential type variable would escape. For example:
402 data T = forall a. MkT { x,y::a }
403 We obviously can't define
405 Nevertheless we *do* put a RecordSelId into the type environment
406 so that if the user tries to use 'x' as a selector we can bleat
407 helpfully, rather than saying unhelpfully that 'x' is not in scope.
408 Hence the sel_naughty flag, to identify record selectors that don't really exist.
410 In general, a field is naughty if its type mentions a type variable that
411 isn't in the result type of the constructor.
413 Note [GADT record selectors]
414 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
415 For GADTs, we require that all constructors with a common field 'f' have the same
416 result type (modulo alpha conversion). [Checked in TcTyClsDecls.checkValidTyCon]
419 T1 { f :: a } :: T [a]
420 T2 { f :: a, y :: b } :: T [a]
421 and now the selector takes that type as its argument:
422 f :: forall a. T [a] -> a
426 Note the forall'd tyvars of the selector are just the free tyvars
427 of the result type; there may be other tyvars in the constructor's
428 type (e.g. 'b' in T2).
432 -- Steps for handling "naughty" vs "non-naughty" selectors:
433 -- 1. Determine naughtiness by comparing field type vs result type
434 -- 2. Install naughty ones with selector_ty of type _|_ and fill in mzero for info
435 -- 3. If it's not naughty, do the normal plan.
437 mkRecordSelId :: TyCon -> FieldLabel -> Id
438 mkRecordSelId tycon field_label
439 -- Assumes that all fields with the same field label have the same type
440 | is_naughty = naughty_id
443 is_naughty = not (tyVarsOfType field_ty `subVarSet` res_tv_set)
444 sel_id_details = RecordSelId tycon field_label is_naughty
446 -- Escapist case here for naughty construcotrs
447 -- We give it no IdInfo, and a type of forall a.a (never looked at)
448 naughty_id = mkGlobalId sel_id_details field_label forall_a_a noCafIdInfo
449 forall_a_a = mkForAllTy alphaTyVar (mkTyVarTy alphaTyVar)
451 -- Normal case starts here
452 sel_id = mkGlobalId sel_id_details field_label selector_ty info
453 data_cons = tyConDataCons tycon
454 data_cons_w_field = filter has_field data_cons -- Can't be empty!
455 has_field con = field_label `elem` dataConFieldLabels con
457 con1 = head data_cons_w_field
458 res_tys = dataConResTys con1
459 res_tv_set = tyVarsOfTypes res_tys
460 res_tvs = varSetElems res_tv_set
461 data_ty = mkTyConApp tycon res_tys
462 field_ty = dataConFieldType con1 field_label
464 -- *Very* tiresomely, the selectors are (unnecessarily!) overloaded over
465 -- just the dictionaries in the types of the constructors that contain
466 -- the relevant field. [The Report says that pattern matching on a
467 -- constructor gives the same constraints as applying it.] Urgh.
469 -- However, not all data cons have all constraints (because of
470 -- BuildTyCl.mkDataConStupidTheta). So we need to find all the data cons
471 -- involved in the pattern match and take the union of their constraints.
472 stupid_dict_tys = mkPredTys (dataConsStupidTheta data_cons_w_field)
473 n_stupid_dicts = length stupid_dict_tys
475 (field_tyvars,pre_field_theta,field_tau) = tcSplitSigmaTy field_ty
477 field_theta = filter (not . isEqPred) pre_field_theta
478 field_dict_tys = mkPredTys field_theta
479 n_field_dict_tys = length field_dict_tys
480 -- If the field has a universally quantified type we have to
481 -- be a bit careful. Suppose we have
482 -- data R = R { op :: forall a. Foo a => a -> a }
483 -- Then we can't give op the type
484 -- op :: R -> forall a. Foo a => a -> a
485 -- because the typechecker doesn't understand foralls to the
486 -- right of an arrow. The "right" type to give it is
487 -- op :: forall a. Foo a => R -> a -> a
488 -- But then we must generate the right unfolding too:
489 -- op = /\a -> \dfoo -> \ r ->
492 -- Note that this is exactly the type we'd infer from a user defn
496 selector_ty = mkForAllTys res_tvs $ mkForAllTys field_tyvars $
497 mkFunTys stupid_dict_tys $ mkFunTys field_dict_tys $
498 mkFunTy data_ty field_tau
500 arity = 1 + n_stupid_dicts + n_field_dict_tys
502 (strict_sig, rhs_w_str) = dmdAnalTopRhs sel_rhs
503 -- Use the demand analyser to work out strictness.
504 -- With all this unpackery it's not easy!
507 `setCafInfo` caf_info
509 `setUnfoldingInfo` mkTopUnfolding rhs_w_str
510 `setAllStrictnessInfo` Just strict_sig
512 -- Allocate Ids. We do it a funny way round because field_dict_tys is
513 -- almost always empty. Also note that we use max_dict_tys
514 -- rather than n_dict_tys, because the latter gives an infinite loop:
515 -- n_dict tys depends on the_alts, which depens on arg_ids, which depends
516 -- on arity, which depends on n_dict tys. Sigh! Mega sigh!
517 stupid_dict_ids = mkTemplateLocalsNum 1 stupid_dict_tys
518 max_stupid_dicts = length (tyConStupidTheta tycon)
519 field_dict_base = max_stupid_dicts + 1
520 field_dict_ids = mkTemplateLocalsNum field_dict_base field_dict_tys
521 dict_id_base = field_dict_base + n_field_dict_tys
522 data_id = mkTemplateLocal dict_id_base data_ty
523 arg_base = dict_id_base + 1
525 the_alts :: [CoreAlt]
526 the_alts = map mk_alt data_cons_w_field -- Already sorted by data-con
527 no_default = length data_cons == length data_cons_w_field -- No default needed
529 default_alt | no_default = []
530 | otherwise = [(DEFAULT, [], error_expr)]
532 -- The default branch may have CAF refs, because it calls recSelError etc.
533 caf_info | no_default = NoCafRefs
534 | otherwise = MayHaveCafRefs
536 sel_rhs = mkLams res_tvs $ mkLams field_tyvars $
537 mkLams stupid_dict_ids $ mkLams field_dict_ids $
538 Lam data_id $ mk_result sel_body
540 -- NB: A newtype always has a vanilla DataCon; no existentials etc
541 -- res_tys will simply be the dataConUnivTyVars
542 sel_body | isNewTyCon tycon = unwrapNewTypeBody tycon res_tys (Var data_id)
543 | otherwise = Case (Var data_id) data_id field_ty (default_alt ++ the_alts)
545 mk_result poly_result = mkVarApps (mkVarApps poly_result field_tyvars) field_dict_ids
546 -- We pull the field lambdas to the top, so we need to
547 -- apply them in the body. For example:
548 -- data T = MkT { foo :: forall a. a->a }
550 -- foo :: forall a. T -> a -> a
551 -- foo = /\a. \t:T. case t of { MkT f -> f a }
554 = -- In the non-vanilla case, the pattern must bind type variables and
555 -- the context stuff; hence the arg_prefix binding below
556 mkReboxingAlt uniqs data_con (arg_prefix ++ arg_ids) rhs
558 (arg_prefix, arg_ids)
559 | isVanillaDataCon data_con -- Instantiate from commmon base
560 = ([], mkTemplateLocalsNum arg_base (dataConInstOrigArgTys data_con res_tys))
561 | otherwise -- The case pattern binds type variables, which are used
562 -- in the types of the arguments of the pattern
563 = (ex_tvs ++ co_tvs ++ dict_vs, field_vs)
565 (ex_tvs, co_tvs, arg_vs) = dataConInstPat uniqs' data_con res_tys
566 (dict_vs, field_vs) = splitAt (length dc_theta) arg_vs
568 (_, pre_dc_theta, dc_arg_tys) = dataConSig data_con
569 dc_theta = filter (not . isEqPred) pre_dc_theta
571 arg_base' = arg_base + length dc_theta
573 unpack_base = arg_base' + length dc_arg_tys
575 uniq_list = map mkBuiltinUnique [unpack_base..]
577 Succeeded refinement = gadtRefine emptyRefinement ex_tvs co_tvs
578 (co_fn, out_ty) = refineType refinement (idType the_arg_id)
580 rhs = ASSERT(out_ty `coreEqType` field_tau) perform_co co_fn (Var the_arg_id)
582 perform_co (ExprCoFn co) expr = Cast expr co
583 perform_co id_co expr = ASSERT(isIdCoercion id_co) expr
585 -- split the uniq_list into two
586 uniqs = takeHalf uniq_list
587 uniqs' = takeHalf (drop 1 uniq_list)
590 takeHalf (h:_:t) = h:(takeHalf t)
593 the_arg_id = assoc "mkRecordSelId:mk_alt" (field_lbls `zip` arg_ids) field_label
594 field_lbls = dataConFieldLabels data_con
596 error_expr = mkRuntimeErrorApp rEC_SEL_ERROR_ID field_tau full_msg
597 full_msg = showSDoc (sep [text "No match in record selector", ppr sel_id])
599 -- unbox a product type...
600 -- we will recurse into newtypes, casting along the way, and unbox at the
601 -- first product data constructor we find. e.g.
603 -- data PairInt = PairInt Int Int
604 -- newtype S = MkS PairInt
607 -- If we have e = MkT (MkS (PairInt 0 1)) and some body expecting a list of
608 -- ids, we get (modulo int passing)
610 -- case (e `cast` (sym CoT)) `cast` (sym CoS) of
611 -- PairInt a b -> body [a,b]
613 -- The Ints passed around are just for creating fresh locals
614 unboxProduct :: Int -> CoreExpr -> Type -> (Int -> [Id] -> CoreExpr) -> Type -> CoreExpr
615 unboxProduct i arg arg_ty body res_ty
618 result = mkUnpackCase the_id arg arg_ty con_args boxing_con rhs
619 (tycon, tycon_args, boxing_con, tys) = deepSplitProductType "unboxProduct" arg_ty
620 ([the_id], i') = mkLocals i [arg_ty]
621 (con_args, i'') = mkLocals i' tys
622 rhs = body i'' con_args
624 mkUnpackCase :: Id -> CoreExpr -> Type -> [Id] -> DataCon -> CoreExpr -> CoreExpr
625 -- (mkUnpackCase x e args Con body)
627 -- case (e `cast` ...) of bndr { Con args -> body }
629 -- the type of the bndr passed in is irrelevent
630 mkUnpackCase bndr arg arg_ty unpk_args boxing_con body
631 = Case cast_arg (setIdType bndr bndr_ty) (exprType body) [(DataAlt boxing_con, unpk_args, body)]
633 (cast_arg, bndr_ty) = go (idType bndr) arg
635 | res@(tycon, tycon_args, _, _) <- splitProductType "mkUnpackCase" ty
636 , isNewTyCon tycon && not (isRecursiveTyCon tycon)
637 = go (newTyConInstRhs tycon tycon_args)
638 (unwrapNewTypeBody tycon tycon_args arg)
639 | otherwise = (arg, ty)
642 reboxProduct :: [Unique] -- uniques to create new local binders
643 -> Type -- type of product to box
644 -> ([Unique], -- remaining uniques
645 CoreExpr, -- boxed product
646 [Id]) -- Ids being boxed into product
649 (tycon, tycon_args, pack_con, con_arg_tys) = deepSplitProductType "reboxProduct" ty
651 us' = dropList con_arg_tys us
653 arg_ids = zipWith (mkSysLocal FSLIT("rb")) us con_arg_tys
655 bind_rhs = mkProductBox arg_ids ty
658 (us', bind_rhs, arg_ids)
660 mkProductBox :: [Id] -> Type -> CoreExpr
661 mkProductBox arg_ids ty
664 (tycon, tycon_args, pack_con, con_arg_tys) = splitProductType "mkProductBox" ty
667 | isNewTyCon tycon && not (isRecursiveTyCon tycon)
668 = wrap (mkProductBox arg_ids (newTyConInstRhs tycon tycon_args))
669 | otherwise = mkConApp pack_con (map Type tycon_args ++ map Var arg_ids)
671 wrap expr = wrapNewTypeBody tycon tycon_args expr
674 -- (mkReboxingAlt us con xs rhs) basically constructs the case
675 -- alternative (con, xs, rhs)
676 -- but it does the reboxing necessary to construct the *source*
677 -- arguments, xs, from the representation arguments ys.
679 -- data T = MkT !(Int,Int) Bool
681 -- mkReboxingAlt MkT [x,b] r
682 -- = (DataAlt MkT, [y::Int,z::Int,b], let x = (y,z) in r)
684 -- mkDataAlt should really be in DataCon, but it can't because
685 -- it manipulates CoreSyn.
688 :: [Unique] -- Uniques for the new Ids
690 -> [Var] -- Source-level args, including existential dicts
694 mkReboxingAlt us con args rhs
695 | not (any isMarkedUnboxed stricts)
696 = (DataAlt con, args, rhs)
700 (binds, args') = go args stricts us
702 (DataAlt con, args', mkLets binds rhs)
705 stricts = dataConExStricts con ++ dataConStrictMarks con
707 go [] stricts us = ([], [])
709 -- Type variable case
710 go (arg:args) stricts us
712 = let (binds, args') = go args stricts us
713 in (binds, arg:args')
715 -- Term variable case
716 go (arg:args) (str:stricts) us
717 | isMarkedUnboxed str
719 let (binds, unpacked_args') = go args stricts us'
720 (us', bind_rhs, unpacked_args) = reboxProduct us (idType arg)
722 (NonRec arg bind_rhs : binds, unpacked_args ++ unpacked_args')
724 = let (binds, args') = go args stricts us
725 in (binds, arg:args')
729 %************************************************************************
731 \subsection{Dictionary selectors}
733 %************************************************************************
735 Selecting a field for a dictionary. If there is just one field, then
736 there's nothing to do.
738 Dictionary selectors may get nested forall-types. Thus:
741 op :: forall b. Ord b => a -> b -> b
743 Then the top-level type for op is
745 op :: forall a. Foo a =>
749 This is unlike ordinary record selectors, which have all the for-alls
750 at the outside. When dealing with classes it's very convenient to
751 recover the original type signature from the class op selector.
754 mkDictSelId :: Name -> Class -> Id
755 mkDictSelId name clas
756 = mkGlobalId (ClassOpId clas) name sel_ty info
758 sel_ty = mkForAllTys tyvars (mkFunTy (idType dict_id) (idType the_arg_id))
759 -- We can't just say (exprType rhs), because that would give a type
761 -- for a single-op class (after all, the selector is the identity)
762 -- But it's type must expose the representation of the dictionary
763 -- to gat (say) C a -> (a -> a)
767 `setUnfoldingInfo` mkTopUnfolding rhs
768 `setAllStrictnessInfo` Just strict_sig
770 -- We no longer use 'must-inline' on record selectors. They'll
771 -- inline like crazy if they scrutinise a constructor
773 -- The strictness signature is of the form U(AAAVAAAA) -> T
774 -- where the V depends on which item we are selecting
775 -- It's worth giving one, so that absence info etc is generated
776 -- even if the selector isn't inlined
777 strict_sig = mkStrictSig (mkTopDmdType [arg_dmd] TopRes)
778 arg_dmd | isNewTyCon tycon = evalDmd
779 | otherwise = Eval (Prod [ if the_arg_id == id then evalDmd else Abs
782 tycon = classTyCon clas
783 [data_con] = tyConDataCons tycon
784 tyvars = dataConUnivTyVars data_con
785 arg_tys = ASSERT( isVanillaDataCon data_con ) dataConRepArgTys data_con
786 the_arg_id = assoc "MkId.mkDictSelId" (map idName (classSelIds clas) `zip` arg_ids) name
788 pred = mkClassPred clas (mkTyVarTys tyvars)
789 (dict_id:arg_ids) = mkTemplateLocals (mkPredTy pred : arg_tys)
791 rhs = mkLams tyvars (Lam dict_id rhs_body)
792 rhs_body | isNewTyCon tycon = unwrapNewTypeBody tycon (map mkTyVarTy tyvars) (Var dict_id)
793 | otherwise = Case (Var dict_id) dict_id (idType the_arg_id)
794 [(DataAlt data_con, arg_ids, Var the_arg_id)]
796 wrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
797 -- The wrapper for the data constructor for a newtype looks like this:
798 -- newtype T a = MkT (a,Int)
799 -- MkT :: forall a. (a,Int) -> T a
800 -- MkT = /\a. \(x:(a,Int)). x `cast` CoT a
801 -- where CoT is the coercion TyCon assoicated with the newtype
803 -- The call (wrapNewTypeBody T [a] e) returns the
804 -- body of the wrapper, namely
807 -- If a coercion constructor is prodivided in the newtype, then we use
808 -- it, otherwise the wrap/unwrap are both no-ops
810 wrapNewTypeBody tycon args result_expr
811 | Just co_con <- newTyConCo tycon
812 = Cast result_expr (mkTyConApp co_con args)
816 unwrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
817 unwrapNewTypeBody tycon args result_expr
818 | Just co_con <- newTyConCo tycon
819 = Cast result_expr (mkSymCoercion (mkTyConApp co_con args))
827 %************************************************************************
829 \subsection{Primitive operations
831 %************************************************************************
834 mkPrimOpId :: PrimOp -> Id
838 (tyvars,arg_tys,res_ty, arity, strict_sig) = primOpSig prim_op
839 ty = mkForAllTys tyvars (mkFunTys arg_tys res_ty)
840 name = mkWiredInName gHC_PRIM (primOpOcc prim_op)
841 (mkPrimOpIdUnique (primOpTag prim_op))
842 Nothing (AnId id) UserSyntax
843 id = mkGlobalId (PrimOpId prim_op) name ty info
846 `setSpecInfo` mkSpecInfo (primOpRules prim_op name)
848 `setAllStrictnessInfo` Just strict_sig
850 -- For each ccall we manufacture a separate CCallOpId, giving it
851 -- a fresh unique, a type that is correct for this particular ccall,
852 -- and a CCall structure that gives the correct details about calling
855 -- The *name* of this Id is a local name whose OccName gives the full
856 -- details of the ccall, type and all. This means that the interface
857 -- file reader can reconstruct a suitable Id
859 mkFCallId :: Unique -> ForeignCall -> Type -> Id
860 mkFCallId uniq fcall ty
861 = ASSERT( isEmptyVarSet (tyVarsOfType ty) )
862 -- A CCallOpId should have no free type variables;
863 -- when doing substitutions won't substitute over it
864 mkGlobalId (FCallId fcall) name ty info
866 occ_str = showSDoc (braces (ppr fcall <+> ppr ty))
867 -- The "occurrence name" of a ccall is the full info about the
868 -- ccall; it is encoded, but may have embedded spaces etc!
870 name = mkFCallName uniq occ_str
874 `setAllStrictnessInfo` Just strict_sig
876 (_, tau) = tcSplitForAllTys ty
877 (arg_tys, _) = tcSplitFunTys tau
878 arity = length arg_tys
879 strict_sig = mkStrictSig (mkTopDmdType (replicate arity evalDmd) TopRes)
883 %************************************************************************
885 \subsection{DictFuns and default methods}
887 %************************************************************************
889 Important notes about dict funs and default methods
890 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
891 Dict funs and default methods are *not* ImplicitIds. Their definition
892 involves user-written code, so we can't figure out their strictness etc
893 based on fixed info, as we can for constructors and record selectors (say).
895 We build them as LocalIds, but with External Names. This ensures that
896 they are taken to account by free-variable finding and dependency
897 analysis (e.g. CoreFVs.exprFreeVars).
899 Why shouldn't they be bound as GlobalIds? Because, in particular, if
900 they are globals, the specialiser floats dict uses above their defns,
901 which prevents good simplifications happening. Also the strictness
902 analyser treats a occurrence of a GlobalId as imported and assumes it
903 contains strictness in its IdInfo, which isn't true if the thing is
904 bound in the same module as the occurrence.
906 It's OK for dfuns to be LocalIds, because we form the instance-env to
907 pass on to the next module (md_insts) in CoreTidy, afer tidying
908 and globalising the top-level Ids.
910 BUT make sure they are *exported* LocalIds (mkExportedLocalId) so
911 that they aren't discarded by the occurrence analyser.
914 mkDefaultMethodId dm_name ty = mkExportedLocalId dm_name ty
916 mkDictFunId :: Name -- Name to use for the dict fun;
923 mkDictFunId dfun_name inst_tyvars dfun_theta clas inst_tys
924 = mkExportedLocalId dfun_name dfun_ty
926 dfun_ty = mkSigmaTy inst_tyvars dfun_theta (mkDictTy clas inst_tys)
928 {- 1 dec 99: disable the Mark Jones optimisation for the sake
929 of compatibility with Hugs.
930 See `types/InstEnv' for a discussion related to this.
932 (class_tyvars, sc_theta, _, _) = classBigSig clas
933 not_const (clas, tys) = not (isEmptyVarSet (tyVarsOfTypes tys))
934 sc_theta' = substClasses (zipTopTvSubst class_tyvars inst_tys) sc_theta
935 dfun_theta = case inst_decl_theta of
936 [] -> [] -- If inst_decl_theta is empty, then we don't
937 -- want to have any dict arguments, so that we can
938 -- expose the constant methods.
940 other -> nub (inst_decl_theta ++ filter not_const sc_theta')
941 -- Otherwise we pass the superclass dictionaries to
942 -- the dictionary function; the Mark Jones optimisation.
944 -- NOTE the "nub". I got caught by this one:
945 -- class Monad m => MonadT t m where ...
946 -- instance Monad m => MonadT (EnvT env) m where ...
947 -- Here, the inst_decl_theta has (Monad m); but so
948 -- does the sc_theta'!
950 -- NOTE the "not_const". I got caught by this one too:
951 -- class Foo a => Baz a b where ...
952 -- instance Wob b => Baz T b where..
953 -- Now sc_theta' has Foo T
958 %************************************************************************
960 \subsection{Un-definable}
962 %************************************************************************
964 These Ids can't be defined in Haskell. They could be defined in
965 unfoldings in the wired-in GHC.Prim interface file, but we'd have to
966 ensure that they were definitely, definitely inlined, because there is
967 no curried identifier for them. That's what mkCompulsoryUnfolding
968 does. If we had a way to get a compulsory unfolding from an interface
969 file, we could do that, but we don't right now.
971 unsafeCoerce# isn't so much a PrimOp as a phantom identifier, that
972 just gets expanded into a type coercion wherever it occurs. Hence we
973 add it as a built-in Id with an unfolding here.
975 The type variables we use here are "open" type variables: this means
976 they can unify with both unlifted and lifted types. Hence we provide
977 another gun with which to shoot yourself in the foot.
980 mkWiredInIdName mod fs uniq id
981 = mkWiredInName mod (mkOccNameFS varName fs) uniq Nothing (AnId id) UserSyntax
983 unsafeCoerceName = mkWiredInIdName gHC_PRIM FSLIT("unsafeCoerce#") unsafeCoerceIdKey unsafeCoerceId
984 nullAddrName = mkWiredInIdName gHC_PRIM FSLIT("nullAddr#") nullAddrIdKey nullAddrId
985 seqName = mkWiredInIdName gHC_PRIM FSLIT("seq") seqIdKey seqId
986 realWorldName = mkWiredInIdName gHC_PRIM FSLIT("realWorld#") realWorldPrimIdKey realWorldPrimId
987 lazyIdName = mkWiredInIdName gHC_BASE FSLIT("lazy") lazyIdKey lazyId
989 errorName = mkWiredInIdName gHC_ERR FSLIT("error") errorIdKey eRROR_ID
990 recSelErrorName = mkWiredInIdName gHC_ERR FSLIT("recSelError") recSelErrorIdKey rEC_SEL_ERROR_ID
991 runtimeErrorName = mkWiredInIdName gHC_ERR FSLIT("runtimeError") runtimeErrorIdKey rUNTIME_ERROR_ID
992 irrefutPatErrorName = mkWiredInIdName gHC_ERR FSLIT("irrefutPatError") irrefutPatErrorIdKey iRREFUT_PAT_ERROR_ID
993 recConErrorName = mkWiredInIdName gHC_ERR FSLIT("recConError") recConErrorIdKey rEC_CON_ERROR_ID
994 patErrorName = mkWiredInIdName gHC_ERR FSLIT("patError") patErrorIdKey pAT_ERROR_ID
995 noMethodBindingErrorName = mkWiredInIdName gHC_ERR FSLIT("noMethodBindingError")
996 noMethodBindingErrorIdKey nO_METHOD_BINDING_ERROR_ID
997 nonExhaustiveGuardsErrorName
998 = mkWiredInIdName gHC_ERR FSLIT("nonExhaustiveGuardsError")
999 nonExhaustiveGuardsErrorIdKey nON_EXHAUSTIVE_GUARDS_ERROR_ID
1003 -- unsafeCoerce# :: forall a b. a -> b
1005 = pcMiscPrelId unsafeCoerceName ty info
1007 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1010 ty = mkForAllTys [openAlphaTyVar,openBetaTyVar]
1011 (mkFunTy openAlphaTy openBetaTy)
1012 [x] = mkTemplateLocals [openAlphaTy]
1013 rhs = mkLams [openAlphaTyVar,openBetaTyVar,x] $
1014 -- Note (Coerce openBetaTy openAlphaTy) (Var x)
1015 Cast (Var x) (mkUnsafeCoercion openAlphaTy openBetaTy)
1017 -- nullAddr# :: Addr#
1018 -- The reason is is here is because we don't provide
1019 -- a way to write this literal in Haskell.
1021 = pcMiscPrelId nullAddrName addrPrimTy info
1023 info = noCafIdInfo `setUnfoldingInfo`
1024 mkCompulsoryUnfolding (Lit nullAddrLit)
1027 = pcMiscPrelId seqName ty info
1029 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
1032 ty = mkForAllTys [alphaTyVar,openBetaTyVar]
1033 (mkFunTy alphaTy (mkFunTy openBetaTy openBetaTy))
1034 [x,y] = mkTemplateLocals [alphaTy, openBetaTy]
1035 rhs = mkLams [alphaTyVar,openBetaTyVar,x,y] (Case (Var x) x openBetaTy [(DEFAULT, [], Var y)])
1037 -- lazy :: forall a?. a? -> a? (i.e. works for unboxed types too)
1038 -- Used to lazify pseq: pseq a b = a `seq` lazy b
1040 -- Also, no strictness: by being a built-in Id, all the info about lazyId comes from here,
1041 -- not from GHC.Base.hi. This is important, because the strictness
1042 -- analyser will spot it as strict!
1044 -- Also no unfolding in lazyId: it gets "inlined" by a HACK in the worker/wrapper pass
1045 -- (see WorkWrap.wwExpr)
1046 -- We could use inline phases to do this, but that would be vulnerable to changes in
1047 -- phase numbering....we must inline precisely after strictness analysis.
1049 = pcMiscPrelId lazyIdName ty info
1052 ty = mkForAllTys [alphaTyVar] (mkFunTy alphaTy alphaTy)
1054 lazyIdUnfolding :: CoreExpr -- Used to expand 'lazyId' after strictness anal
1055 lazyIdUnfolding = mkLams [openAlphaTyVar,x] (Var x)
1057 [x] = mkTemplateLocals [openAlphaTy]
1060 @realWorld#@ used to be a magic literal, \tr{void#}. If things get
1061 nasty as-is, change it back to a literal (@Literal@).
1063 voidArgId is a Local Id used simply as an argument in functions
1064 where we just want an arg to avoid having a thunk of unlifted type.
1066 x = \ void :: State# RealWorld -> (# p, q #)
1068 This comes up in strictness analysis
1071 realWorldPrimId -- :: State# RealWorld
1072 = pcMiscPrelId realWorldName realWorldStatePrimTy
1073 (noCafIdInfo `setUnfoldingInfo` evaldUnfolding)
1074 -- The evaldUnfolding makes it look that realWorld# is evaluated
1075 -- which in turn makes Simplify.interestingArg return True,
1076 -- which in turn makes INLINE things applied to realWorld# likely
1079 voidArgId -- :: State# RealWorld
1080 = mkSysLocal FSLIT("void") voidArgIdKey realWorldStatePrimTy
1084 %************************************************************************
1086 \subsection[PrelVals-error-related]{@error@ and friends; @trace@}
1088 %************************************************************************
1090 GHC randomly injects these into the code.
1092 @patError@ is just a version of @error@ for pattern-matching
1093 failures. It knows various ``codes'' which expand to longer
1094 strings---this saves space!
1096 @absentErr@ is a thing we put in for ``absent'' arguments. They jolly
1097 well shouldn't be yanked on, but if one is, then you will get a
1098 friendly message from @absentErr@ (rather than a totally random
1101 @parError@ is a special version of @error@ which the compiler does
1102 not know to be a bottoming Id. It is used in the @_par_@ and @_seq_@
1103 templates, but we don't ever expect to generate code for it.
1107 :: Id -- Should be of type (forall a. Addr# -> a)
1108 -- where Addr# points to a UTF8 encoded string
1109 -> Type -- The type to instantiate 'a'
1110 -> String -- The string to print
1113 mkRuntimeErrorApp err_id res_ty err_msg
1114 = mkApps (Var err_id) [Type res_ty, err_string]
1116 err_string = Lit (mkStringLit err_msg)
1118 rEC_SEL_ERROR_ID = mkRuntimeErrorId recSelErrorName
1119 rUNTIME_ERROR_ID = mkRuntimeErrorId runtimeErrorName
1120 iRREFUT_PAT_ERROR_ID = mkRuntimeErrorId irrefutPatErrorName
1121 rEC_CON_ERROR_ID = mkRuntimeErrorId recConErrorName
1122 pAT_ERROR_ID = mkRuntimeErrorId patErrorName
1123 nO_METHOD_BINDING_ERROR_ID = mkRuntimeErrorId noMethodBindingErrorName
1124 nON_EXHAUSTIVE_GUARDS_ERROR_ID = mkRuntimeErrorId nonExhaustiveGuardsErrorName
1126 -- The runtime error Ids take a UTF8-encoded string as argument
1127 mkRuntimeErrorId name = pc_bottoming_Id name runtimeErrorTy
1128 runtimeErrorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTy addrPrimTy openAlphaTy)
1132 eRROR_ID = pc_bottoming_Id errorName errorTy
1135 errorTy = mkSigmaTy [openAlphaTyVar] [] (mkFunTys [mkListTy charTy] openAlphaTy)
1136 -- Notice the openAlphaTyVar. It says that "error" can be applied
1137 -- to unboxed as well as boxed types. This is OK because it never
1138 -- returns, so the return type is irrelevant.
1142 %************************************************************************
1144 \subsection{Utilities}
1146 %************************************************************************
1149 pcMiscPrelId :: Name -> Type -> IdInfo -> Id
1150 pcMiscPrelId name ty info
1151 = mkVanillaGlobal name ty info
1152 -- We lie and say the thing is imported; otherwise, we get into
1153 -- a mess with dependency analysis; e.g., core2stg may heave in
1154 -- random calls to GHCbase.unpackPS__. If GHCbase is the module
1155 -- being compiled, then it's just a matter of luck if the definition
1156 -- will be in "the right place" to be in scope.
1158 pc_bottoming_Id name ty
1159 = pcMiscPrelId name ty bottoming_info
1161 bottoming_info = vanillaIdInfo `setAllStrictnessInfo` Just strict_sig
1162 -- Do *not* mark them as NoCafRefs, because they can indeed have
1163 -- CAF refs. For example, pAT_ERROR_ID calls GHC.Err.untangle,
1164 -- which has some CAFs
1165 -- In due course we may arrange that these error-y things are
1166 -- regarded by the GC as permanently live, in which case we
1167 -- can give them NoCaf info. As it is, any function that calls
1168 -- any pc_bottoming_Id will itself have CafRefs, which bloats
1171 strict_sig = mkStrictSig (mkTopDmdType [evalDmd] BotRes)
1172 -- These "bottom" out, no matter what their arguments
1174 (openAlphaTyVar:openBetaTyVar:_) = openAlphaTyVars
1175 openAlphaTy = mkTyVarTy openAlphaTyVar
1176 openBetaTy = mkTyVarTy openBetaTyVar