2 % (c) The University of Glasgow 2006
3 % (c) The AQUA Project, Glasgow University, 1998
6 This module contains definitions for the IdInfo for things that
7 have a standard form, namely:
11 - method and superclass selectors
12 - primitive operations
16 mkDictFunId, mkDefaultMethodId,
20 mkPrimOpId, mkFCallId, mkTickBoxOpId, mkBreakPointOpId,
22 mkReboxingAlt, wrapNewTypeBody, unwrapNewTypeBody,
23 wrapFamInstBody, unwrapFamInstScrut,
24 mkUnpackCase, mkProductBox,
26 -- And some particular Ids; see below for why they are wired in
27 wiredInIds, ghcPrimIds,
28 unsafeCoerceName, unsafeCoerceId, realWorldPrimId,
29 voidArgId, nullAddrId, seqId, lazyId, lazyIdKey
32 #include "HsVersions.h"
41 import CoreUtils ( exprType, mkCoerce )
52 import Var ( Var, TyVar, mkCoVar, mkExportedLocalVar )
58 import BasicTypes hiding ( SuccessFlag(..) )
66 %************************************************************************
68 \subsection{Wired in Ids}
70 %************************************************************************
74 There are several reasons why an Id might appear in the wiredInIds:
76 (1) The ghcPrimIds are wired in because they can't be defined in
77 Haskell at all, although the can be defined in Core. They have
78 compulsory unfoldings, so they are always inlined and they have
79 no definition site. Their home module is GHC.Prim, so they
80 also have a description in primops.txt.pp, where they are called
83 (2) The 'error' function, eRROR_ID, is wired in because we don't yet have
84 a way to express in an interface file that the result type variable
85 is 'open'; that is can be unified with an unboxed type
87 [The interface file format now carry such information, but there's
88 no way yet of expressing at the definition site for these
89 error-reporting functions that they have an 'open'
90 result type. -- sof 1/99]
92 (3) Other error functions (rUNTIME_ERROR_ID) are wired in (a) because
93 the desugarer generates code that mentiones them directly, and
94 (b) for the same reason as eRROR_ID
96 (4) lazyId is wired in because the wired-in version overrides the
97 strictness of the version defined in GHC.Base
99 In cases (2-4), the function has a definition in a library module, and
100 can be called; but the wired-in version means that the details are
101 never read from that module's interface file; instead, the full definition
108 ++ errorIds -- Defined in MkCore
111 -- These Ids are exported from GHC.Prim
114 = [ -- These can't be defined in Haskell, but they have
115 -- perfectly reasonable unfoldings in Core
123 %************************************************************************
125 \subsection{Data constructors}
127 %************************************************************************
129 The wrapper for a constructor is an ordinary top-level binding that evaluates
130 any strict args, unboxes any args that are going to be flattened, and calls
133 We're going to build a constructor that looks like:
135 data (Data a, C b) => T a b = T1 !a !Int b
138 \d1::Data a, d2::C b ->
139 \p q r -> case p of { p ->
141 Con T1 [a,b] [p,q,r]}}
145 * d2 is thrown away --- a context in a data decl is used to make sure
146 one *could* construct dictionaries at the site the constructor
147 is used, but the dictionary isn't actually used.
149 * We have to check that we can construct Data dictionaries for
150 the types a and Int. Once we've done that we can throw d1 away too.
152 * We use (case p of q -> ...) to evaluate p, rather than "seq" because
153 all that matters is that the arguments are evaluated. "seq" is
154 very careful to preserve evaluation order, which we don't need
157 You might think that we could simply give constructors some strictness
158 info, like PrimOps, and let CoreToStg do the let-to-case transformation.
159 But we don't do that because in the case of primops and functions strictness
160 is a *property* not a *requirement*. In the case of constructors we need to
161 do something active to evaluate the argument.
163 Making an explicit case expression allows the simplifier to eliminate
164 it in the (common) case where the constructor arg is already evaluated.
166 Note [Wrappers for data instance tycons]
167 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
168 In the case of data instances, the wrapper also applies the coercion turning
169 the representation type into the family instance type to cast the result of
170 the wrapper. For example, consider the declarations
172 data family Map k :: * -> *
173 data instance Map (a, b) v = MapPair (Map a (Pair b v))
175 The tycon to which the datacon MapPair belongs gets a unique internal
176 name of the form :R123Map, and we call it the representation tycon.
177 In contrast, Map is the family tycon (accessible via
178 tyConFamInst_maybe). A coercion allows you to move between
179 representation and family type. It is accessible from :R123Map via
180 tyConFamilyCoercion_maybe and has kind
182 Co123Map a b v :: {Map (a, b) v ~ :R123Map a b v}
184 The wrapper and worker of MapPair get the types
187 $WMapPair :: forall a b v. Map a (Map a b v) -> Map (a, b) v
188 $WMapPair a b v = MapPair a b v `cast` sym (Co123Map a b v)
191 MapPair :: forall a b v. Map a (Map a b v) -> :R123Map a b v
193 This coercion is conditionally applied by wrapFamInstBody.
195 It's a bit more complicated if the data instance is a GADT as well!
197 data instance T [a] where
198 T1 :: forall b. b -> T [Maybe b]
200 Hence we translate to
203 $WT1 :: forall b. b -> T [Maybe b]
204 $WT1 b v = T1 (Maybe b) b (Maybe b) v
205 `cast` sym (Co7T (Maybe b))
208 T1 :: forall c b. (c ~ Maybe b) => b -> :R7T c
210 -- Coercion from family type to representation type
211 Co7T a :: T [a] ~ :R7T a
214 mkDataConIds :: Name -> Name -> DataCon -> DataConIds
215 mkDataConIds wrap_name wkr_name data_con
216 | isNewTyCon tycon -- Newtype, only has a worker
217 = DCIds Nothing nt_work_id
219 | any isBanged all_strict_marks -- Algebraic, needs wrapper
220 || not (null eq_spec) -- NB: LoadIface.ifaceDeclSubBndrs
221 || isFamInstTyCon tycon -- depends on this test
222 = DCIds (Just alg_wrap_id) wrk_id
224 | otherwise -- Algebraic, no wrapper
225 = DCIds Nothing wrk_id
227 (univ_tvs, ex_tvs, eq_spec,
228 eq_theta, dict_theta, orig_arg_tys, res_ty) = dataConFullSig data_con
229 tycon = dataConTyCon data_con -- The representation TyCon (not family)
231 ----------- Worker (algebraic data types only) --------------
232 -- The *worker* for the data constructor is the function that
233 -- takes the representation arguments and builds the constructor.
234 wrk_id = mkGlobalId (DataConWorkId data_con) wkr_name
235 (dataConRepType data_con) wkr_info
237 wkr_arity = dataConRepArity data_con
238 wkr_info = noCafIdInfo
239 `setArityInfo` wkr_arity
240 `setStrictnessInfo` Just wkr_sig
241 `setUnfoldingInfo` evaldUnfolding -- Record that it's evaluated,
244 wkr_sig = mkStrictSig (mkTopDmdType (replicate wkr_arity topDmd) cpr_info)
245 -- Note [Data-con worker strictness]
246 -- Notice that we do *not* say the worker is strict
247 -- even if the data constructor is declared strict
248 -- e.g. data T = MkT !(Int,Int)
249 -- Why? Because the *wrapper* is strict (and its unfolding has case
250 -- expresssions that do the evals) but the *worker* itself is not.
251 -- If we pretend it is strict then when we see
252 -- case x of y -> $wMkT y
253 -- the simplifier thinks that y is "sure to be evaluated" (because
254 -- $wMkT is strict) and drops the case. No, $wMkT is not strict.
256 -- When the simplifer sees a pattern
257 -- case e of MkT x -> ...
258 -- it uses the dataConRepStrictness of MkT to mark x as evaluated;
259 -- but that's fine... dataConRepStrictness comes from the data con
260 -- not from the worker Id.
262 cpr_info | isProductTyCon tycon &&
265 wkr_arity <= mAX_CPR_SIZE = retCPR
267 -- RetCPR is only true for products that are real data types;
268 -- that is, not unboxed tuples or [non-recursive] newtypes
270 ----------- Workers for newtypes --------------
271 nt_work_id = mkGlobalId (DataConWrapId data_con) wkr_name wrap_ty nt_work_info
272 nt_work_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
273 `setArityInfo` 1 -- Arity 1
274 `setUnfoldingInfo` newtype_unf
275 id_arg1 = mkTemplateLocal 1 (head orig_arg_tys)
276 newtype_unf = ASSERT2( isVanillaDataCon data_con &&
277 isSingleton orig_arg_tys, ppr data_con )
278 -- Note [Newtype datacons]
279 mkCompulsoryUnfolding $
280 mkLams wrap_tvs $ Lam id_arg1 $
281 wrapNewTypeBody tycon res_ty_args (Var id_arg1)
284 ----------- Wrapper --------------
285 -- We used to include the stupid theta in the wrapper's args
286 -- but now we don't. Instead the type checker just injects these
287 -- extra constraints where necessary.
288 wrap_tvs = (univ_tvs `minusList` map fst eq_spec) ++ ex_tvs
289 res_ty_args = substTyVars (mkTopTvSubst eq_spec) univ_tvs
290 eq_tys = mkPredTys eq_theta
291 dict_tys = mkPredTys dict_theta
292 wrap_ty = mkForAllTys wrap_tvs $ mkFunTys eq_tys $ mkFunTys dict_tys $
293 mkFunTys orig_arg_tys $ res_ty
294 -- NB: watch out here if you allow user-written equality
295 -- constraints in data constructor signatures
297 ----------- Wrappers for algebraic data types --------------
298 alg_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty alg_wrap_info
299 alg_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo
300 `setArityInfo` wrap_arity
301 -- It's important to specify the arity, so that partial
302 -- applications are treated as values
303 `setInlinePragInfo` alwaysInlinePragma
304 `setUnfoldingInfo` wrap_unf
305 `setStrictnessInfo` Just wrap_sig
307 all_strict_marks = dataConExStricts data_con ++ dataConStrictMarks data_con
308 wrap_sig = mkStrictSig (mkTopDmdType arg_dmds cpr_info)
309 arg_dmds = map mk_dmd all_strict_marks
310 mk_dmd str | isBanged str = evalDmd
311 | otherwise = lazyDmd
312 -- The Cpr info can be important inside INLINE rhss, where the
313 -- wrapper constructor isn't inlined.
314 -- And the argument strictness can be important too; we
315 -- may not inline a contructor when it is partially applied.
317 -- data W = C !Int !Int !Int
318 -- ...(let w = C x in ...(w p q)...)...
319 -- we want to see that w is strict in its two arguments
321 wrap_unf = mkInlineUnfolding (Just (length dict_args + length id_args)) wrap_rhs
322 wrap_rhs = mkLams wrap_tvs $
324 mkLams dict_args $ mkLams id_args $
325 foldr mk_case con_app
326 (zip (dict_args ++ id_args) all_strict_marks)
329 con_app _ rep_ids = wrapFamInstBody tycon res_ty_args $
330 Var wrk_id `mkTyApps` res_ty_args
332 -- Equality evidence:
333 `mkTyApps` map snd eq_spec
335 `mkVarApps` reverse rep_ids
337 (dict_args,i2) = mkLocals 1 dict_tys
338 (id_args,i3) = mkLocals i2 orig_arg_tys
340 (eq_args,_) = mkCoVarLocals i3 eq_tys
342 mkCoVarLocals i [] = ([],i)
343 mkCoVarLocals i (x:xs) = let (ys,j) = mkCoVarLocals (i+1) xs
344 y = mkCoVar (mkSysTvName (mkBuiltinUnique i)
349 :: (Id, HsBang) -- Arg, strictness
350 -> (Int -> [Id] -> CoreExpr) -- Body
351 -> Int -- Next rep arg id
352 -> [Id] -- Rep args so far, reversed
354 mk_case (arg,strict) body i rep_args
356 HsNoBang -> body i (arg:rep_args)
357 HsUnpack -> unboxProduct i (Var arg) (idType arg) the_body
359 the_body i con_args = body i (reverse con_args ++ rep_args)
360 _other -- HsUnpackFailed and HsStrict
361 | isUnLiftedType (idType arg) -> body i (arg:rep_args)
362 | otherwise -> Case (Var arg) arg res_ty
363 [(DEFAULT,[], body i (arg:rep_args))]
365 mAX_CPR_SIZE :: Arity
367 -- We do not treat very big tuples as CPR-ish:
368 -- a) for a start we get into trouble because there aren't
369 -- "enough" unboxed tuple types (a tiresome restriction,
371 -- b) more importantly, big unboxed tuples get returned mainly
372 -- on the stack, and are often then allocated in the heap
373 -- by the caller. So doing CPR for them may in fact make
376 mkLocals :: Int -> [Type] -> ([Id], Int)
377 mkLocals i tys = (zipWith mkTemplateLocal [i..i+n-1] tys, i+n)
382 Note [Newtype datacons]
383 ~~~~~~~~~~~~~~~~~~~~~~~
384 The "data constructor" for a newtype should always be vanilla. At one
385 point this wasn't true, because the newtype arising from
388 newtype T:D a = D:D (C a)
389 so the data constructor for T:C had a single argument, namely the
390 predicate (C a). But now we treat that as an ordinary argument, not
391 part of the theta-type, so all is well.
394 %************************************************************************
396 \subsection{Dictionary selectors}
398 %************************************************************************
400 Selecting a field for a dictionary. If there is just one field, then
401 there's nothing to do.
403 Dictionary selectors may get nested forall-types. Thus:
406 op :: forall b. Ord b => a -> b -> b
408 Then the top-level type for op is
410 op :: forall a. Foo a =>
414 This is unlike ordinary record selectors, which have all the for-alls
415 at the outside. When dealing with classes it's very convenient to
416 recover the original type signature from the class op selector.
419 mkDictSelId :: Bool -- True <=> don't include the unfolding
420 -- Little point on imports without -O, because the
421 -- dictionary itself won't be visible
422 -> Name -- Name of one of the *value* selectors
423 -- (dictionary superclass or method)
425 mkDictSelId no_unf name clas
426 = mkGlobalId (ClassOpId clas) name sel_ty info
428 sel_ty = mkForAllTys tyvars (mkFunTy (idType dict_id) (idType the_arg_id))
429 -- We can't just say (exprType rhs), because that would give a type
431 -- for a single-op class (after all, the selector is the identity)
432 -- But it's type must expose the representation of the dictionary
433 -- to get (say) C a -> (a -> a)
435 base_info = noCafIdInfo
437 `setStrictnessInfo` Just strict_sig
438 `setUnfoldingInfo` (if no_unf then noUnfolding
439 else mkImplicitUnfolding rhs)
440 -- In module where class op is defined, we must add
441 -- the unfolding, even though it'll never be inlined
442 -- becuase we use that to generate a top-level binding
445 info = base_info `setSpecInfo` mkSpecInfo [rule]
446 `setInlinePragInfo` neverInlinePragma
447 -- Add a magic BuiltinRule, and never inline it
448 -- so that the rule is always available to fire.
449 -- See Note [ClassOp/DFun selection] in TcInstDcls
451 n_ty_args = length tyvars
453 -- This is the built-in rule that goes
454 -- op (dfT d1 d2) ---> opT d1 d2
455 rule = BuiltinRule { ru_name = fsLit "Class op " `appendFS`
456 occNameFS (getOccName name)
458 , ru_nargs = n_ty_args + 1
459 , ru_try = dictSelRule val_index n_ty_args n_eq_args }
461 -- The strictness signature is of the form U(AAAVAAAA) -> T
462 -- where the V depends on which item we are selecting
463 -- It's worth giving one, so that absence info etc is generated
464 -- even if the selector isn't inlined
465 strict_sig = mkStrictSig (mkTopDmdType [arg_dmd] TopRes)
466 arg_dmd | new_tycon = evalDmd
467 | otherwise = Eval (Prod [ if the_arg_id == id then evalDmd else Abs
470 tycon = classTyCon clas
471 new_tycon = isNewTyCon tycon
472 [data_con] = tyConDataCons tycon
473 tyvars = dataConUnivTyVars data_con
474 arg_tys = dataConRepArgTys data_con -- Includes the dictionary superclasses
475 eq_theta = dataConEqTheta data_con
476 n_eq_args = length eq_theta
478 -- 'index' is a 0-index into the *value* arguments of the dictionary
479 val_index = assoc "MkId.mkDictSelId" sel_index_prs name
480 sel_index_prs = map idName (classAllSelIds clas) `zip` [0..]
482 the_arg_id = arg_ids !! val_index
483 pred = mkClassPred clas (mkTyVarTys tyvars)
484 dict_id = mkTemplateLocal 1 $ mkPredTy pred
485 arg_ids = mkTemplateLocalsNum 2 arg_tys
486 eq_ids = map mkWildEvBinder eq_theta
488 rhs = mkLams tyvars (Lam dict_id rhs_body)
489 rhs_body | new_tycon = unwrapNewTypeBody tycon (map mkTyVarTy tyvars) (Var dict_id)
490 | otherwise = Case (Var dict_id) dict_id (idType the_arg_id)
491 [(DataAlt data_con, eq_ids ++ arg_ids, Var the_arg_id)]
493 dictSelRule :: Int -> Arity -> Arity
494 -> IdUnfoldingFun -> [CoreExpr] -> Maybe CoreExpr
496 -- sel_i t1..tk (df s1..sn d1..dm) = op_i_helper s1..sn d1..dm
497 -- sel_i t1..tk (D t1..tk op1 ... opm) = opi
499 -- NB: the data constructor has the same number of type and
500 -- coercion args as the selector
502 -- This only works for *value* superclasses
503 -- There are no selector functions for equality superclasses
504 dictSelRule val_index n_ty_args n_eq_args id_unf args
505 | (dict_arg : _) <- drop n_ty_args args
506 , Just (_, _, con_args) <- exprIsConApp_maybe id_unf dict_arg
507 , let val_args = drop n_eq_args con_args
508 = Just (val_args !! val_index)
514 %************************************************************************
518 %************************************************************************
521 -- unbox a product type...
522 -- we will recurse into newtypes, casting along the way, and unbox at the
523 -- first product data constructor we find. e.g.
525 -- data PairInt = PairInt Int Int
526 -- newtype S = MkS PairInt
529 -- If we have e = MkT (MkS (PairInt 0 1)) and some body expecting a list of
530 -- ids, we get (modulo int passing)
532 -- case (e `cast` CoT) `cast` CoS of
533 -- PairInt a b -> body [a,b]
535 -- The Ints passed around are just for creating fresh locals
536 unboxProduct :: Int -> CoreExpr -> Type -> (Int -> [Id] -> CoreExpr) -> CoreExpr
537 unboxProduct i arg arg_ty body
540 result = mkUnpackCase the_id arg con_args boxing_con rhs
541 (_tycon, _tycon_args, boxing_con, tys) = deepSplitProductType "unboxProduct" arg_ty
542 ([the_id], i') = mkLocals i [arg_ty]
543 (con_args, i'') = mkLocals i' tys
544 rhs = body i'' con_args
546 mkUnpackCase :: Id -> CoreExpr -> [Id] -> DataCon -> CoreExpr -> CoreExpr
547 -- (mkUnpackCase x e args Con body)
549 -- case (e `cast` ...) of bndr { Con args -> body }
551 -- the type of the bndr passed in is irrelevent
552 mkUnpackCase bndr arg unpk_args boxing_con body
553 = Case cast_arg (setIdType bndr bndr_ty) (exprType body) [(DataAlt boxing_con, unpk_args, body)]
555 (cast_arg, bndr_ty) = go (idType bndr) arg
557 | (tycon, tycon_args, _, _) <- splitProductType "mkUnpackCase" ty
558 , isNewTyCon tycon && not (isRecursiveTyCon tycon)
559 = go (newTyConInstRhs tycon tycon_args)
560 (unwrapNewTypeBody tycon tycon_args arg)
561 | otherwise = (arg, ty)
564 reboxProduct :: [Unique] -- uniques to create new local binders
565 -> Type -- type of product to box
566 -> ([Unique], -- remaining uniques
567 CoreExpr, -- boxed product
568 [Id]) -- Ids being boxed into product
571 (_tycon, _tycon_args, _pack_con, con_arg_tys) = deepSplitProductType "reboxProduct" ty
573 us' = dropList con_arg_tys us
575 arg_ids = zipWith (mkSysLocal (fsLit "rb")) us con_arg_tys
577 bind_rhs = mkProductBox arg_ids ty
580 (us', bind_rhs, arg_ids)
582 mkProductBox :: [Id] -> Type -> CoreExpr
583 mkProductBox arg_ids ty
586 (tycon, tycon_args, pack_con, _con_arg_tys) = splitProductType "mkProductBox" ty
589 | isNewTyCon tycon && not (isRecursiveTyCon tycon)
590 = wrap (mkProductBox arg_ids (newTyConInstRhs tycon tycon_args))
591 | otherwise = mkConApp pack_con (map Type tycon_args ++ map Var arg_ids)
593 wrap expr = wrapNewTypeBody tycon tycon_args expr
596 -- (mkReboxingAlt us con xs rhs) basically constructs the case
597 -- alternative (con, xs, rhs)
598 -- but it does the reboxing necessary to construct the *source*
599 -- arguments, xs, from the representation arguments ys.
601 -- data T = MkT !(Int,Int) Bool
603 -- mkReboxingAlt MkT [x,b] r
604 -- = (DataAlt MkT, [y::Int,z::Int,b], let x = (y,z) in r)
606 -- mkDataAlt should really be in DataCon, but it can't because
607 -- it manipulates CoreSyn.
610 :: [Unique] -- Uniques for the new Ids
612 -> [Var] -- Source-level args, including existential dicts
616 mkReboxingAlt us con args rhs
617 | not (any isMarkedUnboxed stricts)
618 = (DataAlt con, args, rhs)
622 (binds, args') = go args stricts us
624 (DataAlt con, args', mkLets binds rhs)
627 stricts = dataConExStricts con ++ dataConStrictMarks con
629 go [] _stricts _us = ([], [])
631 -- Type variable case
632 go (arg:args) stricts us
634 = let (binds, args') = go args stricts us
635 in (binds, arg:args')
637 -- Term variable case
638 go (arg:args) (str:stricts) us
639 | isMarkedUnboxed str
641 let (binds, unpacked_args') = go args stricts us'
642 (us', bind_rhs, unpacked_args) = reboxProduct us (idType arg)
644 (NonRec arg bind_rhs : binds, unpacked_args ++ unpacked_args')
646 = let (binds, args') = go args stricts us
647 in (binds, arg:args')
648 go (_ : _) [] _ = panic "mkReboxingAlt"
652 %************************************************************************
654 Wrapping and unwrapping newtypes and type families
656 %************************************************************************
659 wrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
660 -- The wrapper for the data constructor for a newtype looks like this:
661 -- newtype T a = MkT (a,Int)
662 -- MkT :: forall a. (a,Int) -> T a
663 -- MkT = /\a. \(x:(a,Int)). x `cast` sym (CoT a)
664 -- where CoT is the coercion TyCon assoicated with the newtype
666 -- The call (wrapNewTypeBody T [a] e) returns the
667 -- body of the wrapper, namely
668 -- e `cast` (CoT [a])
670 -- If a coercion constructor is provided in the newtype, then we use
671 -- it, otherwise the wrap/unwrap are both no-ops
673 -- If the we are dealing with a newtype *instance*, we have a second coercion
674 -- identifying the family instance with the constructor of the newtype
675 -- instance. This coercion is applied in any case (ie, composed with the
676 -- coercion constructor of the newtype or applied by itself).
678 wrapNewTypeBody tycon args result_expr
679 = wrapFamInstBody tycon args inner
682 | Just co_con <- newTyConCo_maybe tycon
683 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) result_expr
687 -- When unwrapping, we do *not* apply any family coercion, because this will
688 -- be done via a CoPat by the type checker. We have to do it this way as
689 -- computing the right type arguments for the coercion requires more than just
690 -- a spliting operation (cf, TcPat.tcConPat).
692 unwrapNewTypeBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
693 unwrapNewTypeBody tycon args result_expr
694 | Just co_con <- newTyConCo_maybe tycon
695 = mkCoerce (mkTyConApp co_con args) result_expr
699 -- If the type constructor is a representation type of a data instance, wrap
700 -- the expression into a cast adjusting the expression type, which is an
701 -- instance of the representation type, to the corresponding instance of the
702 -- family instance type.
703 -- See Note [Wrappers for data instance tycons]
704 wrapFamInstBody :: TyCon -> [Type] -> CoreExpr -> CoreExpr
705 wrapFamInstBody tycon args body
706 | Just co_con <- tyConFamilyCoercion_maybe tycon
707 = mkCoerce (mkSymCoercion (mkTyConApp co_con args)) body
711 unwrapFamInstScrut :: TyCon -> [Type] -> CoreExpr -> CoreExpr
712 unwrapFamInstScrut tycon args scrut
713 | Just co_con <- tyConFamilyCoercion_maybe tycon
714 = mkCoerce (mkTyConApp co_con args) scrut
720 %************************************************************************
722 \subsection{Primitive operations}
724 %************************************************************************
727 mkPrimOpId :: PrimOp -> Id
731 (tyvars,arg_tys,res_ty, arity, strict_sig) = primOpSig prim_op
732 ty = mkForAllTys tyvars (mkFunTys arg_tys res_ty)
733 name = mkWiredInName gHC_PRIM (primOpOcc prim_op)
734 (mkPrimOpIdUnique (primOpTag prim_op))
736 id = mkGlobalId (PrimOpId prim_op) name ty info
739 `setSpecInfo` mkSpecInfo (primOpRules prim_op name)
741 `setStrictnessInfo` Just strict_sig
743 -- For each ccall we manufacture a separate CCallOpId, giving it
744 -- a fresh unique, a type that is correct for this particular ccall,
745 -- and a CCall structure that gives the correct details about calling
748 -- The *name* of this Id is a local name whose OccName gives the full
749 -- details of the ccall, type and all. This means that the interface
750 -- file reader can reconstruct a suitable Id
752 mkFCallId :: Unique -> ForeignCall -> Type -> Id
753 mkFCallId uniq fcall ty
754 = ASSERT( isEmptyVarSet (tyVarsOfType ty) )
755 -- A CCallOpId should have no free type variables;
756 -- when doing substitutions won't substitute over it
757 mkGlobalId (FCallId fcall) name ty info
759 occ_str = showSDoc (braces (ppr fcall <+> ppr ty))
760 -- The "occurrence name" of a ccall is the full info about the
761 -- ccall; it is encoded, but may have embedded spaces etc!
763 name = mkFCallName uniq occ_str
767 `setStrictnessInfo` Just strict_sig
769 (_, tau) = tcSplitForAllTys ty
770 (arg_tys, _) = tcSplitFunTys tau
771 arity = length arg_tys
772 strict_sig = mkStrictSig (mkTopDmdType (replicate arity evalDmd) TopRes)
774 -- Tick boxes and breakpoints are both represented as TickBoxOpIds,
775 -- except for the type:
777 -- a plain HPC tick box has type (State# RealWorld)
778 -- a breakpoint Id has type forall a.a
780 -- The breakpoint Id will be applied to a list of arbitrary free variables,
781 -- which is why it needs a polymorphic type.
783 mkTickBoxOpId :: Unique -> Module -> TickBoxId -> Id
784 mkTickBoxOpId uniq mod ix = mkTickBox' uniq mod ix realWorldStatePrimTy
786 mkBreakPointOpId :: Unique -> Module -> TickBoxId -> Id
787 mkBreakPointOpId uniq mod ix = mkTickBox' uniq mod ix ty
788 where ty = mkSigmaTy [openAlphaTyVar] [] openAlphaTy
790 mkTickBox' :: Unique -> Module -> TickBoxId -> Type -> Id
791 mkTickBox' uniq mod ix ty = mkGlobalId (TickBoxOpId tickbox) name ty info
793 tickbox = TickBox mod ix
794 occ_str = showSDoc (braces (ppr tickbox))
795 name = mkTickBoxOpName uniq occ_str
800 %************************************************************************
802 \subsection{DictFuns and default methods}
804 %************************************************************************
806 Important notes about dict funs and default methods
807 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
808 Dict funs and default methods are *not* ImplicitIds. Their definition
809 involves user-written code, so we can't figure out their strictness etc
810 based on fixed info, as we can for constructors and record selectors (say).
812 We build them as LocalIds, but with External Names. This ensures that
813 they are taken to account by free-variable finding and dependency
814 analysis (e.g. CoreFVs.exprFreeVars).
816 Why shouldn't they be bound as GlobalIds? Because, in particular, if
817 they are globals, the specialiser floats dict uses above their defns,
818 which prevents good simplifications happening. Also the strictness
819 analyser treats a occurrence of a GlobalId as imported and assumes it
820 contains strictness in its IdInfo, which isn't true if the thing is
821 bound in the same module as the occurrence.
823 It's OK for dfuns to be LocalIds, because we form the instance-env to
824 pass on to the next module (md_insts) in CoreTidy, afer tidying
825 and globalising the top-level Ids.
827 BUT make sure they are *exported* LocalIds (mkExportedLocalId) so
828 that they aren't discarded by the occurrence analyser.
831 mkDefaultMethodId :: Id -- Selector Id
832 -> Name -- Default method name
833 -> Id -- Default method Id
834 mkDefaultMethodId sel_id dm_name = mkExportedLocalId dm_name (idType sel_id)
836 mkDictFunId :: Name -- Name to use for the dict fun;
843 mkDictFunId dfun_name inst_tyvars dfun_theta clas inst_tys
844 = mkExportedLocalVar (DFunId is_nt) dfun_name dfun_ty vanillaIdInfo
846 is_nt = isNewTyCon (classTyCon clas)
847 dfun_ty = mkSigmaTy inst_tyvars dfun_theta (mkDictTy clas inst_tys)
851 %************************************************************************
853 \subsection{Un-definable}
855 %************************************************************************
857 These Ids can't be defined in Haskell. They could be defined in
858 unfoldings in the wired-in GHC.Prim interface file, but we'd have to
859 ensure that they were definitely, definitely inlined, because there is
860 no curried identifier for them. That's what mkCompulsoryUnfolding
861 does. If we had a way to get a compulsory unfolding from an interface
862 file, we could do that, but we don't right now.
864 unsafeCoerce# isn't so much a PrimOp as a phantom identifier, that
865 just gets expanded into a type coercion wherever it occurs. Hence we
866 add it as a built-in Id with an unfolding here.
868 The type variables we use here are "open" type variables: this means
869 they can unify with both unlifted and lifted types. Hence we provide
870 another gun with which to shoot yourself in the foot.
873 lazyIdName, unsafeCoerceName, nullAddrName, seqName, realWorldName :: Name
874 unsafeCoerceName = mkWiredInIdName gHC_PRIM (fsLit "unsafeCoerce#") unsafeCoerceIdKey unsafeCoerceId
875 nullAddrName = mkWiredInIdName gHC_PRIM (fsLit "nullAddr#") nullAddrIdKey nullAddrId
876 seqName = mkWiredInIdName gHC_PRIM (fsLit "seq") seqIdKey seqId
877 realWorldName = mkWiredInIdName gHC_PRIM (fsLit "realWorld#") realWorldPrimIdKey realWorldPrimId
878 lazyIdName = mkWiredInIdName gHC_BASE (fsLit "lazy") lazyIdKey lazyId
882 ------------------------------------------------
883 -- unsafeCoerce# :: forall a b. a -> b
886 = pcMiscPrelId unsafeCoerceName ty info
888 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
891 ty = mkForAllTys [argAlphaTyVar,openBetaTyVar]
892 (mkFunTy argAlphaTy openBetaTy)
893 [x] = mkTemplateLocals [argAlphaTy]
894 rhs = mkLams [argAlphaTyVar,openBetaTyVar,x] $
895 Cast (Var x) (mkUnsafeCoercion argAlphaTy openBetaTy)
897 ------------------------------------------------
899 -- nullAddr# :: Addr#
900 -- The reason is is here is because we don't provide
901 -- a way to write this literal in Haskell.
902 nullAddrId = pcMiscPrelId nullAddrName addrPrimTy info
904 info = noCafIdInfo `setUnfoldingInfo`
905 mkCompulsoryUnfolding (Lit nullAddrLit)
907 ------------------------------------------------
908 seqId :: Id -- See Note [seqId magic]
909 seqId = pcMiscPrelId seqName ty info
911 info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs
912 `setSpecInfo` mkSpecInfo [seq_cast_rule]
915 ty = mkForAllTys [alphaTyVar,argBetaTyVar]
916 (mkFunTy alphaTy (mkFunTy argBetaTy argBetaTy))
917 [x,y] = mkTemplateLocals [alphaTy, argBetaTy]
918 rhs = mkLams [alphaTyVar,argBetaTyVar,x,y] (Case (Var x) x argBetaTy [(DEFAULT, [], Var y)])
920 -- See Note [Built-in RULES for seq]
921 seq_cast_rule = BuiltinRule { ru_name = fsLit "seq of cast"
924 , ru_try = match_seq_of_cast
927 match_seq_of_cast :: IdUnfoldingFun -> [CoreExpr] -> Maybe CoreExpr
928 -- See Note [Built-in RULES for seq]
929 match_seq_of_cast _ [Type _, Type res_ty, Cast scrut co, expr]
930 = Just (Var seqId `mkApps` [Type (fst (coercionKind co)), Type res_ty,
932 match_seq_of_cast _ _ = Nothing
934 ------------------------------------------------
935 lazyId :: Id -- See Note [lazyId magic]
936 lazyId = pcMiscPrelId lazyIdName ty info
939 ty = mkForAllTys [alphaTyVar] (mkFunTy alphaTy alphaTy)
944 'GHC.Prim.seq' is special in several ways.
946 a) Its second arg can have an unboxed type
949 b) Its fixity is set in LoadIface.ghcPrimIface
951 c) It has quite a bit of desugaring magic.
952 See DsUtils.lhs Note [Desugaring seq (1)] and (2) and (3)
954 d) There is some special rule handing: Note [User-defined RULES for seq]
956 e) See Note [Typing rule for seq] in TcExpr.
958 Note [User-defined RULES for seq]
959 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
960 Roman found situations where he had
962 where he knew that f (which was strict in n) would terminate if n did.
963 Notice that the result of (f n) is discarded. So it makes sense to
967 Rather than attempt some general analysis to support this, I've added
968 enough support that you can do this using a rewrite rule:
970 RULE "f/seq" forall n. seq (f n) e = seq n e
972 You write that rule. When GHC sees a case expression that discards
973 its result, it mentally transforms it to a call to 'seq' and looks for
974 a RULE. (This is done in Simplify.rebuildCase.) As usual, the
975 correctness of the rule is up to you.
977 To make this work, we need to be careful that the magical desugaring
978 done in Note [seqId magic] item (c) is *not* done on the LHS of a rule.
979 Or rather, we arrange to un-do it, in DsBinds.decomposeRuleLhs.
981 Note [Built-in RULES for seq]
982 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
983 We also have the following built-in rule for seq
985 seq (x `cast` co) y = seq x y
987 This eliminates unnecessary casts and also allows other seq rules to
988 match more often. Notably,
990 seq (f x `cast` co) y --> seq (f x) y
992 and now a user-defined rule for seq (see Note [User-defined RULES for seq])
998 lazy :: forall a?. a? -> a? (i.e. works for unboxed types too)
1000 Used to lazify pseq: pseq a b = a `seq` lazy b
1002 Also, no strictness: by being a built-in Id, all the info about lazyId comes from here,
1003 not from GHC.Base.hi. This is important, because the strictness
1004 analyser will spot it as strict!
1006 Also no unfolding in lazyId: it gets "inlined" by a HACK in CorePrep.
1007 It's very important to do this inlining *after* unfoldings are exposed
1008 in the interface file. Otherwise, the unfolding for (say) pseq in the
1009 interface file will not mention 'lazy', so if we inline 'pseq' we'll totally
1010 miss the very thing that 'lazy' was there for in the first place.
1011 See Trac #3259 for a real world example.
1013 lazyId is defined in GHC.Base, so we don't *have* to inline it. If it
1014 appears un-applied, we'll end up just calling it.
1016 -------------------------------------------------------------
1017 @realWorld#@ used to be a magic literal, \tr{void#}. If things get
1018 nasty as-is, change it back to a literal (@Literal@).
1020 voidArgId is a Local Id used simply as an argument in functions
1021 where we just want an arg to avoid having a thunk of unlifted type.
1023 x = \ void :: State# RealWorld -> (# p, q #)
1025 This comes up in strictness analysis
1028 realWorldPrimId :: Id
1029 realWorldPrimId -- :: State# RealWorld
1030 = pcMiscPrelId realWorldName realWorldStatePrimTy
1031 (noCafIdInfo `setUnfoldingInfo` evaldUnfolding)
1032 -- The evaldUnfolding makes it look that realWorld# is evaluated
1033 -- which in turn makes Simplify.interestingArg return True,
1034 -- which in turn makes INLINE things applied to realWorld# likely
1038 voidArgId -- :: State# RealWorld
1039 = mkSysLocal (fsLit "void") voidArgIdKey realWorldStatePrimTy
1044 pcMiscPrelId :: Name -> Type -> IdInfo -> Id
1045 pcMiscPrelId name ty info
1046 = mkVanillaGlobalWithInfo name ty info
1047 -- We lie and say the thing is imported; otherwise, we get into
1048 -- a mess with dependency analysis; e.g., core2stg may heave in
1049 -- random calls to GHCbase.unpackPS__. If GHCbase is the module
1050 -- being compiled, then it's just a matter of luck if the definition
1051 -- will be in "the right place" to be in scope.