2 % (c) The University of Glasgow 2002
4 \section[ByteCodeGen]{Generate bytecode from Core}
7 module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where
9 #include "HsVersions.h"
12 import ByteCodeFFI ( mkMarshalCode, moan64 )
13 import ByteCodeAsm ( CompiledByteCode(..), UnlinkedBCO,
14 assembleBCO, assembleBCOs, iNTERP_STACK_CHECK_THRESH )
15 import ByteCodeLink ( lookupStaticPtr )
18 import Name ( Name, getName, mkSystemVarName )
21 import ForeignCall ( ForeignCall(..), CCallTarget(..), CCallSpec(..) )
22 import HscTypes ( TypeEnv, typeEnvTyCons, typeEnvClasses )
23 import CoreUtils ( exprType )
25 import PprCore ( pprCoreExpr )
26 import Literal ( Literal(..), literalType )
27 import PrimOp ( PrimOp(..) )
28 import CoreFVs ( freeVars )
29 import Type ( isUnLiftedType, splitTyConApp_maybe )
30 import DataCon ( DataCon, dataConTag, fIRST_TAG, dataConTyCon,
31 isUnboxedTupleCon, isNullaryRepDataCon, dataConWorkId,
33 import TyCon ( TyCon, tyConFamilySize, isDataTyCon,
34 tyConDataCons, isUnboxedTupleTyCon )
35 import Class ( Class, classTyCon )
36 import Type ( Type, repType, splitFunTys, dropForAlls, pprType )
38 import DataCon ( dataConRepArity )
39 import Var ( isTyVar )
40 import VarSet ( VarSet, varSetElems )
41 import TysPrim ( arrayPrimTyCon, mutableArrayPrimTyCon,
42 byteArrayPrimTyCon, mutableByteArrayPrimTyCon
44 import DynFlags ( DynFlags, DynFlag(..) )
45 import ErrUtils ( showPass, dumpIfSet_dyn )
46 import Unique ( mkPseudoUniqueE )
47 import FastString ( FastString(..), unpackFS )
48 import Panic ( GhcException(..) )
49 import SMRep ( typeCgRep, arrWordsHdrSize, arrPtrsHdrSize, StgWord,
50 CgRep(..), cgRepSizeW, isFollowableArg, idCgRep )
51 import Bitmap ( intsToReverseBitmap, mkBitmap )
53 import Constants ( wORD_SIZE )
55 import Data.List ( intersperse, sortBy, zip4, zip6, partition )
56 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8,
58 import Foreign.C ( CInt )
59 import Control.Exception ( throwDyn )
61 import GHC.Exts ( Int(..), ByteArray# )
63 import Control.Monad ( when )
64 import Data.Char ( ord, chr )
66 -- -----------------------------------------------------------------------------
67 -- Generating byte code for a complete module
69 byteCodeGen :: DynFlags
72 -> IO CompiledByteCode
73 byteCodeGen dflags binds tycs
74 = do showPass dflags "ByteCodeGen"
76 let flatBinds = [ (bndr, freeVars rhs)
77 | (bndr, rhs) <- flattenBinds binds]
79 (BcM_State final_ctr mallocd, proto_bcos)
80 <- runBc (mapM schemeTopBind flatBinds)
82 when (notNull mallocd)
83 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
85 dumpIfSet_dyn dflags Opt_D_dump_BCOs
86 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
88 assembleBCOs proto_bcos tycs
90 -- -----------------------------------------------------------------------------
91 -- Generating byte code for an expression
93 -- Returns: (the root BCO for this expression,
94 -- a list of auxilary BCOs resulting from compiling closures)
95 coreExprToBCOs :: DynFlags
98 coreExprToBCOs dflags expr
99 = do showPass dflags "ByteCodeGen"
101 -- create a totally bogus name for the top-level BCO; this
102 -- should be harmless, since it's never used for anything
103 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
104 invented_id = mkLocalId invented_name (panic "invented_id's type")
106 (BcM_State final_ctr mallocd, proto_bco)
107 <- runBc (schemeTopBind (invented_id, freeVars expr))
109 when (notNull mallocd)
110 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
112 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
114 assembleBCO proto_bco
117 -- -----------------------------------------------------------------------------
118 -- Compilation schema for the bytecode generator
120 type BCInstrList = OrdList BCInstr
122 type Sequel = Int -- back off to this depth before ENTER
124 -- Maps Ids to the offset from the stack _base_ so we don't have
125 -- to mess with it after each push/pop.
126 type BCEnv = FiniteMap Id Int -- To find vars on the stack
128 ppBCEnv :: BCEnv -> SDoc
131 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
134 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
135 cmp_snd x y = compare (snd x) (snd y)
137 -- Create a BCO and do a spot of peephole optimisation on the insns
142 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
146 -> Bool -- True <=> is a return point, rather than a function
149 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap
150 is_ret mallocd_blocks
153 protoBCOInstrs = maybe_with_stack_check,
154 protoBCOBitmap = bitmap,
155 protoBCOBitmapSize = bitmap_size,
156 protoBCOArity = arity,
157 protoBCOExpr = origin,
158 protoBCOPtrs = mallocd_blocks
161 -- Overestimate the stack usage (in words) of this BCO,
162 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
163 -- stack check. (The interpreter always does a stack check
164 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
165 -- BCO anyway, so we only need to add an explicit on in the
166 -- (hopefully rare) cases when the (overestimated) stack use
167 -- exceeds iNTERP_STACK_CHECK_THRESH.
168 maybe_with_stack_check
170 -- don't do stack checks at return points;
171 -- everything is aggregated up to the top BCO
172 -- (which must be a function)
173 | stack_overest >= 65535
174 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
176 | stack_overest >= iNTERP_STACK_CHECK_THRESH
177 = STKCHECK stack_overest : peep_d
179 = peep_d -- the supposedly common case
181 stack_overest = sum (map bciStackUse peep_d)
183 -- Merge local pushes
184 peep_d = peep (fromOL instrs_ordlist)
186 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
187 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
188 peep (PUSH_L off1 : PUSH_L off2 : rest)
189 = PUSH_LL off1 (off2-1) : peep rest
195 argBits :: [CgRep] -> [Bool]
198 | isFollowableArg rep = False : argBits args
199 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
201 -- -----------------------------------------------------------------------------
204 -- Compile code for the right-hand side of a top-level binding
206 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
209 schemeTopBind (id, rhs)
210 | Just data_con <- isDataConWorkId_maybe id,
211 isNullaryRepDataCon data_con
212 = -- Special case for the worker of a nullary data con.
213 -- It'll look like this: Nil = /\a -> Nil a
214 -- If we feed it into schemeR, we'll get
216 -- because mkConAppCode treats nullary constructor applications
217 -- by just re-using the single top-level definition. So
218 -- for the worker itself, we must allocate it directly.
219 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
220 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
223 = schemeR [{- No free variables -}] (id, rhs)
225 -- -----------------------------------------------------------------------------
228 -- Compile code for a right-hand side, to give a BCO that,
229 -- when executed with the free variables and arguments on top of the stack,
230 -- will return with a pointer to the result on top of the stack, after
231 -- removing the free variables and arguments.
233 -- Park the resulting BCO in the monad. Also requires the
234 -- variable to which this value was bound, so as to give the
235 -- resulting BCO a name.
237 schemeR :: [Id] -- Free vars of the RHS, ordered as they
238 -- will appear in the thunk. Empty for
239 -- top-level things, which have no free vars.
240 -> (Id, AnnExpr Id VarSet)
241 -> BcM (ProtoBCO Name)
242 schemeR fvs (nm, rhs)
246 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
247 $$ pprCoreExpr (deAnnotate rhs)
253 = schemeR_wrk fvs nm rhs (collect [] rhs)
255 collect xs (_, AnnNote note e) = collect xs e
256 collect xs (_, AnnCast e _) = collect xs e
257 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
258 collect xs (_, not_lambda) = (reverse xs, not_lambda)
260 schemeR_wrk fvs nm original_body (args, body)
262 all_args = reverse args ++ fvs
263 arity = length all_args
264 -- all_args are the args in reverse order. We're compiling a function
265 -- \fv1..fvn x1..xn -> e
266 -- i.e. the fvs come first
268 szsw_args = map idSizeW all_args
269 szw_args = sum szsw_args
270 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
272 -- make the arg bitmap
273 bits = argBits (reverse (map idCgRep all_args))
274 bitmap_size = length bits
275 bitmap = mkBitmap bits
277 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
278 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
279 arity bitmap_size bitmap False{-not alts-})
282 fvsToEnv :: BCEnv -> VarSet -> [Id]
283 -- Takes the free variables of a right-hand side, and
284 -- delivers an ordered list of the local variables that will
285 -- be captured in the thunk for the RHS
286 -- The BCEnv argument tells which variables are in the local
287 -- environment: these are the ones that should be captured
289 -- The code that constructs the thunk, and the code that executes
290 -- it, have to agree about this layout
291 fvsToEnv p fvs = [v | v <- varSetElems fvs,
292 isId v, -- Could be a type variable
295 -- -----------------------------------------------------------------------------
298 -- Compile code to apply the given expression to the remaining args
299 -- on the stack, returning a HNF.
300 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
302 -- Delegate tail-calls to schemeT.
303 schemeE d s p e@(AnnApp f a)
306 schemeE d s p e@(AnnVar v)
307 | not (isUnLiftedType v_type)
308 = -- Lifted-type thing; push it in the normal way
312 = -- Returning an unlifted value.
313 -- Heave it on the stack, SLIDE, and RETURN.
314 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
315 returnBc (push -- value onto stack
316 `appOL` mkSLIDE szw (d-s) -- clear to sequel
317 `snocOL` RETURN_UBX v_rep) -- go
320 v_rep = typeCgRep v_type
322 schemeE d s p (AnnLit literal)
323 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
324 let l_rep = typeCgRep (literalType literal)
325 in returnBc (push -- value onto stack
326 `appOL` mkSLIDE szw (d-s) -- clear to sequel
327 `snocOL` RETURN_UBX l_rep) -- go
330 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
331 | (AnnVar v, args_r_to_l) <- splitApp rhs,
332 Just data_con <- isDataConWorkId_maybe v,
333 dataConRepArity data_con == length args_r_to_l
334 = -- Special case for a non-recursive let whose RHS is a
335 -- saturatred constructor application.
336 -- Just allocate the constructor and carry on
337 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
338 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
339 returnBc (alloc_code `appOL` body_code)
341 -- General case for let. Generates correct, if inefficient, code in
343 schemeE d s p (AnnLet binds (_,body))
344 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
345 AnnRec xs_n_rhss -> unzip xs_n_rhss
348 fvss = map (fvsToEnv p' . fst) rhss
350 -- Sizes of free vars
351 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
353 -- the arity of each rhs
354 arities = map (length . fst . collect []) rhss
356 -- This p', d' defn is safe because all the items being pushed
357 -- are ptrs, so all have size 1. d' and p' reflect the stack
358 -- after the closures have been allocated in the heap (but not
359 -- filled in), and pointers to them parked on the stack.
360 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
362 zipE = zipEqual "schemeE"
364 -- ToDo: don't build thunks for things with no free variables
365 build_thunk dd [] size bco off arity
366 = returnBc (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
368 mkap | arity == 0 = MKAP
370 build_thunk dd (fv:fvs) size bco off arity = do
371 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
372 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
373 returnBc (push_code `appOL` more_push_code)
375 alloc_code = toOL (zipWith mkAlloc sizes arities)
376 where mkAlloc sz 0 = ALLOC_AP sz
377 mkAlloc sz arity = ALLOC_PAP arity sz
379 compile_bind d' fvs x rhs size arity off = do
380 bco <- schemeR fvs (x,rhs)
381 build_thunk d' fvs size bco off arity
384 [ compile_bind d' fvs x rhs size arity n
385 | (fvs, x, rhs, size, arity, n) <-
386 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
389 body_code <- schemeE d' s p' body
390 thunk_codes <- sequence compile_binds
391 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
395 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
396 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
398 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
400 -- case .... of a { DEFAULT -> ... }
401 -- becuse the return convention for both are identical.
403 -- Note that it does not matter losing the void-rep thing from the
404 -- envt (it won't be bound now) because we never look such things up.
406 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
407 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
409 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
410 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
411 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
413 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
414 | isUnboxedTupleCon dc
415 -- Similarly, convert
416 -- case .... of x { (# a #) -> ... }
418 -- case .... of a { DEFAULT -> ... }
419 = --trace "automagic mashing of case alts (# a #)" $
420 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
422 schemeE d s p (AnnCase scrut bndr _ alts)
423 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
425 schemeE d s p (AnnNote note (_, body))
428 schemeE d s p (AnnCast (_, body) _)
432 = pprPanic "ByteCodeGen.schemeE: unhandled case"
433 (pprCoreExpr (deAnnotate' other))
436 -- Compile code to do a tail call. Specifically, push the fn,
437 -- slide the on-stack app back down to the sequel depth,
438 -- and enter. Four cases:
441 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
442 -- The int will be on the stack. Generate a code sequence
443 -- to convert it to the relevant constructor, SLIDE and ENTER.
445 -- 1. The fn denotes a ccall. Defer to generateCCall.
447 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
448 -- it simply as b -- since the representations are identical
449 -- (the VoidArg takes up zero stack space). Also, spot
450 -- (# b #) and treat it as b.
452 -- 3. Application of a constructor, by defn saturated.
453 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
454 -- then the ptrs, and then do PACK and RETURN.
456 -- 4. Otherwise, it must be a function call. Push the args
457 -- right to left, SLIDE and ENTER.
459 schemeT :: Int -- Stack depth
460 -> Sequel -- Sequel depth
461 -> BCEnv -- stack env
462 -> AnnExpr' Id VarSet
467 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
468 -- = panic "schemeT ?!?!"
470 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
474 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
475 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
476 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
477 returnBc (push `appOL` tagToId_sequence
478 `appOL` mkSLIDE 1 (d+arg_words-s)
482 | Just (CCall ccall_spec) <- isFCallId_maybe fn
483 = generateCCall d s p ccall_spec fn args_r_to_l
485 -- Case 2: Constructor application
486 | Just con <- maybe_saturated_dcon,
487 isUnboxedTupleCon con
488 = case args_r_to_l of
489 [arg1,arg2] | isVoidArgAtom arg1 ->
490 unboxedTupleReturn d s p arg2
491 [arg1,arg2] | isVoidArgAtom arg2 ->
492 unboxedTupleReturn d s p arg1
493 _other -> unboxedTupleException
495 -- Case 3: Ordinary data constructor
496 | Just con <- maybe_saturated_dcon
497 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
498 returnBc (alloc_con `appOL`
499 mkSLIDE 1 (d - s) `snocOL`
502 -- Case 4: Tail call of function
504 = doTailCall d s p fn args_r_to_l
507 -- Detect and extract relevant info for the tagToEnum kludge.
508 maybe_is_tagToEnum_call
509 = let extract_constr_Names ty
510 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
512 = map (getName . dataConWorkId) (tyConDataCons tyc)
513 -- NOTE: use the worker name, not the source name of
514 -- the DataCon. See DataCon.lhs for details.
516 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
519 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
520 -> case isPrimOpId_maybe v of
521 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
525 -- Extract the args (R->L) and fn
526 -- The function will necessarily be a variable,
527 -- because we are compiling a tail call
528 (AnnVar fn, args_r_to_l) = splitApp app
530 -- Only consider this to be a constructor application iff it is
531 -- saturated. Otherwise, we'll call the constructor wrapper.
532 n_args = length args_r_to_l
534 = case isDataConWorkId_maybe fn of
535 Just con | dataConRepArity con == n_args -> Just con
538 -- -----------------------------------------------------------------------------
539 -- Generate code to build a constructor application,
540 -- leaving it on top of the stack
542 mkConAppCode :: Int -> Sequel -> BCEnv
543 -> DataCon -- The data constructor
544 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
547 mkConAppCode orig_d s p con [] -- Nullary constructor
548 = ASSERT( isNullaryRepDataCon con )
549 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
550 -- Instead of doing a PACK, which would allocate a fresh
551 -- copy of this constructor, use the single shared version.
553 mkConAppCode orig_d s p con args_r_to_l
554 = ASSERT( dataConRepArity con == length args_r_to_l )
555 do_pushery orig_d (non_ptr_args ++ ptr_args)
557 -- The args are already in reverse order, which is the way PACK
558 -- expects them to be. We must push the non-ptrs after the ptrs.
559 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
561 do_pushery d (arg:args)
562 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
563 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
564 returnBc (push `appOL` more_push_code)
566 = returnBc (unitOL (PACK con n_arg_words))
568 n_arg_words = d - orig_d
571 -- -----------------------------------------------------------------------------
572 -- Returning an unboxed tuple with one non-void component (the only
573 -- case we can handle).
575 -- Remember, we don't want to *evaluate* the component that is being
576 -- returned, even if it is a pointed type. We always just return.
579 :: Int -> Sequel -> BCEnv
580 -> AnnExpr' Id VarSet -> BcM BCInstrList
581 unboxedTupleReturn d s p arg = do
582 (push, sz) <- pushAtom d p arg
583 returnBc (push `appOL`
584 mkSLIDE sz (d-s) `snocOL`
585 RETURN_UBX (atomRep arg))
587 -- -----------------------------------------------------------------------------
588 -- Generate code for a tail-call
591 :: Int -> Sequel -> BCEnv
592 -> Id -> [AnnExpr' Id VarSet]
594 doTailCall init_d s p fn args
595 = do_pushes init_d args (map atomRep args)
597 do_pushes d [] reps = do
598 ASSERT( null reps ) return ()
599 (push_fn, sz) <- pushAtom d p (AnnVar fn)
600 ASSERT( sz == 1 ) return ()
601 returnBc (push_fn `appOL` (
602 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
604 do_pushes d args reps = do
605 let (push_apply, n, rest_of_reps) = findPushSeq reps
606 (these_args, rest_of_args) = splitAt n args
607 (next_d, push_code) <- push_seq d these_args
608 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
609 -- ^^^ for the PUSH_APPLY_ instruction
610 returnBc (push_code `appOL` (push_apply `consOL` instrs))
612 push_seq d [] = return (d, nilOL)
613 push_seq d (arg:args) = do
614 (push_code, sz) <- pushAtom d p arg
615 (final_d, more_push_code) <- push_seq (d+sz) args
616 return (final_d, push_code `appOL` more_push_code)
618 -- v. similar to CgStackery.findMatch, ToDo: merge
619 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
620 = (PUSH_APPLY_PPPPPP, 6, rest)
621 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
622 = (PUSH_APPLY_PPPPP, 5, rest)
623 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
624 = (PUSH_APPLY_PPPP, 4, rest)
625 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
626 = (PUSH_APPLY_PPP, 3, rest)
627 findPushSeq (PtrArg: PtrArg: rest)
628 = (PUSH_APPLY_PP, 2, rest)
629 findPushSeq (PtrArg: rest)
630 = (PUSH_APPLY_P, 1, rest)
631 findPushSeq (VoidArg: rest)
632 = (PUSH_APPLY_V, 1, rest)
633 findPushSeq (NonPtrArg: rest)
634 = (PUSH_APPLY_N, 1, rest)
635 findPushSeq (FloatArg: rest)
636 = (PUSH_APPLY_F, 1, rest)
637 findPushSeq (DoubleArg: rest)
638 = (PUSH_APPLY_D, 1, rest)
639 findPushSeq (LongArg: rest)
640 = (PUSH_APPLY_L, 1, rest)
642 = panic "ByteCodeGen.findPushSeq"
644 -- -----------------------------------------------------------------------------
647 doCase :: Int -> Sequel -> BCEnv
648 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
649 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
651 doCase d s p (_,scrut)
652 bndr alts is_unboxed_tuple
654 -- Top of stack is the return itbl, as usual.
655 -- underneath it is the pointer to the alt_code BCO.
656 -- When an alt is entered, it assumes the returned value is
657 -- on top of the itbl.
660 -- An unlifted value gets an extra info table pushed on top
661 -- when it is returned.
662 unlifted_itbl_sizeW | isAlgCase = 0
665 -- depth of stack after the return value has been pushed
666 d_bndr = d + ret_frame_sizeW + idSizeW bndr
668 -- depth of stack after the extra info table for an unboxed return
669 -- has been pushed, if any. This is the stack depth at the
671 d_alts = d_bndr + unlifted_itbl_sizeW
673 -- Env in which to compile the alts, not including
674 -- any vars bound by the alts themselves
675 p_alts = addToFM p bndr (d_bndr - 1)
677 bndr_ty = idType bndr
678 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
680 -- given an alt, return a discr and code for it.
681 codeALt alt@(DEFAULT, _, (_,rhs))
682 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
683 returnBc (NoDiscr, rhs_code)
684 codeAlt alt@(discr, bndrs, (_,rhs))
685 -- primitive or nullary constructor alt: no need to UNPACK
686 | null real_bndrs = do
687 rhs_code <- schemeE d_alts s p_alts rhs
688 returnBc (my_discr alt, rhs_code)
689 -- algebraic alt with some binders
690 | ASSERT(isAlgCase) otherwise =
692 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
693 ptr_sizes = map idSizeW ptrs
694 nptrs_sizes = map idSizeW nptrs
695 bind_sizes = ptr_sizes ++ nptrs_sizes
696 size = sum ptr_sizes + sum nptrs_sizes
697 -- the UNPACK instruction unpacks in reverse order...
698 p' = addListToFM p_alts
699 (zip (reverse (ptrs ++ nptrs))
700 (mkStackOffsets d_alts (reverse bind_sizes)))
702 rhs_code <- schemeE (d_alts+size) s p' rhs
703 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
705 real_bndrs = filter (not.isTyVar) bndrs
708 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
709 my_discr (DataAlt dc, binds, rhs)
710 | isUnboxedTupleCon dc
711 = unboxedTupleException
713 = DiscrP (dataConTag dc - fIRST_TAG)
714 my_discr (LitAlt l, binds, rhs)
715 = case l of MachInt i -> DiscrI (fromInteger i)
716 MachFloat r -> DiscrF (fromRational r)
717 MachDouble r -> DiscrD (fromRational r)
718 MachChar i -> DiscrI (ord i)
719 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
722 | not isAlgCase = Nothing
724 = case [dc | (DataAlt dc, _, _) <- alts] of
726 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
728 -- the bitmap is relative to stack depth d, i.e. before the
729 -- BCO, info table and return value are pushed on.
730 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
731 -- except that here we build the bitmap from the known bindings of
732 -- things that are pointers, whereas in CgBindery the code builds the
733 -- bitmap from the free slots and unboxed bindings.
735 bitmap = intsToReverseBitmap d{-size-} (sortLe (<=) rel_slots)
738 rel_slots = concat (map spread binds)
740 | isFollowableArg (idCgRep id) = [ rel_offset ]
742 where rel_offset = d - offset - 1
745 alt_stuff <- mapM codeAlt alts
746 alt_final <- mkMultiBranch maybe_ncons alt_stuff
748 alt_bco_name = getName bndr
749 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
750 0{-no arity-} d{-bitmap size-} bitmap True{-is alts-}
752 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
753 -- "\n bitmap = " ++ show bitmap) $ do
754 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
755 alt_bco' <- emitBc alt_bco
757 | isAlgCase = PUSH_ALTS alt_bco'
758 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
759 returnBc (push_alts `consOL` scrut_code)
762 -- -----------------------------------------------------------------------------
763 -- Deal with a CCall.
765 -- Taggedly push the args onto the stack R->L,
766 -- deferencing ForeignObj#s and adjusting addrs to point to
767 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
768 -- (machine) code for the ccall, and create bytecodes to call that and
769 -- then return in the right way.
771 generateCCall :: Int -> Sequel -- stack and sequel depths
773 -> CCallSpec -- where to call
774 -> Id -- of target, for type info
775 -> [AnnExpr' Id VarSet] -- args (atoms)
778 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
781 addr_sizeW = cgRepSizeW NonPtrArg
783 -- Get the args on the stack, with tags and suitably
784 -- dereferenced for the CCall. For each arg, return the
785 -- depth to the first word of the bits for that arg, and the
786 -- CgRep of what was actually pushed.
788 pargs d [] = returnBc []
790 = let arg_ty = repType (exprType (deAnnotate' a))
792 in case splitTyConApp_maybe arg_ty of
793 -- Don't push the FO; instead push the Addr# it
796 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
797 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
798 parg_ArrayishRep arrPtrsHdrSize d p a
800 returnBc ((code,NonPtrArg):rest)
802 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
803 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
804 parg_ArrayishRep arrWordsHdrSize d p a
806 returnBc ((code,NonPtrArg):rest)
808 -- Default case: push taggedly, but otherwise intact.
810 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
811 pargs (d+sz_a) az `thenBc` \ rest ->
812 returnBc ((code_a, atomRep a) : rest)
814 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
815 -- the stack but then advance it over the headers, so as to
816 -- point to the payload.
817 parg_ArrayishRep hdrSize d p a
818 = pushAtom d p a `thenBc` \ (push_fo, _) ->
819 -- The ptr points at the header. Advance it over the
820 -- header and then pretend this is an Addr#.
821 returnBc (push_fo `snocOL` SWIZZLE 0 hdrSize)
824 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
826 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
828 push_args = concatOL pushs_arg
829 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
831 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
832 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
834 = reverse (tail a_reps_pushed_r_to_l)
836 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
837 -- push_args is the code to do that.
838 -- d_after_args is the stack depth once the args are on.
840 -- Get the result rep.
841 (returns_void, r_rep)
842 = case maybe_getCCallReturnRep (idType fn) of
843 Nothing -> (True, VoidArg)
844 Just rr -> (False, rr)
846 Because the Haskell stack grows down, the a_reps refer to
847 lowest to highest addresses in that order. The args for the call
848 are on the stack. Now push an unboxed Addr# indicating
849 the C function to call. Then push a dummy placeholder for the
850 result. Finally, emit a CCALL insn with an offset pointing to the
851 Addr# just pushed, and a literal field holding the mallocville
852 address of the piece of marshalling code we generate.
853 So, just prior to the CCALL insn, the stack looks like this
854 (growing down, as usual):
859 Addr# address_of_C_fn
860 <placeholder-for-result#> (must be an unboxed type)
862 The interpreter then calls the marshall code mentioned
863 in the CCALL insn, passing it (& <placeholder-for-result#>),
864 that is, the addr of the topmost word in the stack.
865 When this returns, the placeholder will have been
866 filled in. The placeholder is slid down to the sequel
867 depth, and we RETURN.
869 This arrangement makes it simple to do f-i-dynamic since the Addr#
870 value is the first arg anyway.
872 The marshalling code is generated specifically for this
873 call site, and so knows exactly the (Haskell) stack
874 offsets of the args, fn address and placeholder. It
875 copies the args to the C stack, calls the stacked addr,
876 and parks the result back in the placeholder. The interpreter
877 calls it as a normal C call, assuming it has a signature
878 void marshall_code ( StgWord* ptr_to_top_of_stack )
880 -- resolve static address
884 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
886 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
889 get_target_info `thenBc` \ (is_static, static_target_addr) ->
892 -- Get the arg reps, zapping the leading Addr# in the dynamic case
893 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
894 | is_static = a_reps_pushed_RAW
895 | otherwise = if null a_reps_pushed_RAW
896 then panic "ByteCodeGen.generateCCall: dyn with no args"
897 else tail a_reps_pushed_RAW
900 (push_Addr, d_after_Addr)
902 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
903 d_after_args + addr_sizeW)
904 | otherwise -- is already on the stack
905 = (nilOL, d_after_args)
907 -- Push the return placeholder. For a call returning nothing,
908 -- this is a VoidArg (tag).
909 r_sizeW = cgRepSizeW r_rep
910 d_after_r = d_after_Addr + r_sizeW
911 r_lit = mkDummyLiteral r_rep
912 push_r = (if returns_void
914 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
916 -- generate the marshalling code we're going to call
919 arg1_offW = r_sizeW + addr_sizeW
920 args_offW = map (arg1_offW +)
921 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
923 ioToBc (mkMarshalCode cconv
924 (r_offW, r_rep) addr_offW
925 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
926 recordMallocBc addr_of_marshaller `thenBc_`
928 -- Offset of the next stack frame down the stack. The CCALL
929 -- instruction needs to describe the chunk of stack containing
930 -- the ccall args to the GC, so it needs to know how large it
931 -- is. See comment in Interpreter.c with the CCALL instruction.
932 stk_offset = d_after_r - s
935 do_call = unitOL (CCALL stk_offset (castPtr addr_of_marshaller))
937 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
938 `snocOL` RETURN_UBX r_rep
940 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
943 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
947 -- Make a dummy literal, to be used as a placeholder for FFI return
948 -- values on the stack.
949 mkDummyLiteral :: CgRep -> Literal
952 NonPtrArg -> MachWord 0
953 DoubleArg -> MachDouble 0
954 FloatArg -> MachFloat 0
955 _ -> moan64 "mkDummyLiteral" (ppr pr)
959 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
960 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
963 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
965 -- Alternatively, for call-targets returning nothing, convert
967 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
968 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
972 maybe_getCCallReturnRep :: Type -> Maybe CgRep
973 maybe_getCCallReturnRep fn_ty
974 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
976 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
978 = case splitTyConApp_maybe (repType r_ty) of
979 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
981 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
982 || r_reps == [VoidArg] )
983 && isUnboxedTupleTyCon r_tycon
984 && case maybe_r_rep_to_go of
986 Just r_rep -> r_rep /= PtrArg
987 -- if it was, it would be impossible
988 -- to create a valid return value
989 -- placeholder on the stack
990 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
993 --trace (showSDoc (ppr (a_reps, r_reps))) $
994 if ok then maybe_r_rep_to_go else blargh
996 -- Compile code which expects an unboxed Int on the top of stack,
997 -- (call it i), and pushes the i'th closure in the supplied list
999 implement_tagToId :: [Name] -> BcM BCInstrList
1000 implement_tagToId names
1001 = ASSERT( notNull names )
1002 getLabelsBc (length names) `thenBc` \ labels ->
1003 getLabelBc `thenBc` \ label_fail ->
1004 getLabelBc `thenBc` \ label_exit ->
1005 zip4 labels (tail labels ++ [label_fail])
1006 [0 ..] names `bind` \ infos ->
1007 map (mkStep label_exit) infos `bind` \ steps ->
1008 returnBc (concatOL steps
1010 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1012 mkStep l_exit (my_label, next_label, n, name_for_n)
1013 = toOL [LABEL my_label,
1014 TESTEQ_I n next_label,
1019 -- -----------------------------------------------------------------------------
1022 -- Push an atom onto the stack, returning suitable code & number of
1023 -- stack words used.
1025 -- The env p must map each variable to the highest- numbered stack
1026 -- slot for it. For example, if the stack has depth 4 and we
1027 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1028 -- the tag in stack[5], the stack will have depth 6, and p must map v
1029 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1030 -- depth 6 stack has valid words 0 .. 5.
1032 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1034 pushAtom d p (AnnApp f (_, AnnType _))
1035 = pushAtom d p (snd f)
1037 pushAtom d p (AnnNote note e)
1038 = pushAtom d p (snd e)
1040 pushAtom d p (AnnLam x e)
1042 = pushAtom d p (snd e)
1044 pushAtom d p (AnnVar v)
1046 | idCgRep v == VoidArg
1047 = returnBc (nilOL, 0)
1050 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1052 | Just primop <- isPrimOpId_maybe v
1053 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1055 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1056 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1057 -- d - d_v the number of words between the TOS
1058 -- and the 1st slot of the object
1060 -- d - d_v - 1 the offset from the TOS of the 1st slot
1062 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1065 -- Having found the last slot, we proceed to copy the right number of
1066 -- slots on to the top of the stack.
1068 | otherwise -- v must be a global variable
1070 returnBc (unitOL (PUSH_G (getName v)), sz)
1076 pushAtom d p (AnnLit lit)
1078 MachLabel fs _ -> code NonPtrArg
1079 MachWord w -> code NonPtrArg
1080 MachInt i -> code PtrArg
1081 MachFloat r -> code FloatArg
1082 MachDouble r -> code DoubleArg
1083 MachChar c -> code NonPtrArg
1084 MachStr s -> pushStr s
1087 = let size_host_words = cgRepSizeW rep
1088 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1092 = let getMallocvilleAddr
1094 FastString _ n _ fp _ ->
1095 -- we could grab the Ptr from the ForeignPtr,
1096 -- but then we have no way to control its lifetime.
1097 -- In reality it'll probably stay alive long enoungh
1098 -- by virtue of the global FastString table, but
1099 -- to be on the safe side we copy the string into
1100 -- a malloc'd area of memory.
1101 ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1102 recordMallocBc ptr `thenBc_`
1104 withForeignPtr fp $ \p -> do
1105 memcpy ptr p (fromIntegral n)
1106 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1110 getMallocvilleAddr `thenBc` \ addr ->
1111 -- Get the addr on the stack, untaggedly
1112 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1114 pushAtom d p (AnnCast e _)
1115 = pushAtom d p (snd e)
1118 = pprPanic "ByteCodeGen.pushAtom"
1119 (pprCoreExpr (deAnnotate (undefined, other)))
1121 foreign import ccall unsafe "memcpy"
1122 memcpy :: Ptr a -> Ptr b -> CInt -> IO ()
1125 -- -----------------------------------------------------------------------------
1126 -- Given a bunch of alts code and their discrs, do the donkey work
1127 -- of making a multiway branch using a switch tree.
1128 -- What a load of hassle!
1130 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1131 -- a hint; generates better code
1132 -- Nothing is always safe
1133 -> [(Discr, BCInstrList)]
1135 mkMultiBranch maybe_ncons raw_ways
1136 = let d_way = filter (isNoDiscr.fst) raw_ways
1138 (\w1 w2 -> leAlt (fst w1) (fst w2))
1139 (filter (not.isNoDiscr.fst) raw_ways)
1141 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1142 mkTree [] range_lo range_hi = returnBc the_default
1144 mkTree [val] range_lo range_hi
1145 | range_lo `eqAlt` range_hi
1146 = returnBc (snd val)
1148 = getLabelBc `thenBc` \ label_neq ->
1149 returnBc (mkTestEQ (fst val) label_neq
1151 `appOL` unitOL (LABEL label_neq)
1152 `appOL` the_default))
1154 mkTree vals range_lo range_hi
1155 = let n = length vals `div` 2
1156 vals_lo = take n vals
1157 vals_hi = drop n vals
1158 v_mid = fst (head vals_hi)
1160 getLabelBc `thenBc` \ label_geq ->
1161 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1162 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1163 returnBc (mkTestLT v_mid label_geq
1165 `appOL` unitOL (LABEL label_geq)
1169 = case d_way of [] -> unitOL CASEFAIL
1172 -- None of these will be needed if there are no non-default alts
1173 (mkTestLT, mkTestEQ, init_lo, init_hi)
1175 = panic "mkMultiBranch: awesome foursome"
1177 = case fst (head notd_ways) of {
1178 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1179 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1182 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1183 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1186 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1187 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1190 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1191 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1193 DiscrP algMaxBound )
1196 (algMinBound, algMaxBound)
1197 = case maybe_ncons of
1198 Just n -> (0, n - 1)
1199 Nothing -> (minBound, maxBound)
1201 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1202 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1203 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1204 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1205 NoDiscr `eqAlt` NoDiscr = True
1208 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1209 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1210 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1211 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1212 NoDiscr `leAlt` NoDiscr = True
1215 isNoDiscr NoDiscr = True
1218 dec (DiscrI i) = DiscrI (i-1)
1219 dec (DiscrP i) = DiscrP (i-1)
1220 dec other = other -- not really right, but if you
1221 -- do cases on floating values, you'll get what you deserve
1223 -- same snotty comment applies to the following
1225 minD, maxD :: Double
1231 mkTree notd_ways init_lo init_hi
1234 -- -----------------------------------------------------------------------------
1235 -- Supporting junk for the compilation schemes
1237 -- Describes case alts
1245 instance Outputable Discr where
1246 ppr (DiscrI i) = int i
1247 ppr (DiscrF f) = text (show f)
1248 ppr (DiscrD d) = text (show d)
1249 ppr (DiscrP i) = int i
1250 ppr NoDiscr = text "DEF"
1253 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1254 lookupBCEnv_maybe = lookupFM
1256 idSizeW :: Id -> Int
1257 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1259 unboxedTupleException :: a
1260 unboxedTupleException
1263 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1264 "\tto foreign import/export decls in source. Workaround:\n" ++
1265 "\tcompile this module to a .o file, then restart session."))
1268 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1271 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1272 -- The arguments are returned in *right-to-left* order
1273 splitApp (AnnApp (_,f) (_,a))
1274 | isTypeAtom a = splitApp f
1275 | otherwise = case splitApp f of
1276 (f', as) -> (f', a:as)
1277 splitApp (AnnNote n (_,e)) = splitApp e
1278 splitApp (AnnCast (_,e) _) = splitApp e
1279 splitApp e = (e, [])
1282 isTypeAtom :: AnnExpr' id ann -> Bool
1283 isTypeAtom (AnnType _) = True
1284 isTypeAtom _ = False
1286 isVoidArgAtom :: AnnExpr' id ann -> Bool
1287 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1288 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1289 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1290 isVoidArgAtom _ = False
1292 atomRep :: AnnExpr' Id ann -> CgRep
1293 atomRep (AnnVar v) = typeCgRep (idType v)
1294 atomRep (AnnLit l) = typeCgRep (literalType l)
1295 atomRep (AnnNote n b) = atomRep (snd b)
1296 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1297 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1298 atomRep (AnnCast b _) = atomRep (snd b)
1299 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1301 isPtrAtom :: AnnExpr' Id ann -> Bool
1302 isPtrAtom e = atomRep e == PtrArg
1304 -- Let szsw be the sizes in words of some items pushed onto the stack,
1305 -- which has initial depth d'. Return the values which the stack environment
1306 -- should map these items to.
1307 mkStackOffsets :: Int -> [Int] -> [Int]
1308 mkStackOffsets original_depth szsw
1309 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1311 -- -----------------------------------------------------------------------------
1312 -- The bytecode generator's monad
1316 nextlabel :: Int, -- for generating local labels
1317 malloced :: [Ptr ()] } -- ptrs malloced for current BCO
1318 -- Should be free()d when it is GCd
1320 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1322 ioToBc :: IO a -> BcM a
1323 ioToBc io = BcM $ \st -> do
1327 runBc :: BcM r -> IO (BcM_State, r)
1328 runBc (BcM m) = m (BcM_State 0 [])
1330 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1331 thenBc (BcM expr) cont = BcM $ \st0 -> do
1332 (st1, q) <- expr st0
1337 thenBc_ :: BcM a -> BcM b -> BcM b
1338 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1339 (st1, q) <- expr st0
1340 (st2, r) <- cont st1
1343 returnBc :: a -> BcM a
1344 returnBc result = BcM $ \st -> (return (st, result))
1346 instance Monad BcM where
1351 emitBc :: ([Ptr ()] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1353 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1355 recordMallocBc :: Ptr a -> BcM ()
1357 = BcM $ \st -> return (st{malloced = castPtr a : malloced st}, ())
1359 getLabelBc :: BcM Int
1361 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1363 getLabelsBc :: Int -> BcM [Int]
1365 = BcM $ \st -> let ctr = nextlabel st
1366 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])