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, zip5, 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 (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
257 collect xs (_, not_lambda) = (reverse xs, not_lambda)
259 schemeR_wrk fvs nm original_body (args, body)
261 all_args = reverse args ++ fvs
262 arity = length all_args
263 -- all_args are the args in reverse order. We're compiling a function
264 -- \fv1..fvn x1..xn -> e
265 -- i.e. the fvs come first
267 szsw_args = map idSizeW all_args
268 szw_args = sum szsw_args
269 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
271 -- make the arg bitmap
272 bits = argBits (reverse (map idCgRep all_args))
273 bitmap_size = length bits
274 bitmap = mkBitmap bits
276 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
277 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
278 arity bitmap_size bitmap False{-not alts-})
281 fvsToEnv :: BCEnv -> VarSet -> [Id]
282 -- Takes the free variables of a right-hand side, and
283 -- delivers an ordered list of the local variables that will
284 -- be captured in the thunk for the RHS
285 -- The BCEnv argument tells which variables are in the local
286 -- environment: these are the ones that should be captured
288 -- The code that constructs the thunk, and the code that executes
289 -- it, have to agree about this layout
290 fvsToEnv p fvs = [v | v <- varSetElems fvs,
291 isId v, -- Could be a type variable
294 -- -----------------------------------------------------------------------------
297 -- Compile code to apply the given expression to the remaining args
298 -- on the stack, returning a HNF.
299 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
301 -- Delegate tail-calls to schemeT.
302 schemeE d s p e@(AnnApp f a)
305 schemeE d s p e@(AnnVar v)
306 | not (isUnLiftedType v_type)
307 = -- Lifted-type thing; push it in the normal way
311 = -- Returning an unlifted value.
312 -- Heave it on the stack, SLIDE, and RETURN.
313 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
314 returnBc (push -- value onto stack
315 `appOL` mkSLIDE szw (d-s) -- clear to sequel
316 `snocOL` RETURN_UBX v_rep) -- go
319 v_rep = typeCgRep v_type
321 schemeE d s p (AnnLit literal)
322 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
323 let l_rep = typeCgRep (literalType literal)
324 in returnBc (push -- value onto stack
325 `appOL` mkSLIDE szw (d-s) -- clear to sequel
326 `snocOL` RETURN_UBX l_rep) -- go
329 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
330 | (AnnVar v, args_r_to_l) <- splitApp rhs,
331 Just data_con <- isDataConWorkId_maybe v,
332 dataConRepArity data_con == length args_r_to_l
333 = -- Special case for a non-recursive let whose RHS is a
334 -- saturatred constructor application.
335 -- Just allocate the constructor and carry on
336 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
337 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
338 returnBc (alloc_code `appOL` body_code)
340 -- General case for let. Generates correct, if inefficient, code in
342 schemeE d s p (AnnLet binds (_,body))
343 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
344 AnnRec xs_n_rhss -> unzip xs_n_rhss
347 fvss = map (fvsToEnv p' . fst) rhss
349 -- Sizes of free vars
350 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
352 -- the arity of each rhs
353 arities = map (length . fst . collect []) rhss
355 -- This p', d' defn is safe because all the items being pushed
356 -- are ptrs, so all have size 1. d' and p' reflect the stack
357 -- after the closures have been allocated in the heap (but not
358 -- filled in), and pointers to them parked on the stack.
359 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
361 zipE = zipEqual "schemeE"
363 -- ToDo: don't build thunks for things with no free variables
364 build_thunk dd [] size bco off
365 = returnBc (PUSH_BCO bco
366 `consOL` unitOL (MKAP (off+size) size))
367 build_thunk dd (fv:fvs) size bco off = do
368 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
369 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off
370 returnBc (push_code `appOL` more_push_code)
372 alloc_code = toOL (zipWith mkAlloc sizes arities)
373 where mkAlloc sz 0 = ALLOC_AP sz
374 mkAlloc sz arity = ALLOC_PAP arity sz
376 compile_bind d' fvs x rhs size off = do
377 bco <- schemeR fvs (x,rhs)
378 build_thunk d' fvs size bco off
381 [ compile_bind d' fvs x rhs size n
382 | (fvs, x, rhs, size, n) <-
383 zip5 fvss xs rhss sizes [n_binds, n_binds-1 .. 1]
386 body_code <- schemeE d' s p' body
387 thunk_codes <- sequence compile_binds
388 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
392 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
393 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
395 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
397 -- case .... of a { DEFAULT -> ... }
398 -- becuse the return convention for both are identical.
400 -- Note that it does not matter losing the void-rep thing from the
401 -- envt (it won't be bound now) because we never look such things up.
403 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
404 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
406 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
407 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
408 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
410 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
411 | isUnboxedTupleCon dc
412 -- Similarly, convert
413 -- case .... of x { (# a #) -> ... }
415 -- case .... of a { DEFAULT -> ... }
416 = --trace "automagic mashing of case alts (# a #)" $
417 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
419 schemeE d s p (AnnCase scrut bndr _ alts)
420 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
422 schemeE d s p (AnnNote note (_, body))
426 = pprPanic "ByteCodeGen.schemeE: unhandled case"
427 (pprCoreExpr (deAnnotate' other))
430 -- Compile code to do a tail call. Specifically, push the fn,
431 -- slide the on-stack app back down to the sequel depth,
432 -- and enter. Four cases:
435 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
436 -- The int will be on the stack. Generate a code sequence
437 -- to convert it to the relevant constructor, SLIDE and ENTER.
439 -- 1. The fn denotes a ccall. Defer to generateCCall.
441 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
442 -- it simply as b -- since the representations are identical
443 -- (the VoidArg takes up zero stack space). Also, spot
444 -- (# b #) and treat it as b.
446 -- 3. Application of a constructor, by defn saturated.
447 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
448 -- then the ptrs, and then do PACK and RETURN.
450 -- 4. Otherwise, it must be a function call. Push the args
451 -- right to left, SLIDE and ENTER.
453 schemeT :: Int -- Stack depth
454 -> Sequel -- Sequel depth
455 -> BCEnv -- stack env
456 -> AnnExpr' Id VarSet
461 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
462 -- = panic "schemeT ?!?!"
464 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
468 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
469 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
470 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
471 returnBc (push `appOL` tagToId_sequence
472 `appOL` mkSLIDE 1 (d+arg_words-s)
476 | Just (CCall ccall_spec) <- isFCallId_maybe fn
477 = generateCCall d s p ccall_spec fn args_r_to_l
479 -- Case 2: Constructor application
480 | Just con <- maybe_saturated_dcon,
481 isUnboxedTupleCon con
482 = case args_r_to_l of
483 [arg1,arg2] | isVoidArgAtom arg1 ->
484 unboxedTupleReturn d s p arg2
485 [arg1,arg2] | isVoidArgAtom arg2 ->
486 unboxedTupleReturn d s p arg1
487 _other -> unboxedTupleException
489 -- Case 3: Ordinary data constructor
490 | Just con <- maybe_saturated_dcon
491 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
492 returnBc (alloc_con `appOL`
493 mkSLIDE 1 (d - s) `snocOL`
496 -- Case 4: Tail call of function
498 = doTailCall d s p fn args_r_to_l
501 -- Detect and extract relevant info for the tagToEnum kludge.
502 maybe_is_tagToEnum_call
503 = let extract_constr_Names ty
504 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
506 = map (getName . dataConWorkId) (tyConDataCons tyc)
507 -- NOTE: use the worker name, not the source name of
508 -- the DataCon. See DataCon.lhs for details.
510 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
513 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
514 -> case isPrimOpId_maybe v of
515 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
519 -- Extract the args (R->L) and fn
520 -- The function will necessarily be a variable,
521 -- because we are compiling a tail call
522 (AnnVar fn, args_r_to_l) = splitApp app
524 -- Only consider this to be a constructor application iff it is
525 -- saturated. Otherwise, we'll call the constructor wrapper.
526 n_args = length args_r_to_l
528 = case isDataConWorkId_maybe fn of
529 Just con | dataConRepArity con == n_args -> Just con
532 -- -----------------------------------------------------------------------------
533 -- Generate code to build a constructor application,
534 -- leaving it on top of the stack
536 mkConAppCode :: Int -> Sequel -> BCEnv
537 -> DataCon -- The data constructor
538 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
541 mkConAppCode orig_d s p con [] -- Nullary constructor
542 = ASSERT( isNullaryRepDataCon con )
543 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
544 -- Instead of doing a PACK, which would allocate a fresh
545 -- copy of this constructor, use the single shared version.
547 mkConAppCode orig_d s p con args_r_to_l
548 = ASSERT( dataConRepArity con == length args_r_to_l )
549 do_pushery orig_d (non_ptr_args ++ ptr_args)
551 -- The args are already in reverse order, which is the way PACK
552 -- expects them to be. We must push the non-ptrs after the ptrs.
553 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
555 do_pushery d (arg:args)
556 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
557 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
558 returnBc (push `appOL` more_push_code)
560 = returnBc (unitOL (PACK con n_arg_words))
562 n_arg_words = d - orig_d
565 -- -----------------------------------------------------------------------------
566 -- Returning an unboxed tuple with one non-void component (the only
567 -- case we can handle).
569 -- Remember, we don't want to *evaluate* the component that is being
570 -- returned, even if it is a pointed type. We always just return.
573 :: Int -> Sequel -> BCEnv
574 -> AnnExpr' Id VarSet -> BcM BCInstrList
575 unboxedTupleReturn d s p arg = do
576 (push, sz) <- pushAtom d p arg
577 returnBc (push `appOL`
578 mkSLIDE sz (d-s) `snocOL`
579 RETURN_UBX (atomRep arg))
581 -- -----------------------------------------------------------------------------
582 -- Generate code for a tail-call
585 :: Int -> Sequel -> BCEnv
586 -> Id -> [AnnExpr' Id VarSet]
588 doTailCall init_d s p fn args
589 = do_pushes init_d args (map atomRep args)
591 do_pushes d [] reps = do
592 ASSERT( null reps ) return ()
593 (push_fn, sz) <- pushAtom d p (AnnVar fn)
594 ASSERT( sz == 1 ) return ()
595 returnBc (push_fn `appOL` (
596 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
598 do_pushes d args reps = do
599 let (push_apply, n, rest_of_reps) = findPushSeq reps
600 (these_args, rest_of_args) = splitAt n args
601 (next_d, push_code) <- push_seq d these_args
602 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
603 -- ^^^ for the PUSH_APPLY_ instruction
604 returnBc (push_code `appOL` (push_apply `consOL` instrs))
606 push_seq d [] = return (d, nilOL)
607 push_seq d (arg:args) = do
608 (push_code, sz) <- pushAtom d p arg
609 (final_d, more_push_code) <- push_seq (d+sz) args
610 return (final_d, push_code `appOL` more_push_code)
612 -- v. similar to CgStackery.findMatch, ToDo: merge
613 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
614 = (PUSH_APPLY_PPPPPP, 6, rest)
615 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
616 = (PUSH_APPLY_PPPPP, 5, rest)
617 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
618 = (PUSH_APPLY_PPPP, 4, rest)
619 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
620 = (PUSH_APPLY_PPP, 3, rest)
621 findPushSeq (PtrArg: PtrArg: rest)
622 = (PUSH_APPLY_PP, 2, rest)
623 findPushSeq (PtrArg: rest)
624 = (PUSH_APPLY_P, 1, rest)
625 findPushSeq (VoidArg: rest)
626 = (PUSH_APPLY_V, 1, rest)
627 findPushSeq (NonPtrArg: rest)
628 = (PUSH_APPLY_N, 1, rest)
629 findPushSeq (FloatArg: rest)
630 = (PUSH_APPLY_F, 1, rest)
631 findPushSeq (DoubleArg: rest)
632 = (PUSH_APPLY_D, 1, rest)
633 findPushSeq (LongArg: rest)
634 = (PUSH_APPLY_L, 1, rest)
636 = panic "ByteCodeGen.findPushSeq"
638 -- -----------------------------------------------------------------------------
641 doCase :: Int -> Sequel -> BCEnv
642 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
643 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
645 doCase d s p (_,scrut)
646 bndr alts is_unboxed_tuple
648 -- Top of stack is the return itbl, as usual.
649 -- underneath it is the pointer to the alt_code BCO.
650 -- When an alt is entered, it assumes the returned value is
651 -- on top of the itbl.
654 -- An unlifted value gets an extra info table pushed on top
655 -- when it is returned.
656 unlifted_itbl_sizeW | isAlgCase = 0
659 -- depth of stack after the return value has been pushed
660 d_bndr = d + ret_frame_sizeW + idSizeW bndr
662 -- depth of stack after the extra info table for an unboxed return
663 -- has been pushed, if any. This is the stack depth at the
665 d_alts = d_bndr + unlifted_itbl_sizeW
667 -- Env in which to compile the alts, not including
668 -- any vars bound by the alts themselves
669 p_alts = addToFM p bndr (d_bndr - 1)
671 bndr_ty = idType bndr
672 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
674 -- given an alt, return a discr and code for it.
675 codeALt alt@(DEFAULT, _, (_,rhs))
676 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
677 returnBc (NoDiscr, rhs_code)
678 codeAlt alt@(discr, bndrs, (_,rhs))
679 -- primitive or nullary constructor alt: no need to UNPACK
680 | null real_bndrs = do
681 rhs_code <- schemeE d_alts s p_alts rhs
682 returnBc (my_discr alt, rhs_code)
683 -- algebraic alt with some binders
684 | ASSERT(isAlgCase) otherwise =
686 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
687 ptr_sizes = map idSizeW ptrs
688 nptrs_sizes = map idSizeW nptrs
689 bind_sizes = ptr_sizes ++ nptrs_sizes
690 size = sum ptr_sizes + sum nptrs_sizes
691 -- the UNPACK instruction unpacks in reverse order...
692 p' = addListToFM p_alts
693 (zip (reverse (ptrs ++ nptrs))
694 (mkStackOffsets d_alts (reverse bind_sizes)))
696 rhs_code <- schemeE (d_alts+size) s p' rhs
697 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
699 real_bndrs = filter (not.isTyVar) bndrs
702 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
703 my_discr (DataAlt dc, binds, rhs)
704 | isUnboxedTupleCon dc
705 = unboxedTupleException
707 = DiscrP (dataConTag dc - fIRST_TAG)
708 my_discr (LitAlt l, binds, rhs)
709 = case l of MachInt i -> DiscrI (fromInteger i)
710 MachFloat r -> DiscrF (fromRational r)
711 MachDouble r -> DiscrD (fromRational r)
712 MachChar i -> DiscrI (ord i)
713 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
716 | not isAlgCase = Nothing
718 = case [dc | (DataAlt dc, _, _) <- alts] of
720 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
722 -- the bitmap is relative to stack depth d, i.e. before the
723 -- BCO, info table and return value are pushed on.
724 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
725 -- except that here we build the bitmap from the known bindings of
726 -- things that are pointers, whereas in CgBindery the code builds the
727 -- bitmap from the free slots and unboxed bindings.
729 bitmap = intsToReverseBitmap d{-size-} (sortLe (<=) rel_slots)
732 rel_slots = concat (map spread binds)
734 | isFollowableArg (idCgRep id) = [ rel_offset ]
736 where rel_offset = d - offset - 1
739 alt_stuff <- mapM codeAlt alts
740 alt_final <- mkMultiBranch maybe_ncons alt_stuff
742 alt_bco_name = getName bndr
743 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
744 0{-no arity-} d{-bitmap size-} bitmap True{-is alts-}
746 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
747 -- "\n bitmap = " ++ show bitmap) $ do
748 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
749 alt_bco' <- emitBc alt_bco
751 | isAlgCase = PUSH_ALTS alt_bco'
752 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
753 returnBc (push_alts `consOL` scrut_code)
756 -- -----------------------------------------------------------------------------
757 -- Deal with a CCall.
759 -- Taggedly push the args onto the stack R->L,
760 -- deferencing ForeignObj#s and adjusting addrs to point to
761 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
762 -- (machine) code for the ccall, and create bytecodes to call that and
763 -- then return in the right way.
765 generateCCall :: Int -> Sequel -- stack and sequel depths
767 -> CCallSpec -- where to call
768 -> Id -- of target, for type info
769 -> [AnnExpr' Id VarSet] -- args (atoms)
772 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
775 addr_sizeW = cgRepSizeW NonPtrArg
777 -- Get the args on the stack, with tags and suitably
778 -- dereferenced for the CCall. For each arg, return the
779 -- depth to the first word of the bits for that arg, and the
780 -- CgRep of what was actually pushed.
782 pargs d [] = returnBc []
784 = let arg_ty = repType (exprType (deAnnotate' a))
786 in case splitTyConApp_maybe arg_ty of
787 -- Don't push the FO; instead push the Addr# it
790 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
791 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
792 parg_ArrayishRep arrPtrsHdrSize d p a
794 returnBc ((code,NonPtrArg):rest)
796 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
797 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
798 parg_ArrayishRep arrWordsHdrSize d p a
800 returnBc ((code,NonPtrArg):rest)
802 -- Default case: push taggedly, but otherwise intact.
804 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
805 pargs (d+sz_a) az `thenBc` \ rest ->
806 returnBc ((code_a, atomRep a) : rest)
808 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
809 -- the stack but then advance it over the headers, so as to
810 -- point to the payload.
811 parg_ArrayishRep hdrSize d p a
812 = pushAtom d p a `thenBc` \ (push_fo, _) ->
813 -- The ptr points at the header. Advance it over the
814 -- header and then pretend this is an Addr#.
815 returnBc (push_fo `snocOL` SWIZZLE 0 hdrSize)
818 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
820 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
822 push_args = concatOL pushs_arg
823 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
825 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
826 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
828 = reverse (tail a_reps_pushed_r_to_l)
830 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
831 -- push_args is the code to do that.
832 -- d_after_args is the stack depth once the args are on.
834 -- Get the result rep.
835 (returns_void, r_rep)
836 = case maybe_getCCallReturnRep (idType fn) of
837 Nothing -> (True, VoidArg)
838 Just rr -> (False, rr)
840 Because the Haskell stack grows down, the a_reps refer to
841 lowest to highest addresses in that order. The args for the call
842 are on the stack. Now push an unboxed Addr# indicating
843 the C function to call. Then push a dummy placeholder for the
844 result. Finally, emit a CCALL insn with an offset pointing to the
845 Addr# just pushed, and a literal field holding the mallocville
846 address of the piece of marshalling code we generate.
847 So, just prior to the CCALL insn, the stack looks like this
848 (growing down, as usual):
853 Addr# address_of_C_fn
854 <placeholder-for-result#> (must be an unboxed type)
856 The interpreter then calls the marshall code mentioned
857 in the CCALL insn, passing it (& <placeholder-for-result#>),
858 that is, the addr of the topmost word in the stack.
859 When this returns, the placeholder will have been
860 filled in. The placeholder is slid down to the sequel
861 depth, and we RETURN.
863 This arrangement makes it simple to do f-i-dynamic since the Addr#
864 value is the first arg anyway.
866 The marshalling code is generated specifically for this
867 call site, and so knows exactly the (Haskell) stack
868 offsets of the args, fn address and placeholder. It
869 copies the args to the C stack, calls the stacked addr,
870 and parks the result back in the placeholder. The interpreter
871 calls it as a normal C call, assuming it has a signature
872 void marshall_code ( StgWord* ptr_to_top_of_stack )
874 -- resolve static address
878 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
880 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
883 get_target_info `thenBc` \ (is_static, static_target_addr) ->
886 -- Get the arg reps, zapping the leading Addr# in the dynamic case
887 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
888 | is_static = a_reps_pushed_RAW
889 | otherwise = if null a_reps_pushed_RAW
890 then panic "ByteCodeGen.generateCCall: dyn with no args"
891 else tail a_reps_pushed_RAW
894 (push_Addr, d_after_Addr)
896 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
897 d_after_args + addr_sizeW)
898 | otherwise -- is already on the stack
899 = (nilOL, d_after_args)
901 -- Push the return placeholder. For a call returning nothing,
902 -- this is a VoidArg (tag).
903 r_sizeW = cgRepSizeW r_rep
904 d_after_r = d_after_Addr + r_sizeW
905 r_lit = mkDummyLiteral r_rep
906 push_r = (if returns_void
908 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
910 -- generate the marshalling code we're going to call
913 arg1_offW = r_sizeW + addr_sizeW
914 args_offW = map (arg1_offW +)
915 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
917 ioToBc (mkMarshalCode cconv
918 (r_offW, r_rep) addr_offW
919 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
920 recordMallocBc addr_of_marshaller `thenBc_`
922 -- Offset of the next stack frame down the stack. The CCALL
923 -- instruction needs to describe the chunk of stack containing
924 -- the ccall args to the GC, so it needs to know how large it
925 -- is. See comment in Interpreter.c with the CCALL instruction.
926 stk_offset = d_after_r - s
929 do_call = unitOL (CCALL stk_offset (castPtr addr_of_marshaller))
931 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
932 `snocOL` RETURN_UBX r_rep
934 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
937 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
941 -- Make a dummy literal, to be used as a placeholder for FFI return
942 -- values on the stack.
943 mkDummyLiteral :: CgRep -> Literal
946 NonPtrArg -> MachWord 0
947 DoubleArg -> MachDouble 0
948 FloatArg -> MachFloat 0
949 _ -> moan64 "mkDummyLiteral" (ppr pr)
953 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
954 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
957 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
959 -- Alternatively, for call-targets returning nothing, convert
961 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
962 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
966 maybe_getCCallReturnRep :: Type -> Maybe CgRep
967 maybe_getCCallReturnRep fn_ty
968 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
970 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
972 = case splitTyConApp_maybe (repType r_ty) of
973 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
975 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
976 || r_reps == [VoidArg] )
977 && isUnboxedTupleTyCon r_tycon
978 && case maybe_r_rep_to_go of
980 Just r_rep -> r_rep /= PtrArg
981 -- if it was, it would be impossible
982 -- to create a valid return value
983 -- placeholder on the stack
984 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
987 --trace (showSDoc (ppr (a_reps, r_reps))) $
988 if ok then maybe_r_rep_to_go else blargh
990 -- Compile code which expects an unboxed Int on the top of stack,
991 -- (call it i), and pushes the i'th closure in the supplied list
993 implement_tagToId :: [Name] -> BcM BCInstrList
994 implement_tagToId names
995 = ASSERT( notNull names )
996 getLabelsBc (length names) `thenBc` \ labels ->
997 getLabelBc `thenBc` \ label_fail ->
998 getLabelBc `thenBc` \ label_exit ->
999 zip4 labels (tail labels ++ [label_fail])
1000 [0 ..] names `bind` \ infos ->
1001 map (mkStep label_exit) infos `bind` \ steps ->
1002 returnBc (concatOL steps
1004 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1006 mkStep l_exit (my_label, next_label, n, name_for_n)
1007 = toOL [LABEL my_label,
1008 TESTEQ_I n next_label,
1013 -- -----------------------------------------------------------------------------
1016 -- Push an atom onto the stack, returning suitable code & number of
1017 -- stack words used.
1019 -- The env p must map each variable to the highest- numbered stack
1020 -- slot for it. For example, if the stack has depth 4 and we
1021 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1022 -- the tag in stack[5], the stack will have depth 6, and p must map v
1023 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1024 -- depth 6 stack has valid words 0 .. 5.
1026 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1028 pushAtom d p (AnnApp f (_, AnnType _))
1029 = pushAtom d p (snd f)
1031 pushAtom d p (AnnNote note e)
1032 = pushAtom d p (snd e)
1034 pushAtom d p (AnnLam x e)
1036 = pushAtom d p (snd e)
1038 pushAtom d p (AnnVar v)
1040 | idCgRep v == VoidArg
1041 = returnBc (nilOL, 0)
1044 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1046 | Just primop <- isPrimOpId_maybe v
1047 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1049 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1050 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1051 -- d - d_v the number of words between the TOS
1052 -- and the 1st slot of the object
1054 -- d - d_v - 1 the offset from the TOS of the 1st slot
1056 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1059 -- Having found the last slot, we proceed to copy the right number of
1060 -- slots on to the top of the stack.
1062 | otherwise -- v must be a global variable
1064 returnBc (unitOL (PUSH_G (getName v)), sz)
1070 pushAtom d p (AnnLit lit)
1072 MachLabel fs _ -> code NonPtrArg
1073 MachWord w -> code NonPtrArg
1074 MachInt i -> code PtrArg
1075 MachFloat r -> code FloatArg
1076 MachDouble r -> code DoubleArg
1077 MachChar c -> code NonPtrArg
1078 MachStr s -> pushStr s
1081 = let size_host_words = cgRepSizeW rep
1082 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1086 = let getMallocvilleAddr
1088 FastString _ n _ fp _ ->
1089 -- we could grab the Ptr from the ForeignPtr,
1090 -- but then we have no way to control its lifetime.
1091 -- In reality it'll probably stay alive long enoungh
1092 -- by virtue of the global FastString table, but
1093 -- to be on the safe side we copy the string into
1094 -- a malloc'd area of memory.
1095 ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1096 recordMallocBc ptr `thenBc_`
1098 withForeignPtr fp $ \p -> do
1099 memcpy ptr p (fromIntegral n)
1100 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1103 other -> panic "ByteCodeGen.pushAtom.pushStr"
1105 getMallocvilleAddr `thenBc` \ addr ->
1106 -- Get the addr on the stack, untaggedly
1107 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1110 = pprPanic "ByteCodeGen.pushAtom"
1111 (pprCoreExpr (deAnnotate (undefined, other)))
1113 foreign import ccall unsafe "memcpy"
1114 memcpy :: Ptr a -> Ptr b -> CInt -> IO ()
1117 -- -----------------------------------------------------------------------------
1118 -- Given a bunch of alts code and their discrs, do the donkey work
1119 -- of making a multiway branch using a switch tree.
1120 -- What a load of hassle!
1122 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1123 -- a hint; generates better code
1124 -- Nothing is always safe
1125 -> [(Discr, BCInstrList)]
1127 mkMultiBranch maybe_ncons raw_ways
1128 = let d_way = filter (isNoDiscr.fst) raw_ways
1130 (\w1 w2 -> leAlt (fst w1) (fst w2))
1131 (filter (not.isNoDiscr.fst) raw_ways)
1133 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1134 mkTree [] range_lo range_hi = returnBc the_default
1136 mkTree [val] range_lo range_hi
1137 | range_lo `eqAlt` range_hi
1138 = returnBc (snd val)
1140 = getLabelBc `thenBc` \ label_neq ->
1141 returnBc (mkTestEQ (fst val) label_neq
1143 `appOL` unitOL (LABEL label_neq)
1144 `appOL` the_default))
1146 mkTree vals range_lo range_hi
1147 = let n = length vals `div` 2
1148 vals_lo = take n vals
1149 vals_hi = drop n vals
1150 v_mid = fst (head vals_hi)
1152 getLabelBc `thenBc` \ label_geq ->
1153 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1154 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1155 returnBc (mkTestLT v_mid label_geq
1157 `appOL` unitOL (LABEL label_geq)
1161 = case d_way of [] -> unitOL CASEFAIL
1164 -- None of these will be needed if there are no non-default alts
1165 (mkTestLT, mkTestEQ, init_lo, init_hi)
1167 = panic "mkMultiBranch: awesome foursome"
1169 = case fst (head notd_ways) of {
1170 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1171 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1174 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1175 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1178 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1179 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1182 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1183 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1185 DiscrP algMaxBound )
1188 (algMinBound, algMaxBound)
1189 = case maybe_ncons of
1190 Just n -> (0, n - 1)
1191 Nothing -> (minBound, maxBound)
1193 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1194 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1195 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1196 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1197 NoDiscr `eqAlt` NoDiscr = True
1200 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1201 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1202 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1203 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1204 NoDiscr `leAlt` NoDiscr = True
1207 isNoDiscr NoDiscr = True
1210 dec (DiscrI i) = DiscrI (i-1)
1211 dec (DiscrP i) = DiscrP (i-1)
1212 dec other = other -- not really right, but if you
1213 -- do cases on floating values, you'll get what you deserve
1215 -- same snotty comment applies to the following
1217 minD, maxD :: Double
1223 mkTree notd_ways init_lo init_hi
1226 -- -----------------------------------------------------------------------------
1227 -- Supporting junk for the compilation schemes
1229 -- Describes case alts
1237 instance Outputable Discr where
1238 ppr (DiscrI i) = int i
1239 ppr (DiscrF f) = text (show f)
1240 ppr (DiscrD d) = text (show d)
1241 ppr (DiscrP i) = int i
1242 ppr NoDiscr = text "DEF"
1245 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1246 lookupBCEnv_maybe = lookupFM
1248 idSizeW :: Id -> Int
1249 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1251 unboxedTupleException :: a
1252 unboxedTupleException
1255 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1256 "\tto foreign import/export decls in source. Workaround:\n" ++
1257 "\tcompile this module to a .o file, then restart session."))
1260 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1263 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1264 -- The arguments are returned in *right-to-left* order
1265 splitApp (AnnApp (_,f) (_,a))
1266 | isTypeAtom a = splitApp f
1267 | otherwise = case splitApp f of
1268 (f', as) -> (f', a:as)
1269 splitApp (AnnNote n (_,e)) = splitApp e
1270 splitApp e = (e, [])
1273 isTypeAtom :: AnnExpr' id ann -> Bool
1274 isTypeAtom (AnnType _) = True
1275 isTypeAtom _ = False
1277 isVoidArgAtom :: AnnExpr' id ann -> Bool
1278 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1279 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1280 isVoidArgAtom _ = False
1282 atomRep :: AnnExpr' Id ann -> CgRep
1283 atomRep (AnnVar v) = typeCgRep (idType v)
1284 atomRep (AnnLit l) = typeCgRep (literalType l)
1285 atomRep (AnnNote n b) = atomRep (snd b)
1286 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1287 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1288 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1290 isPtrAtom :: AnnExpr' Id ann -> Bool
1291 isPtrAtom e = atomRep e == PtrArg
1293 -- Let szsw be the sizes in words of some items pushed onto the stack,
1294 -- which has initial depth d'. Return the values which the stack environment
1295 -- should map these items to.
1296 mkStackOffsets :: Int -> [Int] -> [Int]
1297 mkStackOffsets original_depth szsw
1298 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1300 -- -----------------------------------------------------------------------------
1301 -- The bytecode generator's monad
1305 nextlabel :: Int, -- for generating local labels
1306 malloced :: [Ptr ()] } -- ptrs malloced for current BCO
1307 -- Should be free()d when it is GCd
1309 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1311 ioToBc :: IO a -> BcM a
1312 ioToBc io = BcM $ \st -> do
1316 runBc :: BcM r -> IO (BcM_State, r)
1317 runBc (BcM m) = m (BcM_State 0 [])
1319 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1320 thenBc (BcM expr) cont = BcM $ \st0 -> do
1321 (st1, q) <- expr st0
1326 thenBc_ :: BcM a -> BcM b -> BcM b
1327 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1328 (st1, q) <- expr st0
1329 (st2, r) <- cont st1
1332 returnBc :: a -> BcM a
1333 returnBc result = BcM $ \st -> (return (st, result))
1335 instance Monad BcM where
1340 emitBc :: ([Ptr ()] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1342 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1344 recordMallocBc :: Ptr a -> BcM ()
1346 = BcM $ \st -> return (st{malloced = castPtr a : malloced st}, ())
1348 getLabelBc :: BcM Int
1350 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1352 getLabelsBc :: Int -> BcM [Int]
1354 = BcM $ \st -> let ctr = nextlabel st
1355 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])