2 % (c) The University of Glasgow 2002-2006
5 ByteCodeGen: Generate bytecode from Core
8 module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where
10 #include "HsVersions.h"
50 import Data.List ( intersperse, sortBy, zip4, zip6, partition )
51 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8,
52 withForeignPtr, castFunPtrToPtr )
53 import Foreign.C ( CInt )
54 import Control.Exception ( throwDyn )
56 import GHC.Exts ( Int(..), ByteArray# )
58 import Control.Monad ( when )
59 import Data.Char ( ord, chr )
61 -- -----------------------------------------------------------------------------
62 -- Generating byte code for a complete module
64 byteCodeGen :: DynFlags
67 -> IO CompiledByteCode
68 byteCodeGen dflags binds tycs
69 = do showPass dflags "ByteCodeGen"
71 let flatBinds = [ (bndr, freeVars rhs)
72 | (bndr, rhs) <- flattenBinds binds]
74 (BcM_State final_ctr mallocd, proto_bcos)
75 <- runBc (mapM schemeTopBind flatBinds)
77 when (notNull mallocd)
78 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
80 dumpIfSet_dyn dflags Opt_D_dump_BCOs
81 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
83 assembleBCOs proto_bcos tycs
85 -- -----------------------------------------------------------------------------
86 -- Generating byte code for an expression
88 -- Returns: (the root BCO for this expression,
89 -- a list of auxilary BCOs resulting from compiling closures)
90 coreExprToBCOs :: DynFlags
93 coreExprToBCOs dflags expr
94 = do showPass dflags "ByteCodeGen"
96 -- create a totally bogus name for the top-level BCO; this
97 -- should be harmless, since it's never used for anything
98 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
99 invented_id = Id.mkLocalId invented_name (panic "invented_id's type")
101 (BcM_State final_ctr mallocd, proto_bco)
102 <- runBc (schemeTopBind (invented_id, freeVars expr))
104 when (notNull mallocd)
105 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
107 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
109 assembleBCO proto_bco
112 -- -----------------------------------------------------------------------------
113 -- Compilation schema for the bytecode generator
115 type BCInstrList = OrdList BCInstr
117 type Sequel = Int -- back off to this depth before ENTER
119 -- Maps Ids to the offset from the stack _base_ so we don't have
120 -- to mess with it after each push/pop.
121 type BCEnv = FiniteMap Id Int -- To find vars on the stack
123 ppBCEnv :: BCEnv -> SDoc
126 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
129 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
130 cmp_snd x y = compare (snd x) (snd y)
132 -- Create a BCO and do a spot of peephole optimisation on the insns
137 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
141 -> Bool -- True <=> is a return point, rather than a function
144 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap
145 is_ret mallocd_blocks
148 protoBCOInstrs = maybe_with_stack_check,
149 protoBCOBitmap = bitmap,
150 protoBCOBitmapSize = bitmap_size,
151 protoBCOArity = arity,
152 protoBCOExpr = origin,
153 protoBCOPtrs = mallocd_blocks
156 -- Overestimate the stack usage (in words) of this BCO,
157 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
158 -- stack check. (The interpreter always does a stack check
159 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
160 -- BCO anyway, so we only need to add an explicit on in the
161 -- (hopefully rare) cases when the (overestimated) stack use
162 -- exceeds iNTERP_STACK_CHECK_THRESH.
163 maybe_with_stack_check
165 -- don't do stack checks at return points;
166 -- everything is aggregated up to the top BCO
167 -- (which must be a function)
168 | stack_overest >= iNTERP_STACK_CHECK_THRESH
169 = STKCHECK stack_overest : peep_d
171 = peep_d -- the supposedly common case
173 -- We assume that this sum doesn't wrap
174 stack_overest = sum (map bciStackUse peep_d)
176 -- Merge local pushes
177 peep_d = peep (fromOL instrs_ordlist)
179 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
180 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
181 peep (PUSH_L off1 : PUSH_L off2 : rest)
182 = PUSH_LL off1 (off2-1) : peep rest
188 argBits :: [CgRep] -> [Bool]
191 | isFollowableArg rep = False : argBits args
192 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
194 -- -----------------------------------------------------------------------------
197 -- Compile code for the right-hand side of a top-level binding
199 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
202 schemeTopBind (id, rhs)
203 | Just data_con <- isDataConWorkId_maybe id,
204 isNullaryRepDataCon data_con
205 = -- Special case for the worker of a nullary data con.
206 -- It'll look like this: Nil = /\a -> Nil a
207 -- If we feed it into schemeR, we'll get
209 -- because mkConAppCode treats nullary constructor applications
210 -- by just re-using the single top-level definition. So
211 -- for the worker itself, we must allocate it directly.
212 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
213 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
216 = schemeR [{- No free variables -}] (id, rhs)
218 -- -----------------------------------------------------------------------------
221 -- Compile code for a right-hand side, to give a BCO that,
222 -- when executed with the free variables and arguments on top of the stack,
223 -- will return with a pointer to the result on top of the stack, after
224 -- removing the free variables and arguments.
226 -- Park the resulting BCO in the monad. Also requires the
227 -- variable to which this value was bound, so as to give the
228 -- resulting BCO a name.
230 schemeR :: [Id] -- Free vars of the RHS, ordered as they
231 -- will appear in the thunk. Empty for
232 -- top-level things, which have no free vars.
233 -> (Id, AnnExpr Id VarSet)
234 -> BcM (ProtoBCO Name)
235 schemeR fvs (nm, rhs)
239 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
240 $$ pprCoreExpr (deAnnotate rhs)
246 = schemeR_wrk fvs nm rhs (collect [] rhs)
248 collect xs (_, AnnNote note e) = collect xs e
249 collect xs (_, AnnCast e _) = collect xs e
250 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
251 collect xs (_, not_lambda) = (reverse xs, not_lambda)
253 schemeR_wrk fvs nm original_body (args, body)
255 all_args = reverse args ++ fvs
256 arity = length all_args
257 -- all_args are the args in reverse order. We're compiling a function
258 -- \fv1..fvn x1..xn -> e
259 -- i.e. the fvs come first
261 szsw_args = map idSizeW all_args
262 szw_args = sum szsw_args
263 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
265 -- make the arg bitmap
266 bits = argBits (reverse (map idCgRep all_args))
267 bitmap_size = length bits
268 bitmap = mkBitmap bits
270 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
271 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
272 arity bitmap_size bitmap False{-not alts-})
275 fvsToEnv :: BCEnv -> VarSet -> [Id]
276 -- Takes the free variables of a right-hand side, and
277 -- delivers an ordered list of the local variables that will
278 -- be captured in the thunk for the RHS
279 -- The BCEnv argument tells which variables are in the local
280 -- environment: these are the ones that should be captured
282 -- The code that constructs the thunk, and the code that executes
283 -- it, have to agree about this layout
284 fvsToEnv p fvs = [v | v <- varSetElems fvs,
285 isId v, -- Could be a type variable
288 -- -----------------------------------------------------------------------------
291 -- Compile code to apply the given expression to the remaining args
292 -- on the stack, returning a HNF.
293 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
295 -- Delegate tail-calls to schemeT.
296 schemeE d s p e@(AnnApp f a)
299 schemeE d s p e@(AnnVar v)
300 | not (isUnLiftedType v_type)
301 = -- Lifted-type thing; push it in the normal way
305 = -- Returning an unlifted value.
306 -- Heave it on the stack, SLIDE, and RETURN.
307 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
308 returnBc (push -- value onto stack
309 `appOL` mkSLIDE szw (d-s) -- clear to sequel
310 `snocOL` RETURN_UBX v_rep) -- go
313 v_rep = typeCgRep v_type
315 schemeE d s p (AnnLit literal)
316 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
317 let l_rep = typeCgRep (literalType literal)
318 in returnBc (push -- value onto stack
319 `appOL` mkSLIDE szw (d-s) -- clear to sequel
320 `snocOL` RETURN_UBX l_rep) -- go
323 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
324 | (AnnVar v, args_r_to_l) <- splitApp rhs,
325 Just data_con <- isDataConWorkId_maybe v,
326 dataConRepArity data_con == length args_r_to_l
327 = -- Special case for a non-recursive let whose RHS is a
328 -- saturatred constructor application.
329 -- Just allocate the constructor and carry on
330 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
331 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
332 returnBc (alloc_code `appOL` body_code)
334 -- General case for let. Generates correct, if inefficient, code in
336 schemeE d s p (AnnLet binds (_,body))
337 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
338 AnnRec xs_n_rhss -> unzip xs_n_rhss
341 fvss = map (fvsToEnv p' . fst) rhss
343 -- Sizes of free vars
344 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
346 -- the arity of each rhs
347 arities = map (length . fst . collect []) rhss
349 -- This p', d' defn is safe because all the items being pushed
350 -- are ptrs, so all have size 1. d' and p' reflect the stack
351 -- after the closures have been allocated in the heap (but not
352 -- filled in), and pointers to them parked on the stack.
353 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
355 zipE = zipEqual "schemeE"
357 -- ToDo: don't build thunks for things with no free variables
358 build_thunk dd [] size bco off arity
359 = returnBc (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
361 mkap | arity == 0 = MKAP
363 build_thunk dd (fv:fvs) size bco off arity = do
364 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
365 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
366 returnBc (push_code `appOL` more_push_code)
368 alloc_code = toOL (zipWith mkAlloc sizes arities)
369 where mkAlloc sz 0 = ALLOC_AP sz
370 mkAlloc sz arity = ALLOC_PAP arity sz
372 compile_bind d' fvs x rhs size arity off = do
373 bco <- schemeR fvs (x,rhs)
374 build_thunk d' fvs size bco off arity
377 [ compile_bind d' fvs x rhs size arity n
378 | (fvs, x, rhs, size, arity, n) <-
379 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
382 body_code <- schemeE d' s p' body
383 thunk_codes <- sequence compile_binds
384 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
388 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
389 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
391 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
393 -- case .... of a { DEFAULT -> ... }
394 -- becuse the return convention for both are identical.
396 -- Note that it does not matter losing the void-rep thing from the
397 -- envt (it won't be bound now) because we never look such things up.
399 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
400 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
402 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
403 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
404 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
406 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
407 | isUnboxedTupleCon dc
408 -- Similarly, convert
409 -- case .... of x { (# a #) -> ... }
411 -- case .... of a { DEFAULT -> ... }
412 = --trace "automagic mashing of case alts (# a #)" $
413 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
415 schemeE d s p (AnnCase scrut bndr _ alts)
416 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
418 schemeE d s p (AnnNote note (_, body))
421 schemeE d s p (AnnCast (_, body) _)
425 = pprPanic "ByteCodeGen.schemeE: unhandled case"
426 (pprCoreExpr (deAnnotate' other))
429 -- Compile code to do a tail call. Specifically, push the fn,
430 -- slide the on-stack app back down to the sequel depth,
431 -- and enter. Four cases:
434 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
435 -- The int will be on the stack. Generate a code sequence
436 -- to convert it to the relevant constructor, SLIDE and ENTER.
438 -- 1. The fn denotes a ccall. Defer to generateCCall.
440 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
441 -- it simply as b -- since the representations are identical
442 -- (the VoidArg takes up zero stack space). Also, spot
443 -- (# b #) and treat it as b.
445 -- 3. Application of a constructor, by defn saturated.
446 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
447 -- then the ptrs, and then do PACK and RETURN.
449 -- 4. Otherwise, it must be a function call. Push the args
450 -- right to left, SLIDE and ENTER.
452 schemeT :: Int -- Stack depth
453 -> Sequel -- Sequel depth
454 -> BCEnv -- stack env
455 -> AnnExpr' Id VarSet
460 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
461 -- = panic "schemeT ?!?!"
463 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
467 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
468 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
469 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
470 returnBc (push `appOL` tagToId_sequence
471 `appOL` mkSLIDE 1 (d+arg_words-s)
475 | Just (CCall ccall_spec) <- isFCallId_maybe fn
476 = generateCCall d s p ccall_spec fn args_r_to_l
478 -- Case 2: Constructor application
479 | Just con <- maybe_saturated_dcon,
480 isUnboxedTupleCon con
481 = case args_r_to_l of
482 [arg1,arg2] | isVoidArgAtom arg1 ->
483 unboxedTupleReturn d s p arg2
484 [arg1,arg2] | isVoidArgAtom arg2 ->
485 unboxedTupleReturn d s p arg1
486 _other -> unboxedTupleException
488 -- Case 3: Ordinary data constructor
489 | Just con <- maybe_saturated_dcon
490 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
491 returnBc (alloc_con `appOL`
492 mkSLIDE 1 (d - s) `snocOL`
495 -- Case 4: Tail call of function
497 = doTailCall d s p fn args_r_to_l
500 -- Detect and extract relevant info for the tagToEnum kludge.
501 maybe_is_tagToEnum_call
502 = let extract_constr_Names ty
503 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
505 = map (getName . dataConWorkId) (tyConDataCons tyc)
506 -- NOTE: use the worker name, not the source name of
507 -- the DataCon. See DataCon.lhs for details.
509 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
512 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
513 -> case isPrimOpId_maybe v of
514 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
518 -- Extract the args (R->L) and fn
519 -- The function will necessarily be a variable,
520 -- because we are compiling a tail call
521 (AnnVar fn, args_r_to_l) = splitApp app
523 -- Only consider this to be a constructor application iff it is
524 -- saturated. Otherwise, we'll call the constructor wrapper.
525 n_args = length args_r_to_l
527 = case isDataConWorkId_maybe fn of
528 Just con | dataConRepArity con == n_args -> Just con
531 -- -----------------------------------------------------------------------------
532 -- Generate code to build a constructor application,
533 -- leaving it on top of the stack
535 mkConAppCode :: Int -> Sequel -> BCEnv
536 -> DataCon -- The data constructor
537 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
540 mkConAppCode orig_d s p con [] -- Nullary constructor
541 = ASSERT( isNullaryRepDataCon con )
542 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
543 -- Instead of doing a PACK, which would allocate a fresh
544 -- copy of this constructor, use the single shared version.
546 mkConAppCode orig_d s p con args_r_to_l
547 = ASSERT( dataConRepArity con == length args_r_to_l )
548 do_pushery orig_d (non_ptr_args ++ ptr_args)
550 -- The args are already in reverse order, which is the way PACK
551 -- expects them to be. We must push the non-ptrs after the ptrs.
552 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
554 do_pushery d (arg:args)
555 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
556 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
557 returnBc (push `appOL` more_push_code)
559 = returnBc (unitOL (PACK con n_arg_words))
561 n_arg_words = d - orig_d
564 -- -----------------------------------------------------------------------------
565 -- Returning an unboxed tuple with one non-void component (the only
566 -- case we can handle).
568 -- Remember, we don't want to *evaluate* the component that is being
569 -- returned, even if it is a pointed type. We always just return.
572 :: Int -> Sequel -> BCEnv
573 -> AnnExpr' Id VarSet -> BcM BCInstrList
574 unboxedTupleReturn d s p arg = do
575 (push, sz) <- pushAtom d p arg
576 returnBc (push `appOL`
577 mkSLIDE sz (d-s) `snocOL`
578 RETURN_UBX (atomRep arg))
580 -- -----------------------------------------------------------------------------
581 -- Generate code for a tail-call
584 :: Int -> Sequel -> BCEnv
585 -> Id -> [AnnExpr' Id VarSet]
587 doTailCall init_d s p fn args
588 = do_pushes init_d args (map atomRep args)
590 do_pushes d [] reps = do
591 ASSERT( null reps ) return ()
592 (push_fn, sz) <- pushAtom d p (AnnVar fn)
593 ASSERT( sz == 1 ) return ()
594 returnBc (push_fn `appOL` (
595 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
597 do_pushes d args reps = do
598 let (push_apply, n, rest_of_reps) = findPushSeq reps
599 (these_args, rest_of_args) = splitAt n args
600 (next_d, push_code) <- push_seq d these_args
601 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
602 -- ^^^ for the PUSH_APPLY_ instruction
603 returnBc (push_code `appOL` (push_apply `consOL` instrs))
605 push_seq d [] = return (d, nilOL)
606 push_seq d (arg:args) = do
607 (push_code, sz) <- pushAtom d p arg
608 (final_d, more_push_code) <- push_seq (d+sz) args
609 return (final_d, push_code `appOL` more_push_code)
611 -- v. similar to CgStackery.findMatch, ToDo: merge
612 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
613 = (PUSH_APPLY_PPPPPP, 6, rest)
614 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
615 = (PUSH_APPLY_PPPPP, 5, rest)
616 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
617 = (PUSH_APPLY_PPPP, 4, rest)
618 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
619 = (PUSH_APPLY_PPP, 3, rest)
620 findPushSeq (PtrArg: PtrArg: rest)
621 = (PUSH_APPLY_PP, 2, rest)
622 findPushSeq (PtrArg: rest)
623 = (PUSH_APPLY_P, 1, rest)
624 findPushSeq (VoidArg: rest)
625 = (PUSH_APPLY_V, 1, rest)
626 findPushSeq (NonPtrArg: rest)
627 = (PUSH_APPLY_N, 1, rest)
628 findPushSeq (FloatArg: rest)
629 = (PUSH_APPLY_F, 1, rest)
630 findPushSeq (DoubleArg: rest)
631 = (PUSH_APPLY_D, 1, rest)
632 findPushSeq (LongArg: rest)
633 = (PUSH_APPLY_L, 1, rest)
635 = panic "ByteCodeGen.findPushSeq"
637 -- -----------------------------------------------------------------------------
640 doCase :: Int -> Sequel -> BCEnv
641 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
642 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
644 doCase d s p (_,scrut)
645 bndr alts is_unboxed_tuple
647 -- Top of stack is the return itbl, as usual.
648 -- underneath it is the pointer to the alt_code BCO.
649 -- When an alt is entered, it assumes the returned value is
650 -- on top of the itbl.
653 -- An unlifted value gets an extra info table pushed on top
654 -- when it is returned.
655 unlifted_itbl_sizeW | isAlgCase = 0
658 -- depth of stack after the return value has been pushed
659 d_bndr = d + ret_frame_sizeW + idSizeW bndr
661 -- depth of stack after the extra info table for an unboxed return
662 -- has been pushed, if any. This is the stack depth at the
664 d_alts = d_bndr + unlifted_itbl_sizeW
666 -- Env in which to compile the alts, not including
667 -- any vars bound by the alts themselves
668 p_alts = addToFM p bndr (d_bndr - 1)
670 bndr_ty = idType bndr
671 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
673 -- given an alt, return a discr and code for it.
674 codeALt alt@(DEFAULT, _, (_,rhs))
675 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
676 returnBc (NoDiscr, rhs_code)
677 codeAlt alt@(discr, bndrs, (_,rhs))
678 -- primitive or nullary constructor alt: no need to UNPACK
679 | null real_bndrs = do
680 rhs_code <- schemeE d_alts s p_alts rhs
681 returnBc (my_discr alt, rhs_code)
682 -- algebraic alt with some binders
683 | ASSERT(isAlgCase) otherwise =
685 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
686 ptr_sizes = map idSizeW ptrs
687 nptrs_sizes = map idSizeW nptrs
688 bind_sizes = ptr_sizes ++ nptrs_sizes
689 size = sum ptr_sizes + sum nptrs_sizes
690 -- the UNPACK instruction unpacks in reverse order...
691 p' = addListToFM p_alts
692 (zip (reverse (ptrs ++ nptrs))
693 (mkStackOffsets d_alts (reverse bind_sizes)))
695 rhs_code <- schemeE (d_alts+size) s p' rhs
696 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
698 real_bndrs = filter (not.isTyVar) bndrs
701 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
702 my_discr (DataAlt dc, binds, rhs)
703 | isUnboxedTupleCon dc
704 = unboxedTupleException
706 = DiscrP (dataConTag dc - fIRST_TAG)
707 my_discr (LitAlt l, binds, rhs)
708 = case l of MachInt i -> DiscrI (fromInteger i)
709 MachFloat r -> DiscrF (fromRational r)
710 MachDouble r -> DiscrD (fromRational r)
711 MachChar i -> DiscrI (ord i)
712 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
715 | not isAlgCase = Nothing
717 = case [dc | (DataAlt dc, _, _) <- alts] of
719 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
721 -- the bitmap is relative to stack depth d, i.e. before the
722 -- BCO, info table and return value are pushed on.
723 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
724 -- except that here we build the bitmap from the known bindings of
725 -- things that are pointers, whereas in CgBindery the code builds the
726 -- bitmap from the free slots and unboxed bindings.
729 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
730 -- The bitmap must cover the portion of the stack up to the sequel only.
731 -- Previously we were building a bitmap for the whole depth (d), but we
732 -- really want a bitmap up to depth (d-s). This affects compilation of
733 -- case-of-case expressions, which is the only time we can be compiling a
734 -- case expression with s /= 0.
736 bitmap = intsToReverseBitmap bitmap_size{-size-}
737 (sortLe (<=) (filter (< bitmap_size) rel_slots))
740 rel_slots = concat (map spread binds)
742 | isFollowableArg (idCgRep id) = [ rel_offset ]
744 where rel_offset = d - offset - 1
747 alt_stuff <- mapM codeAlt alts
748 alt_final <- mkMultiBranch maybe_ncons alt_stuff
750 alt_bco_name = getName bndr
751 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
752 0{-no arity-} bitmap_size bitmap True{-is alts-}
754 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
755 -- "\n bitmap = " ++ show bitmap) $ do
756 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
757 alt_bco' <- emitBc alt_bco
759 | isAlgCase = PUSH_ALTS alt_bco'
760 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
761 returnBc (push_alts `consOL` scrut_code)
764 -- -----------------------------------------------------------------------------
765 -- Deal with a CCall.
767 -- Taggedly push the args onto the stack R->L,
768 -- deferencing ForeignObj#s and adjusting addrs to point to
769 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
770 -- (machine) code for the ccall, and create bytecodes to call that and
771 -- then return in the right way.
773 generateCCall :: Int -> Sequel -- stack and sequel depths
775 -> CCallSpec -- where to call
776 -> Id -- of target, for type info
777 -> [AnnExpr' Id VarSet] -- args (atoms)
780 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
783 addr_sizeW = cgRepSizeW NonPtrArg
785 -- Get the args on the stack, with tags and suitably
786 -- dereferenced for the CCall. For each arg, return the
787 -- depth to the first word of the bits for that arg, and the
788 -- CgRep of what was actually pushed.
790 pargs d [] = returnBc []
792 = let arg_ty = repType (exprType (deAnnotate' a))
794 in case splitTyConApp_maybe arg_ty of
795 -- Don't push the FO; instead push the Addr# it
798 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
799 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
800 parg_ArrayishRep arrPtrsHdrSize d p a
802 returnBc ((code,NonPtrArg):rest)
804 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
805 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
806 parg_ArrayishRep arrWordsHdrSize d p a
808 returnBc ((code,NonPtrArg):rest)
810 -- Default case: push taggedly, but otherwise intact.
812 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
813 pargs (d+sz_a) az `thenBc` \ rest ->
814 returnBc ((code_a, atomRep a) : rest)
816 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
817 -- the stack but then advance it over the headers, so as to
818 -- point to the payload.
819 parg_ArrayishRep hdrSize d p a
820 = pushAtom d p a `thenBc` \ (push_fo, _) ->
821 -- The ptr points at the header. Advance it over the
822 -- header and then pretend this is an Addr#.
823 returnBc (push_fo `snocOL` SWIZZLE 0 hdrSize)
826 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
828 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
830 push_args = concatOL pushs_arg
831 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
833 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
834 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
836 = reverse (tail a_reps_pushed_r_to_l)
838 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
839 -- push_args is the code to do that.
840 -- d_after_args is the stack depth once the args are on.
842 -- Get the result rep.
843 (returns_void, r_rep)
844 = case maybe_getCCallReturnRep (idType fn) of
845 Nothing -> (True, VoidArg)
846 Just rr -> (False, rr)
848 Because the Haskell stack grows down, the a_reps refer to
849 lowest to highest addresses in that order. The args for the call
850 are on the stack. Now push an unboxed Addr# indicating
851 the C function to call. Then push a dummy placeholder for the
852 result. Finally, emit a CCALL insn with an offset pointing to the
853 Addr# just pushed, and a literal field holding the mallocville
854 address of the piece of marshalling code we generate.
855 So, just prior to the CCALL insn, the stack looks like this
856 (growing down, as usual):
861 Addr# address_of_C_fn
862 <placeholder-for-result#> (must be an unboxed type)
864 The interpreter then calls the marshall code mentioned
865 in the CCALL insn, passing it (& <placeholder-for-result#>),
866 that is, the addr of the topmost word in the stack.
867 When this returns, the placeholder will have been
868 filled in. The placeholder is slid down to the sequel
869 depth, and we RETURN.
871 This arrangement makes it simple to do f-i-dynamic since the Addr#
872 value is the first arg anyway.
874 The marshalling code is generated specifically for this
875 call site, and so knows exactly the (Haskell) stack
876 offsets of the args, fn address and placeholder. It
877 copies the args to the C stack, calls the stacked addr,
878 and parks the result back in the placeholder. The interpreter
879 calls it as a normal C call, assuming it has a signature
880 void marshall_code ( StgWord* ptr_to_top_of_stack )
882 -- resolve static address
886 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
888 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
891 get_target_info `thenBc` \ (is_static, static_target_addr) ->
894 -- Get the arg reps, zapping the leading Addr# in the dynamic case
895 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
896 | is_static = a_reps_pushed_RAW
897 | otherwise = if null a_reps_pushed_RAW
898 then panic "ByteCodeGen.generateCCall: dyn with no args"
899 else tail a_reps_pushed_RAW
902 (push_Addr, d_after_Addr)
904 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
905 d_after_args + addr_sizeW)
906 | otherwise -- is already on the stack
907 = (nilOL, d_after_args)
909 -- Push the return placeholder. For a call returning nothing,
910 -- this is a VoidArg (tag).
911 r_sizeW = cgRepSizeW r_rep
912 d_after_r = d_after_Addr + r_sizeW
913 r_lit = mkDummyLiteral r_rep
914 push_r = (if returns_void
916 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
918 -- generate the marshalling code we're going to call
921 arg1_offW = r_sizeW + addr_sizeW
922 args_offW = map (arg1_offW +)
923 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
925 ioToBc (mkMarshalCode cconv
926 (r_offW, r_rep) addr_offW
927 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
928 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller)) `thenBc_`
930 -- Offset of the next stack frame down the stack. The CCALL
931 -- instruction needs to describe the chunk of stack containing
932 -- the ccall args to the GC, so it needs to know how large it
933 -- is. See comment in Interpreter.c with the CCALL instruction.
934 stk_offset = d_after_r - s
937 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
939 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
940 `snocOL` RETURN_UBX r_rep
942 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
945 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
949 -- Make a dummy literal, to be used as a placeholder for FFI return
950 -- values on the stack.
951 mkDummyLiteral :: CgRep -> Literal
954 NonPtrArg -> MachWord 0
955 DoubleArg -> MachDouble 0
956 FloatArg -> MachFloat 0
957 LongArg -> MachWord64 0
958 _ -> moan64 "mkDummyLiteral" (ppr pr)
962 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
963 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
966 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
968 -- Alternatively, for call-targets returning nothing, convert
970 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
971 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
975 maybe_getCCallReturnRep :: Type -> Maybe CgRep
976 maybe_getCCallReturnRep fn_ty
977 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
979 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
981 = case splitTyConApp_maybe (repType r_ty) of
982 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
984 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
985 || r_reps == [VoidArg] )
986 && isUnboxedTupleTyCon r_tycon
987 && case maybe_r_rep_to_go of
989 Just r_rep -> r_rep /= PtrArg
990 -- if it was, it would be impossible
991 -- to create a valid return value
992 -- placeholder on the stack
993 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
996 --trace (showSDoc (ppr (a_reps, r_reps))) $
997 if ok then maybe_r_rep_to_go else blargh
999 -- Compile code which expects an unboxed Int on the top of stack,
1000 -- (call it i), and pushes the i'th closure in the supplied list
1001 -- as a consequence.
1002 implement_tagToId :: [Name] -> BcM BCInstrList
1003 implement_tagToId names
1004 = ASSERT( notNull names )
1005 getLabelsBc (length names) `thenBc` \ labels ->
1006 getLabelBc `thenBc` \ label_fail ->
1007 getLabelBc `thenBc` \ label_exit ->
1008 zip4 labels (tail labels ++ [label_fail])
1009 [0 ..] names `bind` \ infos ->
1010 map (mkStep label_exit) infos `bind` \ steps ->
1011 returnBc (concatOL steps
1013 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1015 mkStep l_exit (my_label, next_label, n, name_for_n)
1016 = toOL [LABEL my_label,
1017 TESTEQ_I n next_label,
1022 -- -----------------------------------------------------------------------------
1025 -- Push an atom onto the stack, returning suitable code & number of
1026 -- stack words used.
1028 -- The env p must map each variable to the highest- numbered stack
1029 -- slot for it. For example, if the stack has depth 4 and we
1030 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1031 -- the tag in stack[5], the stack will have depth 6, and p must map v
1032 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1033 -- depth 6 stack has valid words 0 .. 5.
1035 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1037 pushAtom d p (AnnApp f (_, AnnType _))
1038 = pushAtom d p (snd f)
1040 pushAtom d p (AnnNote note e)
1041 = pushAtom d p (snd e)
1043 pushAtom d p (AnnLam x e)
1045 = pushAtom d p (snd e)
1047 pushAtom d p (AnnVar v)
1049 | idCgRep v == VoidArg
1050 = returnBc (nilOL, 0)
1053 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1055 | Just primop <- isPrimOpId_maybe v
1056 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1058 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1059 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1060 -- d - d_v the number of words between the TOS
1061 -- and the 1st slot of the object
1063 -- d - d_v - 1 the offset from the TOS of the 1st slot
1065 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1068 -- Having found the last slot, we proceed to copy the right number of
1069 -- slots on to the top of the stack.
1071 | otherwise -- v must be a global variable
1073 returnBc (unitOL (PUSH_G (getName v)), sz)
1079 pushAtom d p (AnnLit lit)
1081 MachLabel fs _ -> code NonPtrArg
1082 MachWord w -> code NonPtrArg
1083 MachInt i -> code PtrArg
1084 MachFloat r -> code FloatArg
1085 MachDouble r -> code DoubleArg
1086 MachChar c -> code NonPtrArg
1087 MachStr s -> pushStr s
1090 = let size_host_words = cgRepSizeW rep
1091 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1095 = let getMallocvilleAddr
1097 FastString _ n _ fp _ ->
1098 -- we could grab the Ptr from the ForeignPtr,
1099 -- but then we have no way to control its lifetime.
1100 -- In reality it'll probably stay alive long enoungh
1101 -- by virtue of the global FastString table, but
1102 -- to be on the safe side we copy the string into
1103 -- a malloc'd area of memory.
1104 ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1105 recordMallocBc ptr `thenBc_`
1107 withForeignPtr fp $ \p -> do
1108 memcpy ptr p (fromIntegral n)
1109 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1113 getMallocvilleAddr `thenBc` \ addr ->
1114 -- Get the addr on the stack, untaggedly
1115 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1117 pushAtom d p (AnnCast e _)
1118 = pushAtom d p (snd e)
1121 = pprPanic "ByteCodeGen.pushAtom"
1122 (pprCoreExpr (deAnnotate (undefined, other)))
1124 foreign import ccall unsafe "memcpy"
1125 memcpy :: Ptr a -> Ptr b -> CInt -> IO ()
1128 -- -----------------------------------------------------------------------------
1129 -- Given a bunch of alts code and their discrs, do the donkey work
1130 -- of making a multiway branch using a switch tree.
1131 -- What a load of hassle!
1133 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1134 -- a hint; generates better code
1135 -- Nothing is always safe
1136 -> [(Discr, BCInstrList)]
1138 mkMultiBranch maybe_ncons raw_ways
1139 = let d_way = filter (isNoDiscr.fst) raw_ways
1141 (\w1 w2 -> leAlt (fst w1) (fst w2))
1142 (filter (not.isNoDiscr.fst) raw_ways)
1144 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1145 mkTree [] range_lo range_hi = returnBc the_default
1147 mkTree [val] range_lo range_hi
1148 | range_lo `eqAlt` range_hi
1149 = returnBc (snd val)
1151 = getLabelBc `thenBc` \ label_neq ->
1152 returnBc (mkTestEQ (fst val) label_neq
1154 `appOL` unitOL (LABEL label_neq)
1155 `appOL` the_default))
1157 mkTree vals range_lo range_hi
1158 = let n = length vals `div` 2
1159 vals_lo = take n vals
1160 vals_hi = drop n vals
1161 v_mid = fst (head vals_hi)
1163 getLabelBc `thenBc` \ label_geq ->
1164 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1165 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1166 returnBc (mkTestLT v_mid label_geq
1168 `appOL` unitOL (LABEL label_geq)
1172 = case d_way of [] -> unitOL CASEFAIL
1175 -- None of these will be needed if there are no non-default alts
1176 (mkTestLT, mkTestEQ, init_lo, init_hi)
1178 = panic "mkMultiBranch: awesome foursome"
1180 = case fst (head notd_ways) of {
1181 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1182 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1185 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1186 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1189 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1190 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1193 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1194 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1196 DiscrP algMaxBound )
1199 (algMinBound, algMaxBound)
1200 = case maybe_ncons of
1201 Just n -> (0, n - 1)
1202 Nothing -> (minBound, maxBound)
1204 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1205 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1206 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1207 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1208 NoDiscr `eqAlt` NoDiscr = True
1211 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1212 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1213 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1214 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1215 NoDiscr `leAlt` NoDiscr = True
1218 isNoDiscr NoDiscr = True
1221 dec (DiscrI i) = DiscrI (i-1)
1222 dec (DiscrP i) = DiscrP (i-1)
1223 dec other = other -- not really right, but if you
1224 -- do cases on floating values, you'll get what you deserve
1226 -- same snotty comment applies to the following
1228 minD, maxD :: Double
1234 mkTree notd_ways init_lo init_hi
1237 -- -----------------------------------------------------------------------------
1238 -- Supporting junk for the compilation schemes
1240 -- Describes case alts
1248 instance Outputable Discr where
1249 ppr (DiscrI i) = int i
1250 ppr (DiscrF f) = text (show f)
1251 ppr (DiscrD d) = text (show d)
1252 ppr (DiscrP i) = int i
1253 ppr NoDiscr = text "DEF"
1256 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1257 lookupBCEnv_maybe = lookupFM
1259 idSizeW :: Id -> Int
1260 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1262 unboxedTupleException :: a
1263 unboxedTupleException
1266 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1267 "\tto foreign import/export decls in source. Workaround:\n" ++
1268 "\tcompile this module to a .o file, then restart session."))
1271 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1274 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1275 -- The arguments are returned in *right-to-left* order
1276 splitApp (AnnApp (_,f) (_,a))
1277 | isTypeAtom a = splitApp f
1278 | otherwise = case splitApp f of
1279 (f', as) -> (f', a:as)
1280 splitApp (AnnNote n (_,e)) = splitApp e
1281 splitApp (AnnCast (_,e) _) = splitApp e
1282 splitApp e = (e, [])
1285 isTypeAtom :: AnnExpr' id ann -> Bool
1286 isTypeAtom (AnnType _) = True
1287 isTypeAtom _ = False
1289 isVoidArgAtom :: AnnExpr' id ann -> Bool
1290 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1291 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1292 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1293 isVoidArgAtom _ = False
1295 atomRep :: AnnExpr' Id ann -> CgRep
1296 atomRep (AnnVar v) = typeCgRep (idType v)
1297 atomRep (AnnLit l) = typeCgRep (literalType l)
1298 atomRep (AnnNote n b) = atomRep (snd b)
1299 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1300 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1301 atomRep (AnnCast b _) = atomRep (snd b)
1302 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1304 isPtrAtom :: AnnExpr' Id ann -> Bool
1305 isPtrAtom e = atomRep e == PtrArg
1307 -- Let szsw be the sizes in words of some items pushed onto the stack,
1308 -- which has initial depth d'. Return the values which the stack environment
1309 -- should map these items to.
1310 mkStackOffsets :: Int -> [Int] -> [Int]
1311 mkStackOffsets original_depth szsw
1312 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1314 -- -----------------------------------------------------------------------------
1315 -- The bytecode generator's monad
1317 type BcPtr = Either ItblPtr (Ptr ())
1321 nextlabel :: Int, -- for generating local labels
1322 malloced :: [BcPtr] } -- thunks malloced for current BCO
1323 -- Should be free()d when it is GCd
1325 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1327 ioToBc :: IO a -> BcM a
1328 ioToBc io = BcM $ \st -> do
1332 runBc :: BcM r -> IO (BcM_State, r)
1333 runBc (BcM m) = m (BcM_State 0 [])
1335 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1336 thenBc (BcM expr) cont = BcM $ \st0 -> do
1337 (st1, q) <- expr st0
1342 thenBc_ :: BcM a -> BcM b -> BcM b
1343 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1344 (st1, q) <- expr st0
1345 (st2, r) <- cont st1
1348 returnBc :: a -> BcM a
1349 returnBc result = BcM $ \st -> (return (st, result))
1351 instance Monad BcM where
1356 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1358 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1360 recordMallocBc :: Ptr a -> BcM ()
1362 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1364 recordItblMallocBc :: ItblPtr -> BcM ()
1365 recordItblMallocBc a
1366 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1368 getLabelBc :: BcM Int
1370 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1372 getLabelsBc :: Int -> BcM [Int]
1374 = BcM $ \st -> let ctr = nextlabel st
1375 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])