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 >= 65535
169 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
171 | stack_overest >= iNTERP_STACK_CHECK_THRESH
172 = STKCHECK stack_overest : peep_d
174 = peep_d -- the supposedly common case
176 stack_overest = sum (map bciStackUse peep_d)
178 -- Merge local pushes
179 peep_d = peep (fromOL instrs_ordlist)
181 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
182 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
183 peep (PUSH_L off1 : PUSH_L off2 : rest)
184 = PUSH_LL off1 (off2-1) : peep rest
190 argBits :: [CgRep] -> [Bool]
193 | isFollowableArg rep = False : argBits args
194 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
196 -- -----------------------------------------------------------------------------
199 -- Compile code for the right-hand side of a top-level binding
201 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
204 schemeTopBind (id, rhs)
205 | Just data_con <- isDataConWorkId_maybe id,
206 isNullaryRepDataCon data_con
207 = -- Special case for the worker of a nullary data con.
208 -- It'll look like this: Nil = /\a -> Nil a
209 -- If we feed it into schemeR, we'll get
211 -- because mkConAppCode treats nullary constructor applications
212 -- by just re-using the single top-level definition. So
213 -- for the worker itself, we must allocate it directly.
214 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
215 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
218 = schemeR [{- No free variables -}] (id, rhs)
220 -- -----------------------------------------------------------------------------
223 -- Compile code for a right-hand side, to give a BCO that,
224 -- when executed with the free variables and arguments on top of the stack,
225 -- will return with a pointer to the result on top of the stack, after
226 -- removing the free variables and arguments.
228 -- Park the resulting BCO in the monad. Also requires the
229 -- variable to which this value was bound, so as to give the
230 -- resulting BCO a name.
232 schemeR :: [Id] -- Free vars of the RHS, ordered as they
233 -- will appear in the thunk. Empty for
234 -- top-level things, which have no free vars.
235 -> (Id, AnnExpr Id VarSet)
236 -> BcM (ProtoBCO Name)
237 schemeR fvs (nm, rhs)
241 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
242 $$ pprCoreExpr (deAnnotate rhs)
248 = schemeR_wrk fvs nm rhs (collect [] rhs)
250 collect xs (_, AnnNote note e) = collect xs e
251 collect xs (_, AnnCast e _) = collect xs e
252 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
253 collect xs (_, not_lambda) = (reverse xs, not_lambda)
255 schemeR_wrk fvs nm original_body (args, body)
257 all_args = reverse args ++ fvs
258 arity = length all_args
259 -- all_args are the args in reverse order. We're compiling a function
260 -- \fv1..fvn x1..xn -> e
261 -- i.e. the fvs come first
263 szsw_args = map idSizeW all_args
264 szw_args = sum szsw_args
265 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
267 -- make the arg bitmap
268 bits = argBits (reverse (map idCgRep all_args))
269 bitmap_size = length bits
270 bitmap = mkBitmap bits
272 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
273 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
274 arity bitmap_size bitmap False{-not alts-})
277 fvsToEnv :: BCEnv -> VarSet -> [Id]
278 -- Takes the free variables of a right-hand side, and
279 -- delivers an ordered list of the local variables that will
280 -- be captured in the thunk for the RHS
281 -- The BCEnv argument tells which variables are in the local
282 -- environment: these are the ones that should be captured
284 -- The code that constructs the thunk, and the code that executes
285 -- it, have to agree about this layout
286 fvsToEnv p fvs = [v | v <- varSetElems fvs,
287 isId v, -- Could be a type variable
290 -- -----------------------------------------------------------------------------
293 -- Compile code to apply the given expression to the remaining args
294 -- on the stack, returning a HNF.
295 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
297 -- Delegate tail-calls to schemeT.
298 schemeE d s p e@(AnnApp f a)
301 schemeE d s p e@(AnnVar v)
302 | not (isUnLiftedType v_type)
303 = -- Lifted-type thing; push it in the normal way
307 = -- Returning an unlifted value.
308 -- Heave it on the stack, SLIDE, and RETURN.
309 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
310 returnBc (push -- value onto stack
311 `appOL` mkSLIDE szw (d-s) -- clear to sequel
312 `snocOL` RETURN_UBX v_rep) -- go
315 v_rep = typeCgRep v_type
317 schemeE d s p (AnnLit literal)
318 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
319 let l_rep = typeCgRep (literalType literal)
320 in returnBc (push -- value onto stack
321 `appOL` mkSLIDE szw (d-s) -- clear to sequel
322 `snocOL` RETURN_UBX l_rep) -- go
325 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
326 | (AnnVar v, args_r_to_l) <- splitApp rhs,
327 Just data_con <- isDataConWorkId_maybe v,
328 dataConRepArity data_con == length args_r_to_l
329 = -- Special case for a non-recursive let whose RHS is a
330 -- saturatred constructor application.
331 -- Just allocate the constructor and carry on
332 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
333 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
334 returnBc (alloc_code `appOL` body_code)
336 -- General case for let. Generates correct, if inefficient, code in
338 schemeE d s p (AnnLet binds (_,body))
339 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
340 AnnRec xs_n_rhss -> unzip xs_n_rhss
343 fvss = map (fvsToEnv p' . fst) rhss
345 -- Sizes of free vars
346 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
348 -- the arity of each rhs
349 arities = map (length . fst . collect []) rhss
351 -- This p', d' defn is safe because all the items being pushed
352 -- are ptrs, so all have size 1. d' and p' reflect the stack
353 -- after the closures have been allocated in the heap (but not
354 -- filled in), and pointers to them parked on the stack.
355 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
357 zipE = zipEqual "schemeE"
359 -- ToDo: don't build thunks for things with no free variables
360 build_thunk dd [] size bco off arity
361 = returnBc (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
363 mkap | arity == 0 = MKAP
365 build_thunk dd (fv:fvs) size bco off arity = do
366 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
367 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
368 returnBc (push_code `appOL` more_push_code)
370 alloc_code = toOL (zipWith mkAlloc sizes arities)
371 where mkAlloc sz 0 = ALLOC_AP sz
372 mkAlloc sz arity = ALLOC_PAP arity sz
374 compile_bind d' fvs x rhs size arity off = do
375 bco <- schemeR fvs (x,rhs)
376 build_thunk d' fvs size bco off arity
379 [ compile_bind d' fvs x rhs size arity n
380 | (fvs, x, rhs, size, arity, n) <-
381 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
384 body_code <- schemeE d' s p' body
385 thunk_codes <- sequence compile_binds
386 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
390 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
391 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
393 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
395 -- case .... of a { DEFAULT -> ... }
396 -- becuse the return convention for both are identical.
398 -- Note that it does not matter losing the void-rep thing from the
399 -- envt (it won't be bound now) because we never look such things up.
401 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
402 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
404 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
405 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
406 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
408 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
409 | isUnboxedTupleCon dc
410 -- Similarly, convert
411 -- case .... of x { (# a #) -> ... }
413 -- case .... of a { DEFAULT -> ... }
414 = --trace "automagic mashing of case alts (# a #)" $
415 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
417 schemeE d s p (AnnCase scrut bndr _ alts)
418 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
420 schemeE d s p (AnnNote note (_, body))
423 schemeE d s p (AnnCast (_, body) _)
427 = pprPanic "ByteCodeGen.schemeE: unhandled case"
428 (pprCoreExpr (deAnnotate' other))
431 -- Compile code to do a tail call. Specifically, push the fn,
432 -- slide the on-stack app back down to the sequel depth,
433 -- and enter. Four cases:
436 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
437 -- The int will be on the stack. Generate a code sequence
438 -- to convert it to the relevant constructor, SLIDE and ENTER.
440 -- 1. The fn denotes a ccall. Defer to generateCCall.
442 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
443 -- it simply as b -- since the representations are identical
444 -- (the VoidArg takes up zero stack space). Also, spot
445 -- (# b #) and treat it as b.
447 -- 3. Application of a constructor, by defn saturated.
448 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
449 -- then the ptrs, and then do PACK and RETURN.
451 -- 4. Otherwise, it must be a function call. Push the args
452 -- right to left, SLIDE and ENTER.
454 schemeT :: Int -- Stack depth
455 -> Sequel -- Sequel depth
456 -> BCEnv -- stack env
457 -> AnnExpr' Id VarSet
462 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
463 -- = panic "schemeT ?!?!"
465 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
469 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
470 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
471 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
472 returnBc (push `appOL` tagToId_sequence
473 `appOL` mkSLIDE 1 (d+arg_words-s)
477 | Just (CCall ccall_spec) <- isFCallId_maybe fn
478 = generateCCall d s p ccall_spec fn args_r_to_l
480 -- Case 2: Constructor application
481 | Just con <- maybe_saturated_dcon,
482 isUnboxedTupleCon con
483 = case args_r_to_l of
484 [arg1,arg2] | isVoidArgAtom arg1 ->
485 unboxedTupleReturn d s p arg2
486 [arg1,arg2] | isVoidArgAtom arg2 ->
487 unboxedTupleReturn d s p arg1
488 _other -> unboxedTupleException
490 -- Case 3: Ordinary data constructor
491 | Just con <- maybe_saturated_dcon
492 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
493 returnBc (alloc_con `appOL`
494 mkSLIDE 1 (d - s) `snocOL`
497 -- Case 4: Tail call of function
499 = doTailCall d s p fn args_r_to_l
502 -- Detect and extract relevant info for the tagToEnum kludge.
503 maybe_is_tagToEnum_call
504 = let extract_constr_Names ty
505 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
507 = map (getName . dataConWorkId) (tyConDataCons tyc)
508 -- NOTE: use the worker name, not the source name of
509 -- the DataCon. See DataCon.lhs for details.
511 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
514 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
515 -> case isPrimOpId_maybe v of
516 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
520 -- Extract the args (R->L) and fn
521 -- The function will necessarily be a variable,
522 -- because we are compiling a tail call
523 (AnnVar fn, args_r_to_l) = splitApp app
525 -- Only consider this to be a constructor application iff it is
526 -- saturated. Otherwise, we'll call the constructor wrapper.
527 n_args = length args_r_to_l
529 = case isDataConWorkId_maybe fn of
530 Just con | dataConRepArity con == n_args -> Just con
533 -- -----------------------------------------------------------------------------
534 -- Generate code to build a constructor application,
535 -- leaving it on top of the stack
537 mkConAppCode :: Int -> Sequel -> BCEnv
538 -> DataCon -- The data constructor
539 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
542 mkConAppCode orig_d s p con [] -- Nullary constructor
543 = ASSERT( isNullaryRepDataCon con )
544 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
545 -- Instead of doing a PACK, which would allocate a fresh
546 -- copy of this constructor, use the single shared version.
548 mkConAppCode orig_d s p con args_r_to_l
549 = ASSERT( dataConRepArity con == length args_r_to_l )
550 do_pushery orig_d (non_ptr_args ++ ptr_args)
552 -- The args are already in reverse order, which is the way PACK
553 -- expects them to be. We must push the non-ptrs after the ptrs.
554 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
556 do_pushery d (arg:args)
557 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
558 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
559 returnBc (push `appOL` more_push_code)
561 = returnBc (unitOL (PACK con n_arg_words))
563 n_arg_words = d - orig_d
566 -- -----------------------------------------------------------------------------
567 -- Returning an unboxed tuple with one non-void component (the only
568 -- case we can handle).
570 -- Remember, we don't want to *evaluate* the component that is being
571 -- returned, even if it is a pointed type. We always just return.
574 :: Int -> Sequel -> BCEnv
575 -> AnnExpr' Id VarSet -> BcM BCInstrList
576 unboxedTupleReturn d s p arg = do
577 (push, sz) <- pushAtom d p arg
578 returnBc (push `appOL`
579 mkSLIDE sz (d-s) `snocOL`
580 RETURN_UBX (atomRep arg))
582 -- -----------------------------------------------------------------------------
583 -- Generate code for a tail-call
586 :: Int -> Sequel -> BCEnv
587 -> Id -> [AnnExpr' Id VarSet]
589 doTailCall init_d s p fn args
590 = do_pushes init_d args (map atomRep args)
592 do_pushes d [] reps = do
593 ASSERT( null reps ) return ()
594 (push_fn, sz) <- pushAtom d p (AnnVar fn)
595 ASSERT( sz == 1 ) return ()
596 returnBc (push_fn `appOL` (
597 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
599 do_pushes d args reps = do
600 let (push_apply, n, rest_of_reps) = findPushSeq reps
601 (these_args, rest_of_args) = splitAt n args
602 (next_d, push_code) <- push_seq d these_args
603 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
604 -- ^^^ for the PUSH_APPLY_ instruction
605 returnBc (push_code `appOL` (push_apply `consOL` instrs))
607 push_seq d [] = return (d, nilOL)
608 push_seq d (arg:args) = do
609 (push_code, sz) <- pushAtom d p arg
610 (final_d, more_push_code) <- push_seq (d+sz) args
611 return (final_d, push_code `appOL` more_push_code)
613 -- v. similar to CgStackery.findMatch, ToDo: merge
614 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
615 = (PUSH_APPLY_PPPPPP, 6, rest)
616 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
617 = (PUSH_APPLY_PPPPP, 5, rest)
618 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
619 = (PUSH_APPLY_PPPP, 4, rest)
620 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
621 = (PUSH_APPLY_PPP, 3, rest)
622 findPushSeq (PtrArg: PtrArg: rest)
623 = (PUSH_APPLY_PP, 2, rest)
624 findPushSeq (PtrArg: rest)
625 = (PUSH_APPLY_P, 1, rest)
626 findPushSeq (VoidArg: rest)
627 = (PUSH_APPLY_V, 1, rest)
628 findPushSeq (NonPtrArg: rest)
629 = (PUSH_APPLY_N, 1, rest)
630 findPushSeq (FloatArg: rest)
631 = (PUSH_APPLY_F, 1, rest)
632 findPushSeq (DoubleArg: rest)
633 = (PUSH_APPLY_D, 1, rest)
634 findPushSeq (LongArg: rest)
635 = (PUSH_APPLY_L, 1, rest)
637 = panic "ByteCodeGen.findPushSeq"
639 -- -----------------------------------------------------------------------------
642 doCase :: Int -> Sequel -> BCEnv
643 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
644 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
646 doCase d s p (_,scrut)
647 bndr alts is_unboxed_tuple
649 -- Top of stack is the return itbl, as usual.
650 -- underneath it is the pointer to the alt_code BCO.
651 -- When an alt is entered, it assumes the returned value is
652 -- on top of the itbl.
655 -- An unlifted value gets an extra info table pushed on top
656 -- when it is returned.
657 unlifted_itbl_sizeW | isAlgCase = 0
660 -- depth of stack after the return value has been pushed
661 d_bndr = d + ret_frame_sizeW + idSizeW bndr
663 -- depth of stack after the extra info table for an unboxed return
664 -- has been pushed, if any. This is the stack depth at the
666 d_alts = d_bndr + unlifted_itbl_sizeW
668 -- Env in which to compile the alts, not including
669 -- any vars bound by the alts themselves
670 p_alts = addToFM p bndr (d_bndr - 1)
672 bndr_ty = idType bndr
673 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
675 -- given an alt, return a discr and code for it.
676 codeALt alt@(DEFAULT, _, (_,rhs))
677 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
678 returnBc (NoDiscr, rhs_code)
679 codeAlt alt@(discr, bndrs, (_,rhs))
680 -- primitive or nullary constructor alt: no need to UNPACK
681 | null real_bndrs = do
682 rhs_code <- schemeE d_alts s p_alts rhs
683 returnBc (my_discr alt, rhs_code)
684 -- algebraic alt with some binders
685 | ASSERT(isAlgCase) otherwise =
687 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
688 ptr_sizes = map idSizeW ptrs
689 nptrs_sizes = map idSizeW nptrs
690 bind_sizes = ptr_sizes ++ nptrs_sizes
691 size = sum ptr_sizes + sum nptrs_sizes
692 -- the UNPACK instruction unpacks in reverse order...
693 p' = addListToFM p_alts
694 (zip (reverse (ptrs ++ nptrs))
695 (mkStackOffsets d_alts (reverse bind_sizes)))
697 rhs_code <- schemeE (d_alts+size) s p' rhs
698 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
700 real_bndrs = filter (not.isTyVar) bndrs
703 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
704 my_discr (DataAlt dc, binds, rhs)
705 | isUnboxedTupleCon dc
706 = unboxedTupleException
708 = DiscrP (dataConTag dc - fIRST_TAG)
709 my_discr (LitAlt l, binds, rhs)
710 = case l of MachInt i -> DiscrI (fromInteger i)
711 MachFloat r -> DiscrF (fromRational r)
712 MachDouble r -> DiscrD (fromRational r)
713 MachChar i -> DiscrI (ord i)
714 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
717 | not isAlgCase = Nothing
719 = case [dc | (DataAlt dc, _, _) <- alts] of
721 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
723 -- the bitmap is relative to stack depth d, i.e. before the
724 -- BCO, info table and return value are pushed on.
725 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
726 -- except that here we build the bitmap from the known bindings of
727 -- things that are pointers, whereas in CgBindery the code builds the
728 -- bitmap from the free slots and unboxed bindings.
731 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
732 -- The bitmap must cover the portion of the stack up to the sequel only.
733 -- Previously we were building a bitmap for the whole depth (d), but we
734 -- really want a bitmap up to depth (d-s). This affects compilation of
735 -- case-of-case expressions, which is the only time we can be compiling a
736 -- case expression with s /= 0.
738 bitmap = intsToReverseBitmap bitmap_size{-size-}
739 (sortLe (<=) (filter (< bitmap_size) rel_slots))
742 rel_slots = concat (map spread binds)
744 | isFollowableArg (idCgRep id) = [ rel_offset ]
746 where rel_offset = d - offset - 1
749 alt_stuff <- mapM codeAlt alts
750 alt_final <- mkMultiBranch maybe_ncons alt_stuff
752 alt_bco_name = getName bndr
753 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
754 0{-no arity-} bitmap_size bitmap True{-is alts-}
756 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
757 -- "\n bitmap = " ++ show bitmap) $ do
758 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
759 alt_bco' <- emitBc alt_bco
761 | isAlgCase = PUSH_ALTS alt_bco'
762 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
763 returnBc (push_alts `consOL` scrut_code)
766 -- -----------------------------------------------------------------------------
767 -- Deal with a CCall.
769 -- Taggedly push the args onto the stack R->L,
770 -- deferencing ForeignObj#s and adjusting addrs to point to
771 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
772 -- (machine) code for the ccall, and create bytecodes to call that and
773 -- then return in the right way.
775 generateCCall :: Int -> Sequel -- stack and sequel depths
777 -> CCallSpec -- where to call
778 -> Id -- of target, for type info
779 -> [AnnExpr' Id VarSet] -- args (atoms)
782 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
785 addr_sizeW = cgRepSizeW NonPtrArg
787 -- Get the args on the stack, with tags and suitably
788 -- dereferenced for the CCall. For each arg, return the
789 -- depth to the first word of the bits for that arg, and the
790 -- CgRep of what was actually pushed.
792 pargs d [] = returnBc []
794 = let arg_ty = repType (exprType (deAnnotate' a))
796 in case splitTyConApp_maybe arg_ty of
797 -- Don't push the FO; instead push the Addr# it
800 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
801 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
802 parg_ArrayishRep arrPtrsHdrSize d p a
804 returnBc ((code,NonPtrArg):rest)
806 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
807 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
808 parg_ArrayishRep arrWordsHdrSize d p a
810 returnBc ((code,NonPtrArg):rest)
812 -- Default case: push taggedly, but otherwise intact.
814 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
815 pargs (d+sz_a) az `thenBc` \ rest ->
816 returnBc ((code_a, atomRep a) : rest)
818 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
819 -- the stack but then advance it over the headers, so as to
820 -- point to the payload.
821 parg_ArrayishRep hdrSize d p a
822 = pushAtom d p a `thenBc` \ (push_fo, _) ->
823 -- The ptr points at the header. Advance it over the
824 -- header and then pretend this is an Addr#.
825 returnBc (push_fo `snocOL` SWIZZLE 0 hdrSize)
828 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
830 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
832 push_args = concatOL pushs_arg
833 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
835 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
836 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
838 = reverse (tail a_reps_pushed_r_to_l)
840 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
841 -- push_args is the code to do that.
842 -- d_after_args is the stack depth once the args are on.
844 -- Get the result rep.
845 (returns_void, r_rep)
846 = case maybe_getCCallReturnRep (idType fn) of
847 Nothing -> (True, VoidArg)
848 Just rr -> (False, rr)
850 Because the Haskell stack grows down, the a_reps refer to
851 lowest to highest addresses in that order. The args for the call
852 are on the stack. Now push an unboxed Addr# indicating
853 the C function to call. Then push a dummy placeholder for the
854 result. Finally, emit a CCALL insn with an offset pointing to the
855 Addr# just pushed, and a literal field holding the mallocville
856 address of the piece of marshalling code we generate.
857 So, just prior to the CCALL insn, the stack looks like this
858 (growing down, as usual):
863 Addr# address_of_C_fn
864 <placeholder-for-result#> (must be an unboxed type)
866 The interpreter then calls the marshall code mentioned
867 in the CCALL insn, passing it (& <placeholder-for-result#>),
868 that is, the addr of the topmost word in the stack.
869 When this returns, the placeholder will have been
870 filled in. The placeholder is slid down to the sequel
871 depth, and we RETURN.
873 This arrangement makes it simple to do f-i-dynamic since the Addr#
874 value is the first arg anyway.
876 The marshalling code is generated specifically for this
877 call site, and so knows exactly the (Haskell) stack
878 offsets of the args, fn address and placeholder. It
879 copies the args to the C stack, calls the stacked addr,
880 and parks the result back in the placeholder. The interpreter
881 calls it as a normal C call, assuming it has a signature
882 void marshall_code ( StgWord* ptr_to_top_of_stack )
884 -- resolve static address
888 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
890 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
893 get_target_info `thenBc` \ (is_static, static_target_addr) ->
896 -- Get the arg reps, zapping the leading Addr# in the dynamic case
897 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
898 | is_static = a_reps_pushed_RAW
899 | otherwise = if null a_reps_pushed_RAW
900 then panic "ByteCodeGen.generateCCall: dyn with no args"
901 else tail a_reps_pushed_RAW
904 (push_Addr, d_after_Addr)
906 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
907 d_after_args + addr_sizeW)
908 | otherwise -- is already on the stack
909 = (nilOL, d_after_args)
911 -- Push the return placeholder. For a call returning nothing,
912 -- this is a VoidArg (tag).
913 r_sizeW = cgRepSizeW r_rep
914 d_after_r = d_after_Addr + r_sizeW
915 r_lit = mkDummyLiteral r_rep
916 push_r = (if returns_void
918 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
920 -- generate the marshalling code we're going to call
923 arg1_offW = r_sizeW + addr_sizeW
924 args_offW = map (arg1_offW +)
925 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
927 ioToBc (mkMarshalCode cconv
928 (r_offW, r_rep) addr_offW
929 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
930 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller)) `thenBc_`
932 -- Offset of the next stack frame down the stack. The CCALL
933 -- instruction needs to describe the chunk of stack containing
934 -- the ccall args to the GC, so it needs to know how large it
935 -- is. See comment in Interpreter.c with the CCALL instruction.
936 stk_offset = d_after_r - s
939 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
941 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
942 `snocOL` RETURN_UBX r_rep
944 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
947 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
951 -- Make a dummy literal, to be used as a placeholder for FFI return
952 -- values on the stack.
953 mkDummyLiteral :: CgRep -> Literal
956 NonPtrArg -> MachWord 0
957 DoubleArg -> MachDouble 0
958 FloatArg -> MachFloat 0
959 _ -> moan64 "mkDummyLiteral" (ppr pr)
963 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
964 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
967 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
969 -- Alternatively, for call-targets returning nothing, convert
971 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
972 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
976 maybe_getCCallReturnRep :: Type -> Maybe CgRep
977 maybe_getCCallReturnRep fn_ty
978 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
980 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
982 = case splitTyConApp_maybe (repType r_ty) of
983 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
985 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
986 || r_reps == [VoidArg] )
987 && isUnboxedTupleTyCon r_tycon
988 && case maybe_r_rep_to_go of
990 Just r_rep -> r_rep /= PtrArg
991 -- if it was, it would be impossible
992 -- to create a valid return value
993 -- placeholder on the stack
994 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
997 --trace (showSDoc (ppr (a_reps, r_reps))) $
998 if ok then maybe_r_rep_to_go else blargh
1000 -- Compile code which expects an unboxed Int on the top of stack,
1001 -- (call it i), and pushes the i'th closure in the supplied list
1002 -- as a consequence.
1003 implement_tagToId :: [Name] -> BcM BCInstrList
1004 implement_tagToId names
1005 = ASSERT( notNull names )
1006 getLabelsBc (length names) `thenBc` \ labels ->
1007 getLabelBc `thenBc` \ label_fail ->
1008 getLabelBc `thenBc` \ label_exit ->
1009 zip4 labels (tail labels ++ [label_fail])
1010 [0 ..] names `bind` \ infos ->
1011 map (mkStep label_exit) infos `bind` \ steps ->
1012 returnBc (concatOL steps
1014 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1016 mkStep l_exit (my_label, next_label, n, name_for_n)
1017 = toOL [LABEL my_label,
1018 TESTEQ_I n next_label,
1023 -- -----------------------------------------------------------------------------
1026 -- Push an atom onto the stack, returning suitable code & number of
1027 -- stack words used.
1029 -- The env p must map each variable to the highest- numbered stack
1030 -- slot for it. For example, if the stack has depth 4 and we
1031 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1032 -- the tag in stack[5], the stack will have depth 6, and p must map v
1033 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1034 -- depth 6 stack has valid words 0 .. 5.
1036 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1038 pushAtom d p (AnnApp f (_, AnnType _))
1039 = pushAtom d p (snd f)
1041 pushAtom d p (AnnNote note e)
1042 = pushAtom d p (snd e)
1044 pushAtom d p (AnnLam x e)
1046 = pushAtom d p (snd e)
1048 pushAtom d p (AnnVar v)
1050 | idCgRep v == VoidArg
1051 = returnBc (nilOL, 0)
1054 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1056 | Just primop <- isPrimOpId_maybe v
1057 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1059 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1060 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1061 -- d - d_v the number of words between the TOS
1062 -- and the 1st slot of the object
1064 -- d - d_v - 1 the offset from the TOS of the 1st slot
1066 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1069 -- Having found the last slot, we proceed to copy the right number of
1070 -- slots on to the top of the stack.
1072 | otherwise -- v must be a global variable
1074 returnBc (unitOL (PUSH_G (getName v)), sz)
1080 pushAtom d p (AnnLit lit)
1082 MachLabel fs _ -> code NonPtrArg
1083 MachWord w -> code NonPtrArg
1084 MachInt i -> code PtrArg
1085 MachFloat r -> code FloatArg
1086 MachDouble r -> code DoubleArg
1087 MachChar c -> code NonPtrArg
1088 MachStr s -> pushStr s
1091 = let size_host_words = cgRepSizeW rep
1092 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1096 = let getMallocvilleAddr
1098 FastString _ n _ fp _ ->
1099 -- we could grab the Ptr from the ForeignPtr,
1100 -- but then we have no way to control its lifetime.
1101 -- In reality it'll probably stay alive long enoungh
1102 -- by virtue of the global FastString table, but
1103 -- to be on the safe side we copy the string into
1104 -- a malloc'd area of memory.
1105 ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1106 recordMallocBc ptr `thenBc_`
1108 withForeignPtr fp $ \p -> do
1109 memcpy ptr p (fromIntegral n)
1110 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1114 getMallocvilleAddr `thenBc` \ addr ->
1115 -- Get the addr on the stack, untaggedly
1116 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1118 pushAtom d p (AnnCast e _)
1119 = pushAtom d p (snd e)
1122 = pprPanic "ByteCodeGen.pushAtom"
1123 (pprCoreExpr (deAnnotate (undefined, other)))
1125 foreign import ccall unsafe "memcpy"
1126 memcpy :: Ptr a -> Ptr b -> CInt -> IO ()
1129 -- -----------------------------------------------------------------------------
1130 -- Given a bunch of alts code and their discrs, do the donkey work
1131 -- of making a multiway branch using a switch tree.
1132 -- What a load of hassle!
1134 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1135 -- a hint; generates better code
1136 -- Nothing is always safe
1137 -> [(Discr, BCInstrList)]
1139 mkMultiBranch maybe_ncons raw_ways
1140 = let d_way = filter (isNoDiscr.fst) raw_ways
1142 (\w1 w2 -> leAlt (fst w1) (fst w2))
1143 (filter (not.isNoDiscr.fst) raw_ways)
1145 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1146 mkTree [] range_lo range_hi = returnBc the_default
1148 mkTree [val] range_lo range_hi
1149 | range_lo `eqAlt` range_hi
1150 = returnBc (snd val)
1152 = getLabelBc `thenBc` \ label_neq ->
1153 returnBc (mkTestEQ (fst val) label_neq
1155 `appOL` unitOL (LABEL label_neq)
1156 `appOL` the_default))
1158 mkTree vals range_lo range_hi
1159 = let n = length vals `div` 2
1160 vals_lo = take n vals
1161 vals_hi = drop n vals
1162 v_mid = fst (head vals_hi)
1164 getLabelBc `thenBc` \ label_geq ->
1165 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1166 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1167 returnBc (mkTestLT v_mid label_geq
1169 `appOL` unitOL (LABEL label_geq)
1173 = case d_way of [] -> unitOL CASEFAIL
1176 -- None of these will be needed if there are no non-default alts
1177 (mkTestLT, mkTestEQ, init_lo, init_hi)
1179 = panic "mkMultiBranch: awesome foursome"
1181 = case fst (head notd_ways) of {
1182 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1183 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1186 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1187 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1190 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1191 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1194 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1195 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1197 DiscrP algMaxBound )
1200 (algMinBound, algMaxBound)
1201 = case maybe_ncons of
1202 Just n -> (0, n - 1)
1203 Nothing -> (minBound, maxBound)
1205 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1206 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1207 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1208 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1209 NoDiscr `eqAlt` NoDiscr = True
1212 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1213 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1214 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1215 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1216 NoDiscr `leAlt` NoDiscr = True
1219 isNoDiscr NoDiscr = True
1222 dec (DiscrI i) = DiscrI (i-1)
1223 dec (DiscrP i) = DiscrP (i-1)
1224 dec other = other -- not really right, but if you
1225 -- do cases on floating values, you'll get what you deserve
1227 -- same snotty comment applies to the following
1229 minD, maxD :: Double
1235 mkTree notd_ways init_lo init_hi
1238 -- -----------------------------------------------------------------------------
1239 -- Supporting junk for the compilation schemes
1241 -- Describes case alts
1249 instance Outputable Discr where
1250 ppr (DiscrI i) = int i
1251 ppr (DiscrF f) = text (show f)
1252 ppr (DiscrD d) = text (show d)
1253 ppr (DiscrP i) = int i
1254 ppr NoDiscr = text "DEF"
1257 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1258 lookupBCEnv_maybe = lookupFM
1260 idSizeW :: Id -> Int
1261 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1263 unboxedTupleException :: a
1264 unboxedTupleException
1267 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1268 "\tto foreign import/export decls in source. Workaround:\n" ++
1269 "\tcompile this module to a .o file, then restart session."))
1272 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1275 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1276 -- The arguments are returned in *right-to-left* order
1277 splitApp (AnnApp (_,f) (_,a))
1278 | isTypeAtom a = splitApp f
1279 | otherwise = case splitApp f of
1280 (f', as) -> (f', a:as)
1281 splitApp (AnnNote n (_,e)) = splitApp e
1282 splitApp (AnnCast (_,e) _) = splitApp e
1283 splitApp e = (e, [])
1286 isTypeAtom :: AnnExpr' id ann -> Bool
1287 isTypeAtom (AnnType _) = True
1288 isTypeAtom _ = False
1290 isVoidArgAtom :: AnnExpr' id ann -> Bool
1291 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1292 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1293 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1294 isVoidArgAtom _ = False
1296 atomRep :: AnnExpr' Id ann -> CgRep
1297 atomRep (AnnVar v) = typeCgRep (idType v)
1298 atomRep (AnnLit l) = typeCgRep (literalType l)
1299 atomRep (AnnNote n b) = atomRep (snd b)
1300 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1301 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1302 atomRep (AnnCast b _) = atomRep (snd b)
1303 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1305 isPtrAtom :: AnnExpr' Id ann -> Bool
1306 isPtrAtom e = atomRep e == PtrArg
1308 -- Let szsw be the sizes in words of some items pushed onto the stack,
1309 -- which has initial depth d'. Return the values which the stack environment
1310 -- should map these items to.
1311 mkStackOffsets :: Int -> [Int] -> [Int]
1312 mkStackOffsets original_depth szsw
1313 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1315 -- -----------------------------------------------------------------------------
1316 -- The bytecode generator's monad
1318 type BcPtr = Either ItblPtr (Ptr ())
1322 nextlabel :: Int, -- for generating local labels
1323 malloced :: [BcPtr] } -- thunks malloced for current BCO
1324 -- Should be free()d when it is GCd
1326 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1328 ioToBc :: IO a -> BcM a
1329 ioToBc io = BcM $ \st -> do
1333 runBc :: BcM r -> IO (BcM_State, r)
1334 runBc (BcM m) = m (BcM_State 0 [])
1336 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1337 thenBc (BcM expr) cont = BcM $ \st0 -> do
1338 (st1, q) <- expr st0
1343 thenBc_ :: BcM a -> BcM b -> BcM b
1344 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1345 (st1, q) <- expr st0
1346 (st2, r) <- cont st1
1349 returnBc :: a -> BcM a
1350 returnBc result = BcM $ \st -> (return (st, result))
1352 instance Monad BcM where
1357 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1359 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1361 recordMallocBc :: Ptr a -> BcM ()
1363 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1365 recordItblMallocBc :: ItblPtr -> BcM ()
1366 recordItblMallocBc a
1367 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1369 getLabelBc :: BcM Int
1371 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1373 getLabelsBc :: Int -> BcM [Int]
1375 = BcM $ \st -> let ctr = nextlabel st
1376 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])