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, mkSystemName )
21 import ForeignCall ( ForeignCall(..), CCallTarget(..), CCallSpec(..) )
22 import HscTypes ( ModGuts(..), ModGuts,
23 TypeEnv, typeEnvTyCons, typeEnvClasses )
24 import CoreUtils ( exprType )
26 import PprCore ( pprCoreExpr )
27 import Literal ( Literal(..), literalPrimRep )
29 import PrimOp ( PrimOp(..) )
30 import CoreFVs ( freeVars )
31 import Type ( typePrimRep, isUnLiftedType, splitTyConApp_maybe,
33 import DataCon ( DataCon, dataConTag, fIRST_TAG, dataConTyCon,
34 isUnboxedTupleCon, isNullaryDataCon, dataConWorkId,
36 import TyCon ( tyConFamilySize, isDataTyCon, tyConDataCons,
37 isFunTyCon, isUnboxedTupleTyCon )
38 import Class ( Class, classTyCon )
39 import Type ( Type, repType, splitFunTys, dropForAlls )
41 import DataCon ( dataConRepArity )
42 import Var ( isTyVar )
43 import VarSet ( VarSet, varSetElems )
44 import TysPrim ( foreignObjPrimTyCon,
45 arrayPrimTyCon, mutableArrayPrimTyCon,
46 byteArrayPrimTyCon, mutableByteArrayPrimTyCon
48 import PrimRep ( isFollowableRep )
49 import CmdLineOpts ( DynFlags, DynFlag(..) )
50 import ErrUtils ( showPass, dumpIfSet_dyn )
51 import Unique ( mkPseudoUnique3 )
52 import FastString ( FastString(..), unpackFS )
53 import Panic ( GhcException(..) )
54 import PprType ( pprType )
55 import SMRep ( arrWordsHdrSize, arrPtrsHdrSize, StgWord )
56 import Bitmap ( intsToReverseBitmap, mkBitmap )
58 import Constants ( wORD_SIZE )
59 import BasicTypes ( TopLevelFlag(..), isTopLevel, isNotTopLevel )
61 import Data.List ( intersperse, sortBy, zip4, zip5, partition )
62 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8 )
63 import Foreign.C ( CInt )
64 import Control.Exception ( throwDyn )
66 import GHC.Exts ( Int(..), ByteArray# )
68 import Control.Monad ( when, mapAndUnzipM )
69 import Data.Char ( ord )
72 -- -----------------------------------------------------------------------------
73 -- Generating byte code for a complete module
75 byteCodeGen :: DynFlags
78 -> IO CompiledByteCode
79 byteCodeGen dflags binds type_env
80 = do showPass dflags "ByteCodeGen"
81 let local_tycons = typeEnvTyCons type_env
82 local_classes = typeEnvClasses type_env
83 tycs = local_tycons ++ map classTyCon local_classes
85 let flatBinds = [ (bndr, freeVars rhs)
86 | (bndr, rhs) <- flattenBinds binds]
88 (BcM_State final_ctr mallocd, proto_bcos)
89 <- runBc (mapM schemeTopBind flatBinds)
91 when (notNull mallocd)
92 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
94 dumpIfSet_dyn dflags Opt_D_dump_BCOs
95 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
97 assembleBCOs proto_bcos tycs
99 -- -----------------------------------------------------------------------------
100 -- Generating byte code for an expression
102 -- Returns: (the root BCO for this expression,
103 -- a list of auxilary BCOs resulting from compiling closures)
104 coreExprToBCOs :: DynFlags
107 coreExprToBCOs dflags expr
108 = do showPass dflags "ByteCodeGen"
110 -- create a totally bogus name for the top-level BCO; this
111 -- should be harmless, since it's never used for anything
112 let invented_name = mkSystemName (mkPseudoUnique3 0) FSLIT("ExprTopLevel")
113 invented_id = mkLocalId invented_name (panic "invented_id's type")
115 (BcM_State final_ctr mallocd, proto_bco)
116 <- runBc (schemeTopBind (invented_id, freeVars expr))
118 when (notNull mallocd)
119 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
121 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
123 assembleBCO proto_bco
126 -- -----------------------------------------------------------------------------
127 -- Compilation schema for the bytecode generator
129 type BCInstrList = OrdList BCInstr
131 type Sequel = Int -- back off to this depth before ENTER
133 -- Maps Ids to the offset from the stack _base_ so we don't have
134 -- to mess with it after each push/pop.
135 type BCEnv = FiniteMap Id Int -- To find vars on the stack
137 ppBCEnv :: BCEnv -> SDoc
140 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
143 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idPrimRep var)
144 cmp_snd x y = compare (snd x) (snd y)
146 -- Create a BCO and do a spot of peephole optimisation on the insns
151 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
155 -> Bool -- True <=> is a return point, rather than a function
158 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap
159 is_ret mallocd_blocks
162 protoBCOInstrs = maybe_with_stack_check,
163 protoBCOBitmap = bitmap,
164 protoBCOBitmapSize = bitmap_size,
165 protoBCOArity = arity,
166 protoBCOExpr = origin,
167 protoBCOPtrs = mallocd_blocks
170 -- Overestimate the stack usage (in words) of this BCO,
171 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
172 -- stack check. (The interpreter always does a stack check
173 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
174 -- BCO anyway, so we only need to add an explicit on in the
175 -- (hopefully rare) cases when the (overestimated) stack use
176 -- exceeds iNTERP_STACK_CHECK_THRESH.
177 maybe_with_stack_check
179 -- don't do stack checks at return points;
180 -- everything is aggregated up to the top BCO
181 -- (which must be a function)
182 | stack_overest >= 65535
183 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
185 | stack_overest >= iNTERP_STACK_CHECK_THRESH
186 = STKCHECK stack_overest : peep_d
188 = peep_d -- the supposedly common case
190 stack_overest = sum (map bciStackUse peep_d)
192 -- Merge local pushes
193 peep_d = peep (fromOL instrs_ordlist)
195 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
196 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
197 peep (PUSH_L off1 : PUSH_L off2 : rest)
198 = PUSH_LL off1 (off2-1) : peep rest
204 argBits :: [PrimRep] -> [Bool]
207 | isFollowableRep rep = False : argBits args
208 | otherwise = take (getPrimRepSize rep) (repeat True) ++ argBits args
210 -- -----------------------------------------------------------------------------
213 -- Compile code for the right-hand side of a top-level binding
215 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
218 schemeTopBind (id, rhs)
219 | Just data_con <- isDataConWorkId_maybe id,
220 isNullaryDataCon data_con
221 = -- Special case for the worker of a nullary data con.
222 -- It'll look like this: Nil = /\a -> Nil a
223 -- If we feed it into schemeR, we'll get
225 -- because mkConAppCode treats nullary constructor applications
226 -- by just re-using the single top-level definition. So
227 -- for the worker itself, we must allocate it directly.
228 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
229 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
232 = schemeR [{- No free variables -}] (id, rhs)
234 -- -----------------------------------------------------------------------------
237 -- Compile code for a right-hand side, to give a BCO that,
238 -- when executed with the free variables and arguments on top of the stack,
239 -- will return with a pointer to the result on top of the stack, after
240 -- removing the free variables and arguments.
242 -- Park the resulting BCO in the monad. Also requires the
243 -- variable to which this value was bound, so as to give the
244 -- resulting BCO a name.
246 schemeR :: [Id] -- Free vars of the RHS, ordered as they
247 -- will appear in the thunk. Empty for
248 -- top-level things, which have no free vars.
249 -> (Id, AnnExpr Id VarSet)
250 -> BcM (ProtoBCO Name)
251 schemeR fvs (nm, rhs)
255 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
256 $$ pprCoreExpr (deAnnotate rhs)
262 = schemeR_wrk fvs nm rhs (collect [] rhs)
264 collect xs (_, AnnNote note e) = collect xs e
265 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
266 collect xs (_, not_lambda) = (reverse xs, not_lambda)
268 schemeR_wrk fvs nm original_body (args, body)
270 all_args = reverse args ++ fvs
271 arity = length all_args
272 -- all_args are the args in reverse order. We're compiling a function
273 -- \fv1..fvn x1..xn -> e
274 -- i.e. the fvs come first
276 szsw_args = map idSizeW all_args
277 szw_args = sum szsw_args
278 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
280 -- make the arg bitmap
281 bits = argBits (reverse (map idPrimRep all_args))
282 bitmap_size = length bits
283 bitmap = mkBitmap bits
285 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
286 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
287 arity bitmap_size bitmap False{-not alts-})
290 fvsToEnv :: BCEnv -> VarSet -> [Id]
291 -- Takes the free variables of a right-hand side, and
292 -- delivers an ordered list of the local variables that will
293 -- be captured in the thunk for the RHS
294 -- The BCEnv argument tells which variables are in the local
295 -- environment: these are the ones that should be captured
297 -- The code that constructs the thunk, and the code that executes
298 -- it, have to agree about this layout
299 fvsToEnv p fvs = [v | v <- varSetElems fvs,
300 isId v, -- Could be a type variable
303 -- -----------------------------------------------------------------------------
306 -- Compile code to apply the given expression to the remaining args
307 -- on the stack, returning a HNF.
308 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
310 -- Delegate tail-calls to schemeT.
311 schemeE d s p e@(AnnApp f a)
314 schemeE d s p e@(AnnVar v)
315 | not (isUnLiftedType v_type)
316 = -- Lifted-type thing; push it in the normal way
320 = -- Returning an unlifted value.
321 -- Heave it on the stack, SLIDE, and RETURN.
322 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
323 returnBc (push -- value onto stack
324 `appOL` mkSLIDE szw (d-s) -- clear to sequel
325 `snocOL` RETURN_UBX v_rep) -- go
328 v_rep = typePrimRep v_type
330 schemeE d s p (AnnLit literal)
331 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
332 let l_rep = literalPrimRep literal
333 in returnBc (push -- value onto stack
334 `appOL` mkSLIDE szw (d-s) -- clear to sequel
335 `snocOL` RETURN_UBX l_rep) -- go
338 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
339 | (AnnVar v, args_r_to_l) <- splitApp rhs,
340 Just data_con <- isDataConWorkId_maybe v,
341 dataConRepArity data_con == length args_r_to_l
342 = -- Special case for a non-recursive let whose RHS is a
343 -- saturatred constructor application.
344 -- Just allocate the constructor and carry on
345 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
346 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
347 returnBc (alloc_code `appOL` body_code)
349 -- General case for let. Generates correct, if inefficient, code in
351 schemeE d s p (AnnLet binds (_,body))
352 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
353 AnnRec xs_n_rhss -> unzip xs_n_rhss
356 fvss = map (fvsToEnv p' . fst) rhss
358 -- Sizes of free vars
359 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
361 -- the arity of each rhs
362 arities = map (length . fst . collect []) rhss
364 -- This p', d' defn is safe because all the items being pushed
365 -- are ptrs, so all have size 1. d' and p' reflect the stack
366 -- after the closures have been allocated in the heap (but not
367 -- filled in), and pointers to them parked on the stack.
368 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
370 zipE = zipEqual "schemeE"
372 -- ToDo: don't build thunks for things with no free variables
373 build_thunk dd [] size bco off
374 = returnBc (PUSH_BCO bco
375 `consOL` unitOL (MKAP (off+size) size))
376 build_thunk dd (fv:fvs) size bco off = do
377 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
378 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off
379 returnBc (push_code `appOL` more_push_code)
381 alloc_code = toOL (zipWith mkAlloc sizes arities)
382 where mkAlloc sz 0 = ALLOC_AP sz
383 mkAlloc sz arity = ALLOC_PAP arity sz
385 compile_bind d' fvs x rhs size off = do
386 bco <- schemeR fvs (x,rhs)
387 build_thunk d' fvs size bco off
390 [ compile_bind d' fvs x rhs size n
391 | (fvs, x, rhs, size, n) <-
392 zip5 fvss xs rhss sizes [n_binds, n_binds-1 .. 1]
395 body_code <- schemeE d' s p' body
396 thunk_codes <- sequence compile_binds
397 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
401 schemeE d s p (AnnCase scrut bndr [(DataAlt dc, [bind1, bind2], rhs)])
402 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind1)
404 -- case .... of x { (# VoidRep'd-thing, a #) -> ... }
406 -- case .... of a { DEFAULT -> ... }
407 -- becuse the return convention for both are identical.
409 -- Note that it does not matter losing the void-rep thing from the
410 -- envt (it won't be bound now) because we never look such things up.
412 = --trace "automagic mashing of case alts (# VoidRep, a #)" $
413 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
415 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind2)
416 = --trace "automagic mashing of case alts (# a, VoidRep #)" $
417 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
419 schemeE d s p (AnnCase scrut bndr [(DataAlt dc, [bind1], rhs)])
420 | isUnboxedTupleCon dc
421 -- Similarly, convert
422 -- case .... of x { (# a #) -> ... }
424 -- case .... of a { DEFAULT -> ... }
425 = --trace "automagic mashing of case alts (# a #)" $
426 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
428 schemeE d s p (AnnCase scrut bndr alts)
429 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
431 schemeE d s p (AnnNote note (_, body))
435 = pprPanic "ByteCodeGen.schemeE: unhandled case"
436 (pprCoreExpr (deAnnotate' other))
439 -- Compile code to do a tail call. Specifically, push the fn,
440 -- slide the on-stack app back down to the sequel depth,
441 -- and enter. Four cases:
444 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
445 -- The int will be on the stack. Generate a code sequence
446 -- to convert it to the relevant constructor, SLIDE and ENTER.
448 -- 1. The fn denotes a ccall. Defer to generateCCall.
450 -- 2. (Another nasty hack). Spot (# a::VoidRep, b #) and treat
451 -- it simply as b -- since the representations are identical
452 -- (the VoidRep takes up zero stack space). Also, spot
453 -- (# b #) and treat it as b.
455 -- 3. Application of a constructor, by defn saturated.
456 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
457 -- then the ptrs, and then do PACK and RETURN.
459 -- 4. Otherwise, it must be a function call. Push the args
460 -- right to left, SLIDE and ENTER.
462 schemeT :: Int -- Stack depth
463 -> Sequel -- Sequel depth
464 -> BCEnv -- stack env
465 -> AnnExpr' Id VarSet
470 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
471 -- = panic "schemeT ?!?!"
473 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
477 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
478 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
479 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
480 returnBc (push `appOL` tagToId_sequence
481 `appOL` mkSLIDE 1 (d+arg_words-s)
485 | Just (CCall ccall_spec) <- isFCallId_maybe fn
486 = generateCCall d s p ccall_spec fn args_r_to_l
488 -- Case 2: Constructor application
489 | Just con <- maybe_saturated_dcon,
490 isUnboxedTupleCon con
491 = case args_r_to_l of
492 [arg1,arg2] | isVoidRepAtom arg1 ->
493 unboxedTupleReturn d s p arg2
494 [arg1,arg2] | isVoidRepAtom arg2 ->
495 unboxedTupleReturn d s p arg1
496 _other -> unboxedTupleException
498 -- Case 3: Ordinary data constructor
499 | Just con <- maybe_saturated_dcon
500 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
501 returnBc (alloc_con `appOL`
502 mkSLIDE 1 (d - s) `snocOL`
505 -- Case 4: Tail call of function
507 = doTailCall d s p fn args_r_to_l
510 -- Detect and extract relevant info for the tagToEnum kludge.
511 maybe_is_tagToEnum_call
512 = let extract_constr_Names ty
513 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
515 = map (getName . dataConWorkId) (tyConDataCons tyc)
516 -- NOTE: use the worker name, not the source name of
517 -- the DataCon. See DataCon.lhs for details.
519 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
522 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
523 -> case isPrimOpId_maybe v of
524 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
528 -- Extract the args (R->L) and fn
529 -- The function will necessarily be a variable,
530 -- because we are compiling a tail call
531 (AnnVar fn, args_r_to_l) = splitApp app
533 -- Only consider this to be a constructor application iff it is
534 -- saturated. Otherwise, we'll call the constructor wrapper.
535 n_args = length args_r_to_l
537 = case isDataConWorkId_maybe fn of
538 Just con | dataConRepArity con == n_args -> Just con
541 -- -----------------------------------------------------------------------------
542 -- Generate code to build a constructor application,
543 -- leaving it on top of the stack
545 mkConAppCode :: Int -> Sequel -> BCEnv
546 -> DataCon -- The data constructor
547 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
550 mkConAppCode orig_d s p con [] -- Nullary constructor
551 = ASSERT( isNullaryDataCon con )
552 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
553 -- Instead of doing a PACK, which would allocate a fresh
554 -- copy of this constructor, use the single shared version.
556 mkConAppCode orig_d s p con args_r_to_l
557 = ASSERT( dataConRepArity con == length args_r_to_l )
558 do_pushery orig_d (non_ptr_args ++ ptr_args)
560 -- The args are already in reverse order, which is the way PACK
561 -- expects them to be. We must push the non-ptrs after the ptrs.
562 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
564 do_pushery d (arg:args)
565 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
566 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
567 returnBc (push `appOL` more_push_code)
569 = returnBc (unitOL (PACK con n_arg_words))
571 n_arg_words = d - orig_d
574 -- -----------------------------------------------------------------------------
575 -- Returning an unboxed tuple with one non-void component (the only
576 -- case we can handle).
578 -- Remember, we don't want to *evaluate* the component that is being
579 -- returned, even if it is a pointed type. We always just return.
582 :: Int -> Sequel -> BCEnv
583 -> AnnExpr' Id VarSet -> BcM BCInstrList
584 unboxedTupleReturn d s p arg = do
585 (push, sz) <- pushAtom d p arg
586 returnBc (push `appOL`
587 mkSLIDE sz (d-s) `snocOL`
588 RETURN_UBX (atomRep arg))
590 -- -----------------------------------------------------------------------------
591 -- Generate code for a tail-call
594 :: Int -> Sequel -> BCEnv
595 -> Id -> [AnnExpr' Id VarSet]
597 doTailCall init_d s p fn args
598 = do_pushes init_d args (map (primRepToArgRep.atomRep) args)
600 do_pushes d [] reps = do
602 (push_fn, sz) <- pushAtom d p (AnnVar fn)
604 returnBc (push_fn `appOL` (
605 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
607 do_pushes d args reps = do
608 let (push_apply, n, rest_of_reps) = findPushSeq reps
609 (these_args, rest_of_args) = splitAt n args
610 (next_d, push_code) <- push_seq d these_args
611 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
612 -- ^^^ for the PUSH_APPLY_ instruction
613 returnBc (push_code `appOL` (push_apply `consOL` instrs))
615 push_seq d [] = return (d, nilOL)
616 push_seq d (arg:args) = do
617 (push_code, sz) <- pushAtom d p arg
618 (final_d, more_push_code) <- push_seq (d+sz) args
619 return (final_d, push_code `appOL` more_push_code)
621 -- v. similar to CgStackery.findMatch, ToDo: merge
622 findPushSeq (RepP: RepP: RepP: RepP: RepP: RepP: RepP: rest)
623 = (PUSH_APPLY_PPPPPPP, 7, rest)
624 findPushSeq (RepP: RepP: RepP: RepP: RepP: RepP: rest)
625 = (PUSH_APPLY_PPPPPP, 6, rest)
626 findPushSeq (RepP: RepP: RepP: RepP: RepP: rest)
627 = (PUSH_APPLY_PPPPP, 5, rest)
628 findPushSeq (RepP: RepP: RepP: RepP: rest)
629 = (PUSH_APPLY_PPPP, 4, rest)
630 findPushSeq (RepP: RepP: RepP: rest)
631 = (PUSH_APPLY_PPP, 3, rest)
632 findPushSeq (RepP: RepP: rest)
633 = (PUSH_APPLY_PP, 2, rest)
634 findPushSeq (RepP: rest)
635 = (PUSH_APPLY_P, 1, rest)
636 findPushSeq (RepV: rest)
637 = (PUSH_APPLY_V, 1, rest)
638 findPushSeq (RepN: rest)
639 = (PUSH_APPLY_N, 1, rest)
640 findPushSeq (RepF: rest)
641 = (PUSH_APPLY_F, 1, rest)
642 findPushSeq (RepD: rest)
643 = (PUSH_APPLY_D, 1, rest)
644 findPushSeq (RepL: rest)
645 = (PUSH_APPLY_L, 1, rest)
647 = panic "ByteCodeGen.findPushSeq"
649 -- -----------------------------------------------------------------------------
652 doCase :: Int -> Sequel -> BCEnv
653 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
654 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
656 doCase d s p (_,scrut)
657 bndr alts is_unboxed_tuple
659 -- Top of stack is the return itbl, as usual.
660 -- underneath it is the pointer to the alt_code BCO.
661 -- When an alt is entered, it assumes the returned value is
662 -- on top of the itbl.
665 -- An unlifted value gets an extra info table pushed on top
666 -- when it is returned.
667 unlifted_itbl_sizeW | isAlgCase = 0
670 -- depth of stack after the return value has been pushed
671 d_bndr = d + ret_frame_sizeW + idSizeW bndr
673 -- depth of stack after the extra info table for an unboxed return
674 -- has been pushed, if any. This is the stack depth at the
676 d_alts = d_bndr + unlifted_itbl_sizeW
678 -- Env in which to compile the alts, not including
679 -- any vars bound by the alts themselves
680 p_alts = addToFM p bndr (d_bndr - 1)
682 bndr_ty = idType bndr
683 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
685 -- given an alt, return a discr and code for it.
686 codeALt alt@(DEFAULT, _, (_,rhs))
687 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
688 returnBc (NoDiscr, rhs_code)
689 codeAlt alt@(discr, bndrs, (_,rhs))
690 -- primitive or nullary constructor alt: no need to UNPACK
691 | null real_bndrs = do
692 rhs_code <- schemeE d_alts s p_alts rhs
693 returnBc (my_discr alt, rhs_code)
694 -- algebraic alt with some binders
695 | ASSERT(isAlgCase) otherwise =
697 (ptrs,nptrs) = partition (isFollowableRep.idPrimRep) real_bndrs
698 ptr_sizes = map idSizeW ptrs
699 nptrs_sizes = map idSizeW nptrs
700 bind_sizes = ptr_sizes ++ nptrs_sizes
701 size = sum ptr_sizes + sum nptrs_sizes
702 -- the UNPACK instruction unpacks in reverse order...
703 p' = addListToFM p_alts
704 (zip (reverse (ptrs ++ nptrs))
705 (mkStackOffsets d_alts (reverse bind_sizes)))
707 rhs_code <- schemeE (d_alts+size) s p' rhs
708 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
710 real_bndrs = filter (not.isTyVar) bndrs
713 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
714 my_discr (DataAlt dc, binds, rhs)
715 | isUnboxedTupleCon dc
716 = unboxedTupleException
718 = DiscrP (dataConTag dc - fIRST_TAG)
719 my_discr (LitAlt l, binds, rhs)
720 = case l of MachInt i -> DiscrI (fromInteger i)
721 MachFloat r -> DiscrF (fromRational r)
722 MachDouble r -> DiscrD (fromRational r)
723 MachChar i -> DiscrI i
724 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
727 | not isAlgCase = Nothing
729 = case [dc | (DataAlt dc, _, _) <- alts] of
731 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
733 -- the bitmap is relative to stack depth d, i.e. before the
734 -- BCO, info table and return value are pushed on.
735 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
736 -- except that here we build the bitmap from the known bindings of
737 -- things that are pointers, whereas in CgBindery the code builds the
738 -- bitmap from the free slots and unboxed bindings.
740 bitmap = intsToReverseBitmap d{-size-} (sortLt (<) rel_slots)
743 rel_slots = concat (map spread binds)
745 | isFollowableRep (idPrimRep id) = [ rel_offset ]
747 where rel_offset = d - offset - 1
750 alt_stuff <- mapM codeAlt alts
751 alt_final <- mkMultiBranch maybe_ncons alt_stuff
753 alt_bco_name = getName bndr
754 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
755 0{-no arity-} d{-bitmap size-} bitmap True{-is alts-}
757 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
758 -- "\n bitmap = " ++ show bitmap) $ do
759 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
760 alt_bco' <- emitBc alt_bco
762 | isAlgCase = PUSH_ALTS alt_bco'
763 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typePrimRep bndr_ty)
764 returnBc (push_alts `consOL` scrut_code)
767 -- -----------------------------------------------------------------------------
768 -- Deal with a CCall.
770 -- Taggedly push the args onto the stack R->L,
771 -- deferencing ForeignObj#s and adjusting addrs to point to
772 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
773 -- (machine) code for the ccall, and create bytecodes to call that and
774 -- then return in the right way.
776 generateCCall :: Int -> Sequel -- stack and sequel depths
778 -> CCallSpec -- where to call
779 -> Id -- of target, for type info
780 -> [AnnExpr' Id VarSet] -- args (atoms)
783 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
786 addr_sizeW = getPrimRepSize AddrRep
788 -- Get the args on the stack, with tags and suitably
789 -- dereferenced for the CCall. For each arg, return the
790 -- depth to the first word of the bits for that arg, and the
791 -- PrimRep of what was actually pushed.
793 pargs d [] = returnBc []
795 = let arg_ty = repType (exprType (deAnnotate' a))
797 in case splitTyConApp_maybe arg_ty of
798 -- Don't push the FO; instead push the Addr# it
801 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
802 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
803 parg_ArrayishRep arrPtrsHdrSize d p a
805 returnBc ((code,AddrRep):rest)
807 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
808 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
809 parg_ArrayishRep arrWordsHdrSize d p a
811 returnBc ((code,AddrRep):rest)
813 -- Default case: push taggedly, but otherwise intact.
815 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
816 pargs (d+sz_a) az `thenBc` \ rest ->
817 returnBc ((code_a, atomRep a) : rest)
819 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
820 -- the stack but then advance it over the headers, so as to
821 -- point to the payload.
822 parg_ArrayishRep hdrSizeW d p a
823 = pushAtom d p a `thenBc` \ (push_fo, _) ->
824 -- The ptr points at the header. Advance it over the
825 -- header and then pretend this is an Addr#.
826 returnBc (push_fo `snocOL`
827 SWIZZLE 0 (hdrSizeW * getPrimRepSize WordRep
831 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
833 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
835 push_args = concatOL pushs_arg
836 d_after_args = d0 + sum (map getPrimRepSize a_reps_pushed_r_to_l)
838 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
839 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
841 = reverse (tail a_reps_pushed_r_to_l)
843 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
844 -- push_args is the code to do that.
845 -- d_after_args is the stack depth once the args are on.
847 -- Get the result rep.
848 (returns_void, r_rep)
849 = case maybe_getCCallReturnRep (idType fn) of
850 Nothing -> (True, VoidRep)
851 Just rr -> (False, rr)
853 Because the Haskell stack grows down, the a_reps refer to
854 lowest to highest addresses in that order. The args for the call
855 are on the stack. Now push an unboxed Addr# indicating
856 the C function to call. Then push a dummy placeholder for the
857 result. Finally, emit a CCALL insn with an offset pointing to the
858 Addr# just pushed, and a literal field holding the mallocville
859 address of the piece of marshalling code we generate.
860 So, just prior to the CCALL insn, the stack looks like this
861 (growing down, as usual):
866 Addr# address_of_C_fn
867 <placeholder-for-result#> (must be an unboxed type)
869 The interpreter then calls the marshall code mentioned
870 in the CCALL insn, passing it (& <placeholder-for-result#>),
871 that is, the addr of the topmost word in the stack.
872 When this returns, the placeholder will have been
873 filled in. The placeholder is slid down to the sequel
874 depth, and we RETURN.
876 This arrangement makes it simple to do f-i-dynamic since the Addr#
877 value is the first arg anyway.
879 The marshalling code is generated specifically for this
880 call site, and so knows exactly the (Haskell) stack
881 offsets of the args, fn address and placeholder. It
882 copies the args to the C stack, calls the stacked addr,
883 and parks the result back in the placeholder. The interpreter
884 calls it as a normal C call, assuming it has a signature
885 void marshall_code ( StgWord* ptr_to_top_of_stack )
887 -- resolve static address
891 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
893 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
896 -> pprPanic "ByteCodeGen.generateCCall: casm" (ppr ccall_spec)
898 get_target_info `thenBc` \ (is_static, static_target_addr) ->
901 -- Get the arg reps, zapping the leading Addr# in the dynamic case
902 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
903 | is_static = a_reps_pushed_RAW
904 | otherwise = if null a_reps_pushed_RAW
905 then panic "ByteCodeGen.generateCCall: dyn with no args"
906 else tail a_reps_pushed_RAW
909 (push_Addr, d_after_Addr)
911 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
912 d_after_args + addr_sizeW)
913 | otherwise -- is already on the stack
914 = (nilOL, d_after_args)
916 -- Push the return placeholder. For a call returning nothing,
917 -- this is a VoidRep (tag).
918 r_sizeW = getPrimRepSize r_rep
919 d_after_r = d_after_Addr + r_sizeW
920 r_lit = mkDummyLiteral r_rep
921 push_r = (if returns_void
923 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
925 -- generate the marshalling code we're going to call
928 arg1_offW = r_sizeW + addr_sizeW
929 args_offW = map (arg1_offW +)
930 (init (scanl (+) 0 (map getPrimRepSize a_reps)))
932 ioToBc (mkMarshalCode cconv
933 (r_offW, r_rep) addr_offW
934 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
935 recordMallocBc addr_of_marshaller `thenBc_`
937 -- Offset of the next stack frame down the stack. The CCALL
938 -- instruction needs to describe the chunk of stack containing
939 -- the ccall args to the GC, so it needs to know how large it
940 -- is. See comment in Interpreter.c with the CCALL instruction.
941 stk_offset = d_after_r - s
944 do_call = unitOL (CCALL stk_offset (castPtr addr_of_marshaller))
946 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
947 `snocOL` RETURN_UBX r_rep
949 --trace (show (arg1_offW, args_offW , (map getPrimRepSize a_reps) )) $
952 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
956 -- Make a dummy literal, to be used as a placeholder for FFI return
957 -- values on the stack.
958 mkDummyLiteral :: PrimRep -> Literal
961 CharRep -> MachChar 0
963 WordRep -> MachWord 0
964 DoubleRep -> MachDouble 0
965 FloatRep -> MachFloat 0
966 AddrRep | getPrimRepSize AddrRep == getPrimRepSize WordRep -> MachWord 0
967 _ -> moan64 "mkDummyLiteral" (ppr pr)
971 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
972 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
975 -- and check that an unboxed pair is returned wherein the first arg is VoidRep'd.
977 -- Alternatively, for call-targets returning nothing, convert
979 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
980 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
984 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
985 maybe_getCCallReturnRep fn_ty
986 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
988 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
990 = case splitTyConApp_maybe (repType r_ty) of
991 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
993 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
994 || r_reps == [VoidRep] )
995 && isUnboxedTupleTyCon r_tycon
996 && case maybe_r_rep_to_go of
998 Just r_rep -> r_rep /= PtrRep
999 -- if it was, it would be impossible
1000 -- to create a valid return value
1001 -- placeholder on the stack
1002 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1005 --trace (showSDoc (ppr (a_reps, r_reps))) $
1006 if ok then maybe_r_rep_to_go else blargh
1008 -- Compile code which expects an unboxed Int on the top of stack,
1009 -- (call it i), and pushes the i'th closure in the supplied list
1010 -- as a consequence.
1011 implement_tagToId :: [Name] -> BcM BCInstrList
1012 implement_tagToId names
1013 = ASSERT( notNull names )
1014 getLabelsBc (length names) `thenBc` \ labels ->
1015 getLabelBc `thenBc` \ label_fail ->
1016 getLabelBc `thenBc` \ label_exit ->
1017 zip4 labels (tail labels ++ [label_fail])
1018 [0 ..] names `bind` \ infos ->
1019 map (mkStep label_exit) infos `bind` \ steps ->
1020 returnBc (concatOL steps
1022 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1024 mkStep l_exit (my_label, next_label, n, name_for_n)
1025 = toOL [LABEL my_label,
1026 TESTEQ_I n next_label,
1031 -- -----------------------------------------------------------------------------
1034 -- Push an atom onto the stack, returning suitable code & number of
1035 -- stack words used.
1037 -- The env p must map each variable to the highest- numbered stack
1038 -- slot for it. For example, if the stack has depth 4 and we
1039 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1040 -- the tag in stack[5], the stack will have depth 6, and p must map v
1041 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1042 -- depth 6 stack has valid words 0 .. 5.
1044 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1046 pushAtom d p (AnnApp f (_, AnnType _))
1047 = pushAtom d p (snd f)
1049 pushAtom d p (AnnNote note e)
1050 = pushAtom d p (snd e)
1052 pushAtom d p (AnnLam x e)
1054 = pushAtom d p (snd e)
1056 pushAtom d p (AnnVar v)
1058 | idPrimRep v == VoidRep
1059 = returnBc (nilOL, 0)
1062 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1064 | Just primop <- isPrimOpId_maybe v
1065 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1067 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1068 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1069 -- d - d_v the number of words between the TOS
1070 -- and the 1st slot of the object
1072 -- d - d_v - 1 the offset from the TOS of the 1st slot
1074 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1077 -- Having found the last slot, we proceed to copy the right number of
1078 -- slots on to the top of the stack.
1080 | otherwise -- v must be a global variable
1082 returnBc (unitOL (PUSH_G (getName v)), sz)
1088 pushAtom d p (AnnLit lit)
1090 MachLabel fs _ -> code CodePtrRep
1091 MachWord w -> code WordRep
1092 MachInt i -> code IntRep
1093 MachFloat r -> code FloatRep
1094 MachDouble r -> code DoubleRep
1095 MachChar c -> code CharRep
1096 MachStr s -> pushStr s
1099 = let size_host_words = getPrimRepSize rep
1100 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1104 = let getMallocvilleAddr
1106 FastString _ l ba ->
1107 -- sigh, a string in the heap is no good to us.
1108 -- We need a static C pointer, since the type of
1109 -- a string literal is Addr#. So, copy the string
1110 -- into C land and remember the pointer so we can
1113 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1114 in ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1115 recordMallocBc ptr `thenBc_`
1117 do memcpy ptr ba (fromIntegral n)
1118 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1121 other -> panic "ByteCodeGen.pushAtom.pushStr"
1123 getMallocvilleAddr `thenBc` \ addr ->
1124 -- Get the addr on the stack, untaggedly
1125 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1128 = pprPanic "ByteCodeGen.pushAtom"
1129 (pprCoreExpr (deAnnotate (undefined, other)))
1131 foreign import ccall unsafe "memcpy"
1132 memcpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1135 -- -----------------------------------------------------------------------------
1136 -- Given a bunch of alts code and their discrs, do the donkey work
1137 -- of making a multiway branch using a switch tree.
1138 -- What a load of hassle!
1140 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1141 -- a hint; generates better code
1142 -- Nothing is always safe
1143 -> [(Discr, BCInstrList)]
1145 mkMultiBranch maybe_ncons raw_ways
1146 = let d_way = filter (isNoDiscr.fst) raw_ways
1147 notd_ways = naturalMergeSortLe
1148 (\w1 w2 -> leAlt (fst w1) (fst w2))
1149 (filter (not.isNoDiscr.fst) raw_ways)
1151 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1152 mkTree [] range_lo range_hi = returnBc the_default
1154 mkTree [val] range_lo range_hi
1155 | range_lo `eqAlt` range_hi
1156 = returnBc (snd val)
1158 = getLabelBc `thenBc` \ label_neq ->
1159 returnBc (mkTestEQ (fst val) label_neq
1161 `appOL` unitOL (LABEL label_neq)
1162 `appOL` the_default))
1164 mkTree vals range_lo range_hi
1165 = let n = length vals `div` 2
1166 vals_lo = take n vals
1167 vals_hi = drop n vals
1168 v_mid = fst (head vals_hi)
1170 getLabelBc `thenBc` \ label_geq ->
1171 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1172 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1173 returnBc (mkTestLT v_mid label_geq
1175 `appOL` unitOL (LABEL label_geq)
1179 = case d_way of [] -> unitOL CASEFAIL
1182 -- None of these will be needed if there are no non-default alts
1183 (mkTestLT, mkTestEQ, init_lo, init_hi)
1185 = panic "mkMultiBranch: awesome foursome"
1187 = case fst (head notd_ways) of {
1188 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1189 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1192 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1193 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1196 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1197 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1200 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1201 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1203 DiscrP algMaxBound )
1206 (algMinBound, algMaxBound)
1207 = case maybe_ncons of
1208 Just n -> (0, n - 1)
1209 Nothing -> (minBound, maxBound)
1211 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1212 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1213 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1214 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1215 NoDiscr `eqAlt` NoDiscr = True
1218 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1219 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1220 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1221 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1222 NoDiscr `leAlt` NoDiscr = True
1225 isNoDiscr NoDiscr = True
1228 dec (DiscrI i) = DiscrI (i-1)
1229 dec (DiscrP i) = DiscrP (i-1)
1230 dec other = other -- not really right, but if you
1231 -- do cases on floating values, you'll get what you deserve
1233 -- same snotty comment applies to the following
1235 minD, maxD :: Double
1241 mkTree notd_ways init_lo init_hi
1244 -- -----------------------------------------------------------------------------
1245 -- Supporting junk for the compilation schemes
1247 -- Describes case alts
1255 instance Outputable Discr where
1256 ppr (DiscrI i) = int i
1257 ppr (DiscrF f) = text (show f)
1258 ppr (DiscrD d) = text (show d)
1259 ppr (DiscrP i) = int i
1260 ppr NoDiscr = text "DEF"
1263 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1264 lookupBCEnv_maybe = lookupFM
1266 idSizeW :: Id -> Int
1267 idSizeW id = getPrimRepSize (typePrimRep (idType id))
1269 unboxedTupleException :: a
1270 unboxedTupleException
1273 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1274 "\tto foreign import/export decls in source. Workaround:\n" ++
1275 "\tcompile this module to a .o file, then restart session."))
1278 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1281 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1282 -- The arguments are returned in *right-to-left* order
1283 splitApp (AnnApp (_,f) (_,a))
1284 | isTypeAtom a = splitApp f
1285 | otherwise = case splitApp f of
1286 (f', as) -> (f', a:as)
1287 splitApp (AnnNote n (_,e)) = splitApp e
1288 splitApp e = (e, [])
1291 isTypeAtom :: AnnExpr' id ann -> Bool
1292 isTypeAtom (AnnType _) = True
1293 isTypeAtom _ = False
1295 isVoidRepAtom :: AnnExpr' id ann -> Bool
1296 isVoidRepAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1297 isVoidRepAtom (AnnNote n (_,e)) = isVoidRepAtom e
1298 isVoidRepAtom _ = False
1300 atomRep :: AnnExpr' Id ann -> PrimRep
1301 atomRep (AnnVar v) = typePrimRep (idType v)
1302 atomRep (AnnLit l) = literalPrimRep l
1303 atomRep (AnnNote n b) = atomRep (snd b)
1304 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1305 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1306 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1308 isPtrAtom :: AnnExpr' Id ann -> Bool
1309 isPtrAtom e = isFollowableRep (atomRep e)
1311 -- Let szsw be the sizes in words of some items pushed onto the stack,
1312 -- which has initial depth d'. Return the values which the stack environment
1313 -- should map these items to.
1314 mkStackOffsets :: Int -> [Int] -> [Int]
1315 mkStackOffsets original_depth szsw
1316 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1318 -- -----------------------------------------------------------------------------
1319 -- The bytecode generator's monad
1323 nextlabel :: Int, -- for generating local labels
1324 malloced :: [Ptr ()] } -- ptrs malloced for current BCO
1325 -- Should be free()d when it is GCd
1327 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1329 ioToBc :: IO a -> BcM a
1330 ioToBc io = BcM $ \st -> do
1334 runBc :: BcM r -> IO (BcM_State, r)
1335 runBc (BcM m) = m (BcM_State 0 [])
1337 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1338 thenBc (BcM expr) cont = BcM $ \st0 -> do
1339 (st1, q) <- expr st0
1344 thenBc_ :: BcM a -> BcM b -> BcM b
1345 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1346 (st1, q) <- expr st0
1347 (st2, r) <- cont st1
1350 returnBc :: a -> BcM a
1351 returnBc result = BcM $ \st -> (return (st, result))
1353 instance Monad BcM where
1358 emitBc :: ([Ptr ()] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1360 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1362 recordMallocBc :: Ptr a -> BcM ()
1364 = BcM $ \st -> return (st{malloced = castPtr a : malloced st}, ())
1366 getLabelBc :: BcM Int
1368 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1370 getLabelsBc :: Int -> BcM [Int]
1372 = BcM $ \st -> let ctr = nextlabel st
1373 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])