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 ( TypeEnv, typeEnvTyCons, typeEnvClasses )
23 import CoreUtils ( exprType )
25 import PprCore ( pprCoreExpr )
26 import Literal ( Literal(..), literalType )
27 import PrimOp ( PrimOp(..) )
28 import CoreFVs ( freeVars )
29 import Type ( isUnLiftedType, splitTyConApp_maybe )
30 import DataCon ( DataCon, dataConTag, fIRST_TAG, dataConTyCon,
31 isUnboxedTupleCon, isNullaryDataCon, dataConWorkId,
33 import TyCon ( tyConFamilySize, isDataTyCon, tyConDataCons,
35 import Class ( Class, classTyCon )
36 import Type ( Type, repType, splitFunTys, dropForAlls, pprType )
38 import DataCon ( dataConRepArity )
39 import Var ( isTyVar )
40 import VarSet ( VarSet, varSetElems )
41 import TysPrim ( arrayPrimTyCon, mutableArrayPrimTyCon,
42 byteArrayPrimTyCon, mutableByteArrayPrimTyCon
44 import CmdLineOpts ( DynFlags, DynFlag(..) )
45 import ErrUtils ( showPass, dumpIfSet_dyn )
46 import Unique ( mkPseudoUniqueE )
47 import FastString ( FastString(..), unpackFS )
48 import Panic ( GhcException(..) )
49 import SMRep ( typeCgRep, arrWordsHdrSize, arrPtrsHdrSize, StgWord,
50 CgRep(..), cgRepSizeW, isFollowableArg, idCgRep )
51 import Bitmap ( intsToReverseBitmap, mkBitmap )
53 import Constants ( wORD_SIZE )
55 import Data.List ( intersperse, sortBy, zip4, zip5, partition )
56 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8 )
57 import Foreign.C ( CInt )
58 import Control.Exception ( throwDyn )
60 import GHC.Exts ( Int(..), ByteArray# )
62 import Control.Monad ( when )
63 import Data.Char ( ord, chr )
65 -- -----------------------------------------------------------------------------
66 -- Generating byte code for a complete module
68 byteCodeGen :: DynFlags
71 -> IO CompiledByteCode
72 byteCodeGen dflags binds type_env
73 = do showPass dflags "ByteCodeGen"
74 let local_tycons = typeEnvTyCons type_env
75 local_classes = typeEnvClasses type_env
76 tycs = local_tycons ++ map classTyCon local_classes
78 let flatBinds = [ (bndr, freeVars rhs)
79 | (bndr, rhs) <- flattenBinds binds]
81 (BcM_State final_ctr mallocd, proto_bcos)
82 <- runBc (mapM schemeTopBind flatBinds)
84 when (notNull mallocd)
85 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
87 dumpIfSet_dyn dflags Opt_D_dump_BCOs
88 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
90 assembleBCOs proto_bcos tycs
92 -- -----------------------------------------------------------------------------
93 -- Generating byte code for an expression
95 -- Returns: (the root BCO for this expression,
96 -- a list of auxilary BCOs resulting from compiling closures)
97 coreExprToBCOs :: DynFlags
100 coreExprToBCOs dflags expr
101 = do showPass dflags "ByteCodeGen"
103 -- create a totally bogus name for the top-level BCO; this
104 -- should be harmless, since it's never used for anything
105 let invented_name = mkSystemName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
106 invented_id = mkLocalId invented_name (panic "invented_id's type")
108 (BcM_State final_ctr mallocd, proto_bco)
109 <- runBc (schemeTopBind (invented_id, freeVars expr))
111 when (notNull mallocd)
112 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
114 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
116 assembleBCO proto_bco
119 -- -----------------------------------------------------------------------------
120 -- Compilation schema for the bytecode generator
122 type BCInstrList = OrdList BCInstr
124 type Sequel = Int -- back off to this depth before ENTER
126 -- Maps Ids to the offset from the stack _base_ so we don't have
127 -- to mess with it after each push/pop.
128 type BCEnv = FiniteMap Id Int -- To find vars on the stack
130 ppBCEnv :: BCEnv -> SDoc
133 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
136 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
137 cmp_snd x y = compare (snd x) (snd y)
139 -- Create a BCO and do a spot of peephole optimisation on the insns
144 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
148 -> Bool -- True <=> is a return point, rather than a function
151 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap
152 is_ret mallocd_blocks
155 protoBCOInstrs = maybe_with_stack_check,
156 protoBCOBitmap = bitmap,
157 protoBCOBitmapSize = bitmap_size,
158 protoBCOArity = arity,
159 protoBCOExpr = origin,
160 protoBCOPtrs = mallocd_blocks
163 -- Overestimate the stack usage (in words) of this BCO,
164 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
165 -- stack check. (The interpreter always does a stack check
166 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
167 -- BCO anyway, so we only need to add an explicit on in the
168 -- (hopefully rare) cases when the (overestimated) stack use
169 -- exceeds iNTERP_STACK_CHECK_THRESH.
170 maybe_with_stack_check
172 -- don't do stack checks at return points;
173 -- everything is aggregated up to the top BCO
174 -- (which must be a function)
175 | stack_overest >= 65535
176 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
178 | stack_overest >= iNTERP_STACK_CHECK_THRESH
179 = STKCHECK stack_overest : peep_d
181 = peep_d -- the supposedly common case
183 stack_overest = sum (map bciStackUse peep_d)
185 -- Merge local pushes
186 peep_d = peep (fromOL instrs_ordlist)
188 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
189 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
190 peep (PUSH_L off1 : PUSH_L off2 : rest)
191 = PUSH_LL off1 (off2-1) : peep rest
197 argBits :: [CgRep] -> [Bool]
200 | isFollowableArg rep = False : argBits args
201 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
203 -- -----------------------------------------------------------------------------
206 -- Compile code for the right-hand side of a top-level binding
208 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
211 schemeTopBind (id, rhs)
212 | Just data_con <- isDataConWorkId_maybe id,
213 isNullaryDataCon data_con
214 = -- Special case for the worker of a nullary data con.
215 -- It'll look like this: Nil = /\a -> Nil a
216 -- If we feed it into schemeR, we'll get
218 -- because mkConAppCode treats nullary constructor applications
219 -- by just re-using the single top-level definition. So
220 -- for the worker itself, we must allocate it directly.
221 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
222 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
225 = schemeR [{- No free variables -}] (id, rhs)
227 -- -----------------------------------------------------------------------------
230 -- Compile code for a right-hand side, to give a BCO that,
231 -- when executed with the free variables and arguments on top of the stack,
232 -- will return with a pointer to the result on top of the stack, after
233 -- removing the free variables and arguments.
235 -- Park the resulting BCO in the monad. Also requires the
236 -- variable to which this value was bound, so as to give the
237 -- resulting BCO a name.
239 schemeR :: [Id] -- Free vars of the RHS, ordered as they
240 -- will appear in the thunk. Empty for
241 -- top-level things, which have no free vars.
242 -> (Id, AnnExpr Id VarSet)
243 -> BcM (ProtoBCO Name)
244 schemeR fvs (nm, rhs)
248 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
249 $$ pprCoreExpr (deAnnotate rhs)
255 = schemeR_wrk fvs nm rhs (collect [] rhs)
257 collect xs (_, AnnNote note e) = collect xs e
258 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
259 collect xs (_, not_lambda) = (reverse xs, not_lambda)
261 schemeR_wrk fvs nm original_body (args, body)
263 all_args = reverse args ++ fvs
264 arity = length all_args
265 -- all_args are the args in reverse order. We're compiling a function
266 -- \fv1..fvn x1..xn -> e
267 -- i.e. the fvs come first
269 szsw_args = map idSizeW all_args
270 szw_args = sum szsw_args
271 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
273 -- make the arg bitmap
274 bits = argBits (reverse (map idCgRep all_args))
275 bitmap_size = length bits
276 bitmap = mkBitmap bits
278 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
279 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
280 arity bitmap_size bitmap False{-not alts-})
283 fvsToEnv :: BCEnv -> VarSet -> [Id]
284 -- Takes the free variables of a right-hand side, and
285 -- delivers an ordered list of the local variables that will
286 -- be captured in the thunk for the RHS
287 -- The BCEnv argument tells which variables are in the local
288 -- environment: these are the ones that should be captured
290 -- The code that constructs the thunk, and the code that executes
291 -- it, have to agree about this layout
292 fvsToEnv p fvs = [v | v <- varSetElems fvs,
293 isId v, -- Could be a type variable
296 -- -----------------------------------------------------------------------------
299 -- Compile code to apply the given expression to the remaining args
300 -- on the stack, returning a HNF.
301 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
303 -- Delegate tail-calls to schemeT.
304 schemeE d s p e@(AnnApp f a)
307 schemeE d s p e@(AnnVar v)
308 | not (isUnLiftedType v_type)
309 = -- Lifted-type thing; push it in the normal way
313 = -- Returning an unlifted value.
314 -- Heave it on the stack, SLIDE, and RETURN.
315 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
316 returnBc (push -- value onto stack
317 `appOL` mkSLIDE szw (d-s) -- clear to sequel
318 `snocOL` RETURN_UBX v_rep) -- go
321 v_rep = typeCgRep v_type
323 schemeE d s p (AnnLit literal)
324 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
325 let l_rep = typeCgRep (literalType literal)
326 in returnBc (push -- value onto stack
327 `appOL` mkSLIDE szw (d-s) -- clear to sequel
328 `snocOL` RETURN_UBX l_rep) -- go
331 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
332 | (AnnVar v, args_r_to_l) <- splitApp rhs,
333 Just data_con <- isDataConWorkId_maybe v,
334 dataConRepArity data_con == length args_r_to_l
335 = -- Special case for a non-recursive let whose RHS is a
336 -- saturatred constructor application.
337 -- Just allocate the constructor and carry on
338 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
339 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
340 returnBc (alloc_code `appOL` body_code)
342 -- General case for let. Generates correct, if inefficient, code in
344 schemeE d s p (AnnLet binds (_,body))
345 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
346 AnnRec xs_n_rhss -> unzip xs_n_rhss
349 fvss = map (fvsToEnv p' . fst) rhss
351 -- Sizes of free vars
352 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
354 -- the arity of each rhs
355 arities = map (length . fst . collect []) rhss
357 -- This p', d' defn is safe because all the items being pushed
358 -- are ptrs, so all have size 1. d' and p' reflect the stack
359 -- after the closures have been allocated in the heap (but not
360 -- filled in), and pointers to them parked on the stack.
361 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
363 zipE = zipEqual "schemeE"
365 -- ToDo: don't build thunks for things with no free variables
366 build_thunk dd [] size bco off
367 = returnBc (PUSH_BCO bco
368 `consOL` unitOL (MKAP (off+size) size))
369 build_thunk dd (fv:fvs) size bco off = do
370 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
371 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off
372 returnBc (push_code `appOL` more_push_code)
374 alloc_code = toOL (zipWith mkAlloc sizes arities)
375 where mkAlloc sz 0 = ALLOC_AP sz
376 mkAlloc sz arity = ALLOC_PAP arity sz
378 compile_bind d' fvs x rhs size off = do
379 bco <- schemeR fvs (x,rhs)
380 build_thunk d' fvs size bco off
383 [ compile_bind d' fvs x rhs size n
384 | (fvs, x, rhs, size, n) <-
385 zip5 fvss xs rhss sizes [n_binds, n_binds-1 .. 1]
388 body_code <- schemeE d' s p' body
389 thunk_codes <- sequence compile_binds
390 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
394 schemeE d s p (AnnCase scrut bndr [(DataAlt dc, [bind1, bind2], rhs)])
395 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
397 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
399 -- case .... of a { DEFAULT -> ... }
400 -- becuse the return convention for both are identical.
402 -- Note that it does not matter losing the void-rep thing from the
403 -- envt (it won't be bound now) because we never look such things up.
405 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
406 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
408 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
409 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
410 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
412 schemeE d s p (AnnCase scrut bndr [(DataAlt dc, [bind1], rhs)])
413 | isUnboxedTupleCon dc
414 -- Similarly, convert
415 -- case .... of x { (# a #) -> ... }
417 -- case .... of a { DEFAULT -> ... }
418 = --trace "automagic mashing of case alts (# a #)" $
419 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
421 schemeE d s p (AnnCase scrut bndr alts)
422 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
424 schemeE d s p (AnnNote note (_, body))
428 = pprPanic "ByteCodeGen.schemeE: unhandled case"
429 (pprCoreExpr (deAnnotate' other))
432 -- Compile code to do a tail call. Specifically, push the fn,
433 -- slide the on-stack app back down to the sequel depth,
434 -- and enter. Four cases:
437 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
438 -- The int will be on the stack. Generate a code sequence
439 -- to convert it to the relevant constructor, SLIDE and ENTER.
441 -- 1. The fn denotes a ccall. Defer to generateCCall.
443 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
444 -- it simply as b -- since the representations are identical
445 -- (the VoidArg takes up zero stack space). Also, spot
446 -- (# b #) and treat it as b.
448 -- 3. Application of a constructor, by defn saturated.
449 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
450 -- then the ptrs, and then do PACK and RETURN.
452 -- 4. Otherwise, it must be a function call. Push the args
453 -- right to left, SLIDE and ENTER.
455 schemeT :: Int -- Stack depth
456 -> Sequel -- Sequel depth
457 -> BCEnv -- stack env
458 -> AnnExpr' Id VarSet
463 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
464 -- = panic "schemeT ?!?!"
466 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
470 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
471 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
472 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
473 returnBc (push `appOL` tagToId_sequence
474 `appOL` mkSLIDE 1 (d+arg_words-s)
478 | Just (CCall ccall_spec) <- isFCallId_maybe fn
479 = generateCCall d s p ccall_spec fn args_r_to_l
481 -- Case 2: Constructor application
482 | Just con <- maybe_saturated_dcon,
483 isUnboxedTupleCon con
484 = case args_r_to_l of
485 [arg1,arg2] | isVoidArgAtom arg1 ->
486 unboxedTupleReturn d s p arg2
487 [arg1,arg2] | isVoidArgAtom arg2 ->
488 unboxedTupleReturn d s p arg1
489 _other -> unboxedTupleException
491 -- Case 3: Ordinary data constructor
492 | Just con <- maybe_saturated_dcon
493 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
494 returnBc (alloc_con `appOL`
495 mkSLIDE 1 (d - s) `snocOL`
498 -- Case 4: Tail call of function
500 = doTailCall d s p fn args_r_to_l
503 -- Detect and extract relevant info for the tagToEnum kludge.
504 maybe_is_tagToEnum_call
505 = let extract_constr_Names ty
506 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
508 = map (getName . dataConWorkId) (tyConDataCons tyc)
509 -- NOTE: use the worker name, not the source name of
510 -- the DataCon. See DataCon.lhs for details.
512 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
515 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
516 -> case isPrimOpId_maybe v of
517 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
521 -- Extract the args (R->L) and fn
522 -- The function will necessarily be a variable,
523 -- because we are compiling a tail call
524 (AnnVar fn, args_r_to_l) = splitApp app
526 -- Only consider this to be a constructor application iff it is
527 -- saturated. Otherwise, we'll call the constructor wrapper.
528 n_args = length args_r_to_l
530 = case isDataConWorkId_maybe fn of
531 Just con | dataConRepArity con == n_args -> Just con
534 -- -----------------------------------------------------------------------------
535 -- Generate code to build a constructor application,
536 -- leaving it on top of the stack
538 mkConAppCode :: Int -> Sequel -> BCEnv
539 -> DataCon -- The data constructor
540 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
543 mkConAppCode orig_d s p con [] -- Nullary constructor
544 = ASSERT( isNullaryDataCon con )
545 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
546 -- Instead of doing a PACK, which would allocate a fresh
547 -- copy of this constructor, use the single shared version.
549 mkConAppCode orig_d s p con args_r_to_l
550 = ASSERT( dataConRepArity con == length args_r_to_l )
551 do_pushery orig_d (non_ptr_args ++ ptr_args)
553 -- The args are already in reverse order, which is the way PACK
554 -- expects them to be. We must push the non-ptrs after the ptrs.
555 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
557 do_pushery d (arg:args)
558 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
559 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
560 returnBc (push `appOL` more_push_code)
562 = returnBc (unitOL (PACK con n_arg_words))
564 n_arg_words = d - orig_d
567 -- -----------------------------------------------------------------------------
568 -- Returning an unboxed tuple with one non-void component (the only
569 -- case we can handle).
571 -- Remember, we don't want to *evaluate* the component that is being
572 -- returned, even if it is a pointed type. We always just return.
575 :: Int -> Sequel -> BCEnv
576 -> AnnExpr' Id VarSet -> BcM BCInstrList
577 unboxedTupleReturn d s p arg = do
578 (push, sz) <- pushAtom d p arg
579 returnBc (push `appOL`
580 mkSLIDE sz (d-s) `snocOL`
581 RETURN_UBX (atomRep arg))
583 -- -----------------------------------------------------------------------------
584 -- Generate code for a tail-call
587 :: Int -> Sequel -> BCEnv
588 -> Id -> [AnnExpr' Id VarSet]
590 doTailCall init_d s p fn args
591 = do_pushes init_d args (map atomRep args)
593 do_pushes d [] reps = do
595 (push_fn, sz) <- pushAtom d p (AnnVar fn)
597 returnBc (push_fn `appOL` (
598 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
600 do_pushes d args reps = do
601 let (push_apply, n, rest_of_reps) = findPushSeq reps
602 (these_args, rest_of_args) = splitAt n args
603 (next_d, push_code) <- push_seq d these_args
604 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
605 -- ^^^ for the PUSH_APPLY_ instruction
606 returnBc (push_code `appOL` (push_apply `consOL` instrs))
608 push_seq d [] = return (d, nilOL)
609 push_seq d (arg:args) = do
610 (push_code, sz) <- pushAtom d p arg
611 (final_d, more_push_code) <- push_seq (d+sz) args
612 return (final_d, push_code `appOL` more_push_code)
614 -- v. similar to CgStackery.findMatch, ToDo: merge
615 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
616 = (PUSH_APPLY_PPPPPPP, 7, rest)
617 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
618 = (PUSH_APPLY_PPPPPP, 6, rest)
619 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
620 = (PUSH_APPLY_PPPPP, 5, rest)
621 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
622 = (PUSH_APPLY_PPPP, 4, rest)
623 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
624 = (PUSH_APPLY_PPP, 3, rest)
625 findPushSeq (PtrArg: PtrArg: rest)
626 = (PUSH_APPLY_PP, 2, rest)
627 findPushSeq (PtrArg: rest)
628 = (PUSH_APPLY_P, 1, rest)
629 findPushSeq (VoidArg: rest)
630 = (PUSH_APPLY_V, 1, rest)
631 findPushSeq (NonPtrArg: rest)
632 = (PUSH_APPLY_N, 1, rest)
633 findPushSeq (FloatArg: rest)
634 = (PUSH_APPLY_F, 1, rest)
635 findPushSeq (DoubleArg: rest)
636 = (PUSH_APPLY_D, 1, rest)
637 findPushSeq (LongArg: rest)
638 = (PUSH_APPLY_L, 1, rest)
640 = panic "ByteCodeGen.findPushSeq"
642 -- -----------------------------------------------------------------------------
645 doCase :: Int -> Sequel -> BCEnv
646 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
647 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
649 doCase d s p (_,scrut)
650 bndr alts is_unboxed_tuple
652 -- Top of stack is the return itbl, as usual.
653 -- underneath it is the pointer to the alt_code BCO.
654 -- When an alt is entered, it assumes the returned value is
655 -- on top of the itbl.
658 -- An unlifted value gets an extra info table pushed on top
659 -- when it is returned.
660 unlifted_itbl_sizeW | isAlgCase = 0
663 -- depth of stack after the return value has been pushed
664 d_bndr = d + ret_frame_sizeW + idSizeW bndr
666 -- depth of stack after the extra info table for an unboxed return
667 -- has been pushed, if any. This is the stack depth at the
669 d_alts = d_bndr + unlifted_itbl_sizeW
671 -- Env in which to compile the alts, not including
672 -- any vars bound by the alts themselves
673 p_alts = addToFM p bndr (d_bndr - 1)
675 bndr_ty = idType bndr
676 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
678 -- given an alt, return a discr and code for it.
679 codeALt alt@(DEFAULT, _, (_,rhs))
680 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
681 returnBc (NoDiscr, rhs_code)
682 codeAlt alt@(discr, bndrs, (_,rhs))
683 -- primitive or nullary constructor alt: no need to UNPACK
684 | null real_bndrs = do
685 rhs_code <- schemeE d_alts s p_alts rhs
686 returnBc (my_discr alt, rhs_code)
687 -- algebraic alt with some binders
688 | ASSERT(isAlgCase) otherwise =
690 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
691 ptr_sizes = map idSizeW ptrs
692 nptrs_sizes = map idSizeW nptrs
693 bind_sizes = ptr_sizes ++ nptrs_sizes
694 size = sum ptr_sizes + sum nptrs_sizes
695 -- the UNPACK instruction unpacks in reverse order...
696 p' = addListToFM p_alts
697 (zip (reverse (ptrs ++ nptrs))
698 (mkStackOffsets d_alts (reverse bind_sizes)))
700 rhs_code <- schemeE (d_alts+size) s p' rhs
701 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
703 real_bndrs = filter (not.isTyVar) bndrs
706 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
707 my_discr (DataAlt dc, binds, rhs)
708 | isUnboxedTupleCon dc
709 = unboxedTupleException
711 = DiscrP (dataConTag dc - fIRST_TAG)
712 my_discr (LitAlt l, binds, rhs)
713 = case l of MachInt i -> DiscrI (fromInteger i)
714 MachFloat r -> DiscrF (fromRational r)
715 MachDouble r -> DiscrD (fromRational r)
716 MachChar i -> DiscrI (ord i)
717 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
720 | not isAlgCase = Nothing
722 = case [dc | (DataAlt dc, _, _) <- alts] of
724 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
726 -- the bitmap is relative to stack depth d, i.e. before the
727 -- BCO, info table and return value are pushed on.
728 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
729 -- except that here we build the bitmap from the known bindings of
730 -- things that are pointers, whereas in CgBindery the code builds the
731 -- bitmap from the free slots and unboxed bindings.
733 bitmap = intsToReverseBitmap d{-size-} (sortLt (<) rel_slots)
736 rel_slots = concat (map spread binds)
738 | isFollowableArg (idCgRep id) = [ rel_offset ]
740 where rel_offset = d - offset - 1
743 alt_stuff <- mapM codeAlt alts
744 alt_final <- mkMultiBranch maybe_ncons alt_stuff
746 alt_bco_name = getName bndr
747 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
748 0{-no arity-} d{-bitmap size-} bitmap True{-is alts-}
750 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
751 -- "\n bitmap = " ++ show bitmap) $ do
752 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
753 alt_bco' <- emitBc alt_bco
755 | isAlgCase = PUSH_ALTS alt_bco'
756 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
757 returnBc (push_alts `consOL` scrut_code)
760 -- -----------------------------------------------------------------------------
761 -- Deal with a CCall.
763 -- Taggedly push the args onto the stack R->L,
764 -- deferencing ForeignObj#s and adjusting addrs to point to
765 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
766 -- (machine) code for the ccall, and create bytecodes to call that and
767 -- then return in the right way.
769 generateCCall :: Int -> Sequel -- stack and sequel depths
771 -> CCallSpec -- where to call
772 -> Id -- of target, for type info
773 -> [AnnExpr' Id VarSet] -- args (atoms)
776 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
779 addr_sizeW = cgRepSizeW NonPtrArg
781 -- Get the args on the stack, with tags and suitably
782 -- dereferenced for the CCall. For each arg, return the
783 -- depth to the first word of the bits for that arg, and the
784 -- CgRep of what was actually pushed.
786 pargs d [] = returnBc []
788 = let arg_ty = repType (exprType (deAnnotate' a))
790 in case splitTyConApp_maybe arg_ty of
791 -- Don't push the FO; instead push the Addr# it
794 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
795 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
796 parg_ArrayishRep arrPtrsHdrSize d p a
798 returnBc ((code,NonPtrArg):rest)
800 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
801 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
802 parg_ArrayishRep arrWordsHdrSize d p a
804 returnBc ((code,NonPtrArg):rest)
806 -- Default case: push taggedly, but otherwise intact.
808 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
809 pargs (d+sz_a) az `thenBc` \ rest ->
810 returnBc ((code_a, atomRep a) : rest)
812 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
813 -- the stack but then advance it over the headers, so as to
814 -- point to the payload.
815 parg_ArrayishRep hdrSize d p a
816 = pushAtom d p a `thenBc` \ (push_fo, _) ->
817 -- The ptr points at the header. Advance it over the
818 -- header and then pretend this is an Addr#.
819 returnBc (push_fo `snocOL` SWIZZLE 0 hdrSize)
822 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
824 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
826 push_args = concatOL pushs_arg
827 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
829 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
830 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
832 = reverse (tail a_reps_pushed_r_to_l)
834 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
835 -- push_args is the code to do that.
836 -- d_after_args is the stack depth once the args are on.
838 -- Get the result rep.
839 (returns_void, r_rep)
840 = case maybe_getCCallReturnRep (idType fn) of
841 Nothing -> (True, VoidArg)
842 Just rr -> (False, rr)
844 Because the Haskell stack grows down, the a_reps refer to
845 lowest to highest addresses in that order. The args for the call
846 are on the stack. Now push an unboxed Addr# indicating
847 the C function to call. Then push a dummy placeholder for the
848 result. Finally, emit a CCALL insn with an offset pointing to the
849 Addr# just pushed, and a literal field holding the mallocville
850 address of the piece of marshalling code we generate.
851 So, just prior to the CCALL insn, the stack looks like this
852 (growing down, as usual):
857 Addr# address_of_C_fn
858 <placeholder-for-result#> (must be an unboxed type)
860 The interpreter then calls the marshall code mentioned
861 in the CCALL insn, passing it (& <placeholder-for-result#>),
862 that is, the addr of the topmost word in the stack.
863 When this returns, the placeholder will have been
864 filled in. The placeholder is slid down to the sequel
865 depth, and we RETURN.
867 This arrangement makes it simple to do f-i-dynamic since the Addr#
868 value is the first arg anyway.
870 The marshalling code is generated specifically for this
871 call site, and so knows exactly the (Haskell) stack
872 offsets of the args, fn address and placeholder. It
873 copies the args to the C stack, calls the stacked addr,
874 and parks the result back in the placeholder. The interpreter
875 calls it as a normal C call, assuming it has a signature
876 void marshall_code ( StgWord* ptr_to_top_of_stack )
878 -- resolve static address
882 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
884 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
887 get_target_info `thenBc` \ (is_static, static_target_addr) ->
890 -- Get the arg reps, zapping the leading Addr# in the dynamic case
891 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
892 | is_static = a_reps_pushed_RAW
893 | otherwise = if null a_reps_pushed_RAW
894 then panic "ByteCodeGen.generateCCall: dyn with no args"
895 else tail a_reps_pushed_RAW
898 (push_Addr, d_after_Addr)
900 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
901 d_after_args + addr_sizeW)
902 | otherwise -- is already on the stack
903 = (nilOL, d_after_args)
905 -- Push the return placeholder. For a call returning nothing,
906 -- this is a VoidArg (tag).
907 r_sizeW = cgRepSizeW r_rep
908 d_after_r = d_after_Addr + r_sizeW
909 r_lit = mkDummyLiteral r_rep
910 push_r = (if returns_void
912 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
914 -- generate the marshalling code we're going to call
917 arg1_offW = r_sizeW + addr_sizeW
918 args_offW = map (arg1_offW +)
919 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
921 ioToBc (mkMarshalCode cconv
922 (r_offW, r_rep) addr_offW
923 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
924 recordMallocBc addr_of_marshaller `thenBc_`
926 -- Offset of the next stack frame down the stack. The CCALL
927 -- instruction needs to describe the chunk of stack containing
928 -- the ccall args to the GC, so it needs to know how large it
929 -- is. See comment in Interpreter.c with the CCALL instruction.
930 stk_offset = d_after_r - s
933 do_call = unitOL (CCALL stk_offset (castPtr addr_of_marshaller))
935 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
936 `snocOL` RETURN_UBX r_rep
938 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
941 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
945 -- Make a dummy literal, to be used as a placeholder for FFI return
946 -- values on the stack.
947 mkDummyLiteral :: CgRep -> Literal
950 NonPtrArg -> MachWord 0
951 DoubleArg -> MachDouble 0
952 FloatArg -> MachFloat 0
953 _ -> moan64 "mkDummyLiteral" (ppr pr)
957 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
958 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
961 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
963 -- Alternatively, for call-targets returning nothing, convert
965 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
966 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
970 maybe_getCCallReturnRep :: Type -> Maybe CgRep
971 maybe_getCCallReturnRep fn_ty
972 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
974 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
976 = case splitTyConApp_maybe (repType r_ty) of
977 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
979 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
980 || r_reps == [VoidArg] )
981 && isUnboxedTupleTyCon r_tycon
982 && case maybe_r_rep_to_go of
984 Just r_rep -> r_rep /= PtrArg
985 -- if it was, it would be impossible
986 -- to create a valid return value
987 -- placeholder on the stack
988 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
991 --trace (showSDoc (ppr (a_reps, r_reps))) $
992 if ok then maybe_r_rep_to_go else blargh
994 -- Compile code which expects an unboxed Int on the top of stack,
995 -- (call it i), and pushes the i'th closure in the supplied list
997 implement_tagToId :: [Name] -> BcM BCInstrList
998 implement_tagToId names
999 = ASSERT( notNull names )
1000 getLabelsBc (length names) `thenBc` \ labels ->
1001 getLabelBc `thenBc` \ label_fail ->
1002 getLabelBc `thenBc` \ label_exit ->
1003 zip4 labels (tail labels ++ [label_fail])
1004 [0 ..] names `bind` \ infos ->
1005 map (mkStep label_exit) infos `bind` \ steps ->
1006 returnBc (concatOL steps
1008 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1010 mkStep l_exit (my_label, next_label, n, name_for_n)
1011 = toOL [LABEL my_label,
1012 TESTEQ_I n next_label,
1017 -- -----------------------------------------------------------------------------
1020 -- Push an atom onto the stack, returning suitable code & number of
1021 -- stack words used.
1023 -- The env p must map each variable to the highest- numbered stack
1024 -- slot for it. For example, if the stack has depth 4 and we
1025 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1026 -- the tag in stack[5], the stack will have depth 6, and p must map v
1027 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1028 -- depth 6 stack has valid words 0 .. 5.
1030 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1032 pushAtom d p (AnnApp f (_, AnnType _))
1033 = pushAtom d p (snd f)
1035 pushAtom d p (AnnNote note e)
1036 = pushAtom d p (snd e)
1038 pushAtom d p (AnnLam x e)
1040 = pushAtom d p (snd e)
1042 pushAtom d p (AnnVar v)
1044 | idCgRep v == VoidArg
1045 = returnBc (nilOL, 0)
1048 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1050 | Just primop <- isPrimOpId_maybe v
1051 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1053 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1054 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1055 -- d - d_v the number of words between the TOS
1056 -- and the 1st slot of the object
1058 -- d - d_v - 1 the offset from the TOS of the 1st slot
1060 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1063 -- Having found the last slot, we proceed to copy the right number of
1064 -- slots on to the top of the stack.
1066 | otherwise -- v must be a global variable
1068 returnBc (unitOL (PUSH_G (getName v)), sz)
1074 pushAtom d p (AnnLit lit)
1076 MachLabel fs _ -> code NonPtrArg
1077 MachWord w -> code NonPtrArg
1078 MachInt i -> code PtrArg
1079 MachFloat r -> code FloatArg
1080 MachDouble r -> code DoubleArg
1081 MachChar c -> code NonPtrArg
1082 MachStr s -> pushStr s
1085 = let size_host_words = cgRepSizeW rep
1086 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1090 = let getMallocvilleAddr
1092 FastString _ l ba ->
1093 -- sigh, a string in the heap is no good to us.
1094 -- We need a static C pointer, since the type of
1095 -- a string literal is Addr#. So, copy the string
1096 -- into C land and remember the pointer so we can
1099 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1100 in ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1101 recordMallocBc ptr `thenBc_`
1103 do memcpy ptr ba (fromIntegral n)
1104 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1107 other -> panic "ByteCodeGen.pushAtom.pushStr"
1109 getMallocvilleAddr `thenBc` \ addr ->
1110 -- Get the addr on the stack, untaggedly
1111 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1114 = pprPanic "ByteCodeGen.pushAtom"
1115 (pprCoreExpr (deAnnotate (undefined, other)))
1117 foreign import ccall unsafe "memcpy"
1118 memcpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1121 -- -----------------------------------------------------------------------------
1122 -- Given a bunch of alts code and their discrs, do the donkey work
1123 -- of making a multiway branch using a switch tree.
1124 -- What a load of hassle!
1126 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1127 -- a hint; generates better code
1128 -- Nothing is always safe
1129 -> [(Discr, BCInstrList)]
1131 mkMultiBranch maybe_ncons raw_ways
1132 = let d_way = filter (isNoDiscr.fst) raw_ways
1133 notd_ways = naturalMergeSortLe
1134 (\w1 w2 -> leAlt (fst w1) (fst w2))
1135 (filter (not.isNoDiscr.fst) raw_ways)
1137 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1138 mkTree [] range_lo range_hi = returnBc the_default
1140 mkTree [val] range_lo range_hi
1141 | range_lo `eqAlt` range_hi
1142 = returnBc (snd val)
1144 = getLabelBc `thenBc` \ label_neq ->
1145 returnBc (mkTestEQ (fst val) label_neq
1147 `appOL` unitOL (LABEL label_neq)
1148 `appOL` the_default))
1150 mkTree vals range_lo range_hi
1151 = let n = length vals `div` 2
1152 vals_lo = take n vals
1153 vals_hi = drop n vals
1154 v_mid = fst (head vals_hi)
1156 getLabelBc `thenBc` \ label_geq ->
1157 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1158 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1159 returnBc (mkTestLT v_mid label_geq
1161 `appOL` unitOL (LABEL label_geq)
1165 = case d_way of [] -> unitOL CASEFAIL
1168 -- None of these will be needed if there are no non-default alts
1169 (mkTestLT, mkTestEQ, init_lo, init_hi)
1171 = panic "mkMultiBranch: awesome foursome"
1173 = case fst (head notd_ways) of {
1174 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1175 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1178 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1179 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1182 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1183 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1186 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1187 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1189 DiscrP algMaxBound )
1192 (algMinBound, algMaxBound)
1193 = case maybe_ncons of
1194 Just n -> (0, n - 1)
1195 Nothing -> (minBound, maxBound)
1197 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1198 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1199 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1200 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1201 NoDiscr `eqAlt` NoDiscr = True
1204 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1205 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1206 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1207 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1208 NoDiscr `leAlt` NoDiscr = True
1211 isNoDiscr NoDiscr = True
1214 dec (DiscrI i) = DiscrI (i-1)
1215 dec (DiscrP i) = DiscrP (i-1)
1216 dec other = other -- not really right, but if you
1217 -- do cases on floating values, you'll get what you deserve
1219 -- same snotty comment applies to the following
1221 minD, maxD :: Double
1227 mkTree notd_ways init_lo init_hi
1230 -- -----------------------------------------------------------------------------
1231 -- Supporting junk for the compilation schemes
1233 -- Describes case alts
1241 instance Outputable Discr where
1242 ppr (DiscrI i) = int i
1243 ppr (DiscrF f) = text (show f)
1244 ppr (DiscrD d) = text (show d)
1245 ppr (DiscrP i) = int i
1246 ppr NoDiscr = text "DEF"
1249 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1250 lookupBCEnv_maybe = lookupFM
1252 idSizeW :: Id -> Int
1253 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1255 unboxedTupleException :: a
1256 unboxedTupleException
1259 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1260 "\tto foreign import/export decls in source. Workaround:\n" ++
1261 "\tcompile this module to a .o file, then restart session."))
1264 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1267 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1268 -- The arguments are returned in *right-to-left* order
1269 splitApp (AnnApp (_,f) (_,a))
1270 | isTypeAtom a = splitApp f
1271 | otherwise = case splitApp f of
1272 (f', as) -> (f', a:as)
1273 splitApp (AnnNote n (_,e)) = splitApp e
1274 splitApp e = (e, [])
1277 isTypeAtom :: AnnExpr' id ann -> Bool
1278 isTypeAtom (AnnType _) = True
1279 isTypeAtom _ = False
1281 isVoidArgAtom :: AnnExpr' id ann -> Bool
1282 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1283 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1284 isVoidArgAtom _ = False
1286 atomRep :: AnnExpr' Id ann -> CgRep
1287 atomRep (AnnVar v) = typeCgRep (idType v)
1288 atomRep (AnnLit l) = typeCgRep (literalType l)
1289 atomRep (AnnNote n b) = atomRep (snd b)
1290 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1291 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1292 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1294 isPtrAtom :: AnnExpr' Id ann -> Bool
1295 isPtrAtom e = atomRep e == PtrArg
1297 -- Let szsw be the sizes in words of some items pushed onto the stack,
1298 -- which has initial depth d'. Return the values which the stack environment
1299 -- should map these items to.
1300 mkStackOffsets :: Int -> [Int] -> [Int]
1301 mkStackOffsets original_depth szsw
1302 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1304 -- -----------------------------------------------------------------------------
1305 -- The bytecode generator's monad
1309 nextlabel :: Int, -- for generating local labels
1310 malloced :: [Ptr ()] } -- ptrs malloced for current BCO
1311 -- Should be free()d when it is GCd
1313 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1315 ioToBc :: IO a -> BcM a
1316 ioToBc io = BcM $ \st -> do
1320 runBc :: BcM r -> IO (BcM_State, r)
1321 runBc (BcM m) = m (BcM_State 0 [])
1323 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1324 thenBc (BcM expr) cont = BcM $ \st0 -> do
1325 (st1, q) <- expr st0
1330 thenBc_ :: BcM a -> BcM b -> BcM b
1331 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1332 (st1, q) <- expr st0
1333 (st2, r) <- cont st1
1336 returnBc :: a -> BcM a
1337 returnBc result = BcM $ \st -> (return (st, result))
1339 instance Monad BcM where
1344 emitBc :: ([Ptr ()] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1346 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1348 recordMallocBc :: Ptr a -> BcM ()
1350 = BcM $ \st -> return (st{malloced = castPtr a : malloced st}, ())
1352 getLabelBc :: BcM Int
1354 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1356 getLabelsBc :: Int -> BcM [Int]
1358 = BcM $ \st -> let ctr = nextlabel st
1359 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])