2 % (c) The University of Glasgow 2002
4 \section[ByteCodeGen]{Generate bytecode from Core}
7 module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where
9 #include "HsVersions.h"
12 import ByteCodeFFI ( mkMarshalCode, moan64 )
13 import ByteCodeAsm ( CompiledByteCode(..), UnlinkedBCO,
14 assembleBCO, assembleBCOs, iNTERP_STACK_CHECK_THRESH )
15 import ByteCodeLink ( lookupStaticPtr )
18 import Name ( Name, getName, mkSystemVarName )
21 import ForeignCall ( ForeignCall(..), CCallTarget(..), CCallSpec(..) )
22 import HscTypes ( TypeEnv, typeEnvTyCons, typeEnvClasses )
23 import CoreUtils ( exprType )
25 import PprCore ( pprCoreExpr )
26 import Literal ( Literal(..), literalType )
27 import PrimOp ( PrimOp(..) )
28 import CoreFVs ( freeVars )
29 import Type ( isUnLiftedType, splitTyConApp_maybe )
30 import DataCon ( DataCon, dataConTag, fIRST_TAG, dataConTyCon,
31 isUnboxedTupleCon, isNullaryRepDataCon, dataConWorkId,
33 import TyCon ( TyCon, tyConFamilySize, isDataTyCon,
34 tyConDataCons, isUnboxedTupleTyCon )
35 import Class ( Class, classTyCon )
36 import Type ( Type, repType, splitFunTys, dropForAlls, pprType )
38 import DataCon ( dataConRepArity )
39 import Var ( isTyVar )
40 import VarSet ( VarSet, varSetElems )
41 import TysPrim ( arrayPrimTyCon, mutableArrayPrimTyCon,
42 byteArrayPrimTyCon, mutableByteArrayPrimTyCon
44 import DynFlags ( DynFlags, DynFlag(..) )
45 import ErrUtils ( showPass, dumpIfSet_dyn )
46 import Unique ( mkPseudoUniqueE )
47 import FastString ( FastString(..), unpackFS )
48 import Panic ( GhcException(..) )
49 import SMRep ( typeCgRep, arrWordsHdrSize, arrPtrsHdrSize, StgWord,
50 CgRep(..), cgRepSizeW, isFollowableArg, idCgRep )
51 import Bitmap ( intsToReverseBitmap, mkBitmap )
53 import Constants ( wORD_SIZE )
55 import Data.List ( intersperse, sortBy, zip4, zip6, partition )
56 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8,
58 import Foreign.C ( CInt )
59 import Control.Exception ( throwDyn )
61 import GHC.Exts ( Int(..), ByteArray# )
63 import Control.Monad ( when )
64 import Data.Char ( ord, chr )
66 -- -----------------------------------------------------------------------------
67 -- Generating byte code for a complete module
69 byteCodeGen :: DynFlags
72 -> IO CompiledByteCode
73 byteCodeGen dflags binds tycs
74 = do showPass dflags "ByteCodeGen"
76 let flatBinds = [ (bndr, freeVars rhs)
77 | (bndr, rhs) <- flattenBinds binds]
79 (BcM_State final_ctr mallocd, proto_bcos)
80 <- runBc (mapM schemeTopBind flatBinds)
82 when (notNull mallocd)
83 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
85 dumpIfSet_dyn dflags Opt_D_dump_BCOs
86 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
88 assembleBCOs proto_bcos tycs
90 -- -----------------------------------------------------------------------------
91 -- Generating byte code for an expression
93 -- Returns: (the root BCO for this expression,
94 -- a list of auxilary BCOs resulting from compiling closures)
95 coreExprToBCOs :: DynFlags
98 coreExprToBCOs dflags expr
99 = do showPass dflags "ByteCodeGen"
101 -- create a totally bogus name for the top-level BCO; this
102 -- should be harmless, since it's never used for anything
103 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
104 invented_id = mkLocalId invented_name (panic "invented_id's type")
106 (BcM_State final_ctr mallocd, proto_bco)
107 <- runBc (schemeTopBind (invented_id, freeVars expr))
109 when (notNull mallocd)
110 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
112 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
114 assembleBCO proto_bco
117 -- -----------------------------------------------------------------------------
118 -- Compilation schema for the bytecode generator
120 type BCInstrList = OrdList BCInstr
122 type Sequel = Int -- back off to this depth before ENTER
124 -- Maps Ids to the offset from the stack _base_ so we don't have
125 -- to mess with it after each push/pop.
126 type BCEnv = FiniteMap Id Int -- To find vars on the stack
128 ppBCEnv :: BCEnv -> SDoc
131 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
134 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
135 cmp_snd x y = compare (snd x) (snd y)
137 -- Create a BCO and do a spot of peephole optimisation on the insns
142 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
146 -> Bool -- True <=> is a return point, rather than a function
149 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap
150 is_ret mallocd_blocks
153 protoBCOInstrs = maybe_with_stack_check,
154 protoBCOBitmap = bitmap,
155 protoBCOBitmapSize = bitmap_size,
156 protoBCOArity = arity,
157 protoBCOExpr = origin,
158 protoBCOPtrs = mallocd_blocks
161 -- Overestimate the stack usage (in words) of this BCO,
162 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
163 -- stack check. (The interpreter always does a stack check
164 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
165 -- BCO anyway, so we only need to add an explicit on in the
166 -- (hopefully rare) cases when the (overestimated) stack use
167 -- exceeds iNTERP_STACK_CHECK_THRESH.
168 maybe_with_stack_check
170 -- don't do stack checks at return points;
171 -- everything is aggregated up to the top BCO
172 -- (which must be a function)
173 | stack_overest >= 65535
174 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
176 | stack_overest >= iNTERP_STACK_CHECK_THRESH
177 = STKCHECK stack_overest : peep_d
179 = peep_d -- the supposedly common case
181 stack_overest = sum (map bciStackUse peep_d)
183 -- Merge local pushes
184 peep_d = peep (fromOL instrs_ordlist)
186 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
187 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
188 peep (PUSH_L off1 : PUSH_L off2 : rest)
189 = PUSH_LL off1 (off2-1) : peep rest
195 argBits :: [CgRep] -> [Bool]
198 | isFollowableArg rep = False : argBits args
199 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
201 -- -----------------------------------------------------------------------------
204 -- Compile code for the right-hand side of a top-level binding
206 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
209 schemeTopBind (id, rhs)
210 | Just data_con <- isDataConWorkId_maybe id,
211 isNullaryRepDataCon data_con
212 = -- Special case for the worker of a nullary data con.
213 -- It'll look like this: Nil = /\a -> Nil a
214 -- If we feed it into schemeR, we'll get
216 -- because mkConAppCode treats nullary constructor applications
217 -- by just re-using the single top-level definition. So
218 -- for the worker itself, we must allocate it directly.
219 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
220 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
223 = schemeR [{- No free variables -}] (id, rhs)
225 -- -----------------------------------------------------------------------------
228 -- Compile code for a right-hand side, to give a BCO that,
229 -- when executed with the free variables and arguments on top of the stack,
230 -- will return with a pointer to the result on top of the stack, after
231 -- removing the free variables and arguments.
233 -- Park the resulting BCO in the monad. Also requires the
234 -- variable to which this value was bound, so as to give the
235 -- resulting BCO a name.
237 schemeR :: [Id] -- Free vars of the RHS, ordered as they
238 -- will appear in the thunk. Empty for
239 -- top-level things, which have no free vars.
240 -> (Id, AnnExpr Id VarSet)
241 -> BcM (ProtoBCO Name)
242 schemeR fvs (nm, rhs)
246 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
247 $$ pprCoreExpr (deAnnotate rhs)
253 = schemeR_wrk fvs nm rhs (collect [] rhs)
255 collect xs (_, AnnNote note e) = collect xs e
256 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
257 collect xs (_, not_lambda) = (reverse xs, not_lambda)
259 schemeR_wrk fvs nm original_body (args, body)
261 all_args = reverse args ++ fvs
262 arity = length all_args
263 -- all_args are the args in reverse order. We're compiling a function
264 -- \fv1..fvn x1..xn -> e
265 -- i.e. the fvs come first
267 szsw_args = map idSizeW all_args
268 szw_args = sum szsw_args
269 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
271 -- make the arg bitmap
272 bits = argBits (reverse (map idCgRep all_args))
273 bitmap_size = length bits
274 bitmap = mkBitmap bits
276 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
277 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
278 arity bitmap_size bitmap False{-not alts-})
281 fvsToEnv :: BCEnv -> VarSet -> [Id]
282 -- Takes the free variables of a right-hand side, and
283 -- delivers an ordered list of the local variables that will
284 -- be captured in the thunk for the RHS
285 -- The BCEnv argument tells which variables are in the local
286 -- environment: these are the ones that should be captured
288 -- The code that constructs the thunk, and the code that executes
289 -- it, have to agree about this layout
290 fvsToEnv p fvs = [v | v <- varSetElems fvs,
291 isId v, -- Could be a type variable
294 -- -----------------------------------------------------------------------------
297 -- Compile code to apply the given expression to the remaining args
298 -- on the stack, returning a HNF.
299 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
301 -- Delegate tail-calls to schemeT.
302 schemeE d s p e@(AnnApp f a)
305 schemeE d s p e@(AnnVar v)
306 | not (isUnLiftedType v_type)
307 = -- Lifted-type thing; push it in the normal way
311 = -- Returning an unlifted value.
312 -- Heave it on the stack, SLIDE, and RETURN.
313 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
314 returnBc (push -- value onto stack
315 `appOL` mkSLIDE szw (d-s) -- clear to sequel
316 `snocOL` RETURN_UBX v_rep) -- go
319 v_rep = typeCgRep v_type
321 schemeE d s p (AnnLit literal)
322 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
323 let l_rep = typeCgRep (literalType literal)
324 in returnBc (push -- value onto stack
325 `appOL` mkSLIDE szw (d-s) -- clear to sequel
326 `snocOL` RETURN_UBX l_rep) -- go
329 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
330 | (AnnVar v, args_r_to_l) <- splitApp rhs,
331 Just data_con <- isDataConWorkId_maybe v,
332 dataConRepArity data_con == length args_r_to_l
333 = -- Special case for a non-recursive let whose RHS is a
334 -- saturatred constructor application.
335 -- Just allocate the constructor and carry on
336 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
337 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
338 returnBc (alloc_code `appOL` body_code)
340 -- General case for let. Generates correct, if inefficient, code in
342 schemeE d s p (AnnLet binds (_,body))
343 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
344 AnnRec xs_n_rhss -> unzip xs_n_rhss
347 fvss = map (fvsToEnv p' . fst) rhss
349 -- Sizes of free vars
350 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
352 -- the arity of each rhs
353 arities = map (length . fst . collect []) rhss
355 -- This p', d' defn is safe because all the items being pushed
356 -- are ptrs, so all have size 1. d' and p' reflect the stack
357 -- after the closures have been allocated in the heap (but not
358 -- filled in), and pointers to them parked on the stack.
359 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
361 zipE = zipEqual "schemeE"
363 -- ToDo: don't build thunks for things with no free variables
364 build_thunk dd [] size bco off arity
365 = returnBc (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
367 mkap | arity == 0 = MKAP
369 build_thunk dd (fv:fvs) size bco off arity = do
370 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
371 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
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 arity off = do
379 bco <- schemeR fvs (x,rhs)
380 build_thunk d' fvs size bco off arity
383 [ compile_bind d' fvs x rhs size arity n
384 | (fvs, x, rhs, size, arity, n) <-
385 zip6 fvss xs rhss sizes arities [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))
427 schemeE d s p (AnnCast (_, body) _)
431 = pprPanic "ByteCodeGen.schemeE: unhandled case"
432 (pprCoreExpr (deAnnotate' other))
435 -- Compile code to do a tail call. Specifically, push the fn,
436 -- slide the on-stack app back down to the sequel depth,
437 -- and enter. Four cases:
440 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
441 -- The int will be on the stack. Generate a code sequence
442 -- to convert it to the relevant constructor, SLIDE and ENTER.
444 -- 1. The fn denotes a ccall. Defer to generateCCall.
446 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
447 -- it simply as b -- since the representations are identical
448 -- (the VoidArg takes up zero stack space). Also, spot
449 -- (# b #) and treat it as b.
451 -- 3. Application of a constructor, by defn saturated.
452 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
453 -- then the ptrs, and then do PACK and RETURN.
455 -- 4. Otherwise, it must be a function call. Push the args
456 -- right to left, SLIDE and ENTER.
458 schemeT :: Int -- Stack depth
459 -> Sequel -- Sequel depth
460 -> BCEnv -- stack env
461 -> AnnExpr' Id VarSet
466 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
467 -- = panic "schemeT ?!?!"
469 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
473 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
474 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
475 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
476 returnBc (push `appOL` tagToId_sequence
477 `appOL` mkSLIDE 1 (d+arg_words-s)
481 | Just (CCall ccall_spec) <- isFCallId_maybe fn
482 = generateCCall d s p ccall_spec fn args_r_to_l
484 -- Case 2: Constructor application
485 | Just con <- maybe_saturated_dcon,
486 isUnboxedTupleCon con
487 = case args_r_to_l of
488 [arg1,arg2] | isVoidArgAtom arg1 ->
489 unboxedTupleReturn d s p arg2
490 [arg1,arg2] | isVoidArgAtom arg2 ->
491 unboxedTupleReturn d s p arg1
492 _other -> unboxedTupleException
494 -- Case 3: Ordinary data constructor
495 | Just con <- maybe_saturated_dcon
496 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
497 returnBc (alloc_con `appOL`
498 mkSLIDE 1 (d - s) `snocOL`
501 -- Case 4: Tail call of function
503 = doTailCall d s p fn args_r_to_l
506 -- Detect and extract relevant info for the tagToEnum kludge.
507 maybe_is_tagToEnum_call
508 = let extract_constr_Names ty
509 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
511 = map (getName . dataConWorkId) (tyConDataCons tyc)
512 -- NOTE: use the worker name, not the source name of
513 -- the DataCon. See DataCon.lhs for details.
515 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
518 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
519 -> case isPrimOpId_maybe v of
520 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
524 -- Extract the args (R->L) and fn
525 -- The function will necessarily be a variable,
526 -- because we are compiling a tail call
527 (AnnVar fn, args_r_to_l) = splitApp app
529 -- Only consider this to be a constructor application iff it is
530 -- saturated. Otherwise, we'll call the constructor wrapper.
531 n_args = length args_r_to_l
533 = case isDataConWorkId_maybe fn of
534 Just con | dataConRepArity con == n_args -> Just con
537 -- -----------------------------------------------------------------------------
538 -- Generate code to build a constructor application,
539 -- leaving it on top of the stack
541 mkConAppCode :: Int -> Sequel -> BCEnv
542 -> DataCon -- The data constructor
543 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
546 mkConAppCode orig_d s p con [] -- Nullary constructor
547 = ASSERT( isNullaryRepDataCon con )
548 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
549 -- Instead of doing a PACK, which would allocate a fresh
550 -- copy of this constructor, use the single shared version.
552 mkConAppCode orig_d s p con args_r_to_l
553 = ASSERT( dataConRepArity con == length args_r_to_l )
554 do_pushery orig_d (non_ptr_args ++ ptr_args)
556 -- The args are already in reverse order, which is the way PACK
557 -- expects them to be. We must push the non-ptrs after the ptrs.
558 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
560 do_pushery d (arg:args)
561 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
562 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
563 returnBc (push `appOL` more_push_code)
565 = returnBc (unitOL (PACK con n_arg_words))
567 n_arg_words = d - orig_d
570 -- -----------------------------------------------------------------------------
571 -- Returning an unboxed tuple with one non-void component (the only
572 -- case we can handle).
574 -- Remember, we don't want to *evaluate* the component that is being
575 -- returned, even if it is a pointed type. We always just return.
578 :: Int -> Sequel -> BCEnv
579 -> AnnExpr' Id VarSet -> BcM BCInstrList
580 unboxedTupleReturn d s p arg = do
581 (push, sz) <- pushAtom d p arg
582 returnBc (push `appOL`
583 mkSLIDE sz (d-s) `snocOL`
584 RETURN_UBX (atomRep arg))
586 -- -----------------------------------------------------------------------------
587 -- Generate code for a tail-call
590 :: Int -> Sequel -> BCEnv
591 -> Id -> [AnnExpr' Id VarSet]
593 doTailCall init_d s p fn args
594 = do_pushes init_d args (map atomRep args)
596 do_pushes d [] reps = do
597 ASSERT( null reps ) return ()
598 (push_fn, sz) <- pushAtom d p (AnnVar fn)
599 ASSERT( sz == 1 ) return ()
600 returnBc (push_fn `appOL` (
601 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
603 do_pushes d args reps = do
604 let (push_apply, n, rest_of_reps) = findPushSeq reps
605 (these_args, rest_of_args) = splitAt n args
606 (next_d, push_code) <- push_seq d these_args
607 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
608 -- ^^^ for the PUSH_APPLY_ instruction
609 returnBc (push_code `appOL` (push_apply `consOL` instrs))
611 push_seq d [] = return (d, nilOL)
612 push_seq d (arg:args) = do
613 (push_code, sz) <- pushAtom d p arg
614 (final_d, more_push_code) <- push_seq (d+sz) args
615 return (final_d, push_code `appOL` more_push_code)
617 -- v. similar to CgStackery.findMatch, ToDo: merge
618 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
619 = (PUSH_APPLY_PPPPPP, 6, rest)
620 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
621 = (PUSH_APPLY_PPPPP, 5, rest)
622 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
623 = (PUSH_APPLY_PPPP, 4, rest)
624 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
625 = (PUSH_APPLY_PPP, 3, rest)
626 findPushSeq (PtrArg: PtrArg: rest)
627 = (PUSH_APPLY_PP, 2, rest)
628 findPushSeq (PtrArg: rest)
629 = (PUSH_APPLY_P, 1, rest)
630 findPushSeq (VoidArg: rest)
631 = (PUSH_APPLY_V, 1, rest)
632 findPushSeq (NonPtrArg: rest)
633 = (PUSH_APPLY_N, 1, rest)
634 findPushSeq (FloatArg: rest)
635 = (PUSH_APPLY_F, 1, rest)
636 findPushSeq (DoubleArg: rest)
637 = (PUSH_APPLY_D, 1, rest)
638 findPushSeq (LongArg: rest)
639 = (PUSH_APPLY_L, 1, rest)
641 = panic "ByteCodeGen.findPushSeq"
643 -- -----------------------------------------------------------------------------
646 doCase :: Int -> Sequel -> BCEnv
647 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
648 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
650 doCase d s p (_,scrut)
651 bndr alts is_unboxed_tuple
653 -- Top of stack is the return itbl, as usual.
654 -- underneath it is the pointer to the alt_code BCO.
655 -- When an alt is entered, it assumes the returned value is
656 -- on top of the itbl.
659 -- An unlifted value gets an extra info table pushed on top
660 -- when it is returned.
661 unlifted_itbl_sizeW | isAlgCase = 0
664 -- depth of stack after the return value has been pushed
665 d_bndr = d + ret_frame_sizeW + idSizeW bndr
667 -- depth of stack after the extra info table for an unboxed return
668 -- has been pushed, if any. This is the stack depth at the
670 d_alts = d_bndr + unlifted_itbl_sizeW
672 -- Env in which to compile the alts, not including
673 -- any vars bound by the alts themselves
674 p_alts = addToFM p bndr (d_bndr - 1)
676 bndr_ty = idType bndr
677 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
679 -- given an alt, return a discr and code for it.
680 codeALt alt@(DEFAULT, _, (_,rhs))
681 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
682 returnBc (NoDiscr, rhs_code)
683 codeAlt alt@(discr, bndrs, (_,rhs))
684 -- primitive or nullary constructor alt: no need to UNPACK
685 | null real_bndrs = do
686 rhs_code <- schemeE d_alts s p_alts rhs
687 returnBc (my_discr alt, rhs_code)
688 -- algebraic alt with some binders
689 | ASSERT(isAlgCase) otherwise =
691 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
692 ptr_sizes = map idSizeW ptrs
693 nptrs_sizes = map idSizeW nptrs
694 bind_sizes = ptr_sizes ++ nptrs_sizes
695 size = sum ptr_sizes + sum nptrs_sizes
696 -- the UNPACK instruction unpacks in reverse order...
697 p' = addListToFM p_alts
698 (zip (reverse (ptrs ++ nptrs))
699 (mkStackOffsets d_alts (reverse bind_sizes)))
701 rhs_code <- schemeE (d_alts+size) s p' rhs
702 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
704 real_bndrs = filter (not.isTyVar) bndrs
707 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
708 my_discr (DataAlt dc, binds, rhs)
709 | isUnboxedTupleCon dc
710 = unboxedTupleException
712 = DiscrP (dataConTag dc - fIRST_TAG)
713 my_discr (LitAlt l, binds, rhs)
714 = case l of MachInt i -> DiscrI (fromInteger i)
715 MachFloat r -> DiscrF (fromRational r)
716 MachDouble r -> DiscrD (fromRational r)
717 MachChar i -> DiscrI (ord i)
718 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
721 | not isAlgCase = Nothing
723 = case [dc | (DataAlt dc, _, _) <- alts] of
725 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
727 -- the bitmap is relative to stack depth d, i.e. before the
728 -- BCO, info table and return value are pushed on.
729 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
730 -- except that here we build the bitmap from the known bindings of
731 -- things that are pointers, whereas in CgBindery the code builds the
732 -- bitmap from the free slots and unboxed bindings.
734 bitmap = intsToReverseBitmap d{-size-} (sortLe (<=) rel_slots)
737 rel_slots = concat (map spread binds)
739 | isFollowableArg (idCgRep id) = [ rel_offset ]
741 where rel_offset = d - offset - 1
744 alt_stuff <- mapM codeAlt alts
745 alt_final <- mkMultiBranch maybe_ncons alt_stuff
747 alt_bco_name = getName bndr
748 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
749 0{-no arity-} d{-bitmap size-} bitmap True{-is alts-}
751 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
752 -- "\n bitmap = " ++ show bitmap) $ do
753 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
754 alt_bco' <- emitBc alt_bco
756 | isAlgCase = PUSH_ALTS alt_bco'
757 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
758 returnBc (push_alts `consOL` scrut_code)
761 -- -----------------------------------------------------------------------------
762 -- Deal with a CCall.
764 -- Taggedly push the args onto the stack R->L,
765 -- deferencing ForeignObj#s and adjusting addrs to point to
766 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
767 -- (machine) code for the ccall, and create bytecodes to call that and
768 -- then return in the right way.
770 generateCCall :: Int -> Sequel -- stack and sequel depths
772 -> CCallSpec -- where to call
773 -> Id -- of target, for type info
774 -> [AnnExpr' Id VarSet] -- args (atoms)
777 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
780 addr_sizeW = cgRepSizeW NonPtrArg
782 -- Get the args on the stack, with tags and suitably
783 -- dereferenced for the CCall. For each arg, return the
784 -- depth to the first word of the bits for that arg, and the
785 -- CgRep of what was actually pushed.
787 pargs d [] = returnBc []
789 = let arg_ty = repType (exprType (deAnnotate' a))
791 in case splitTyConApp_maybe arg_ty of
792 -- Don't push the FO; instead push the Addr# it
795 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
796 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
797 parg_ArrayishRep arrPtrsHdrSize d p a
799 returnBc ((code,NonPtrArg):rest)
801 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
802 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
803 parg_ArrayishRep arrWordsHdrSize d p a
805 returnBc ((code,NonPtrArg):rest)
807 -- Default case: push taggedly, but otherwise intact.
809 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
810 pargs (d+sz_a) az `thenBc` \ rest ->
811 returnBc ((code_a, atomRep a) : rest)
813 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
814 -- the stack but then advance it over the headers, so as to
815 -- point to the payload.
816 parg_ArrayishRep hdrSize d p a
817 = pushAtom d p a `thenBc` \ (push_fo, _) ->
818 -- The ptr points at the header. Advance it over the
819 -- header and then pretend this is an Addr#.
820 returnBc (push_fo `snocOL` SWIZZLE 0 hdrSize)
823 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
825 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
827 push_args = concatOL pushs_arg
828 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
830 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
831 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
833 = reverse (tail a_reps_pushed_r_to_l)
835 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
836 -- push_args is the code to do that.
837 -- d_after_args is the stack depth once the args are on.
839 -- Get the result rep.
840 (returns_void, r_rep)
841 = case maybe_getCCallReturnRep (idType fn) of
842 Nothing -> (True, VoidArg)
843 Just rr -> (False, rr)
845 Because the Haskell stack grows down, the a_reps refer to
846 lowest to highest addresses in that order. The args for the call
847 are on the stack. Now push an unboxed Addr# indicating
848 the C function to call. Then push a dummy placeholder for the
849 result. Finally, emit a CCALL insn with an offset pointing to the
850 Addr# just pushed, and a literal field holding the mallocville
851 address of the piece of marshalling code we generate.
852 So, just prior to the CCALL insn, the stack looks like this
853 (growing down, as usual):
858 Addr# address_of_C_fn
859 <placeholder-for-result#> (must be an unboxed type)
861 The interpreter then calls the marshall code mentioned
862 in the CCALL insn, passing it (& <placeholder-for-result#>),
863 that is, the addr of the topmost word in the stack.
864 When this returns, the placeholder will have been
865 filled in. The placeholder is slid down to the sequel
866 depth, and we RETURN.
868 This arrangement makes it simple to do f-i-dynamic since the Addr#
869 value is the first arg anyway.
871 The marshalling code is generated specifically for this
872 call site, and so knows exactly the (Haskell) stack
873 offsets of the args, fn address and placeholder. It
874 copies the args to the C stack, calls the stacked addr,
875 and parks the result back in the placeholder. The interpreter
876 calls it as a normal C call, assuming it has a signature
877 void marshall_code ( StgWord* ptr_to_top_of_stack )
879 -- resolve static address
883 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
885 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
888 get_target_info `thenBc` \ (is_static, static_target_addr) ->
891 -- Get the arg reps, zapping the leading Addr# in the dynamic case
892 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
893 | is_static = a_reps_pushed_RAW
894 | otherwise = if null a_reps_pushed_RAW
895 then panic "ByteCodeGen.generateCCall: dyn with no args"
896 else tail a_reps_pushed_RAW
899 (push_Addr, d_after_Addr)
901 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
902 d_after_args + addr_sizeW)
903 | otherwise -- is already on the stack
904 = (nilOL, d_after_args)
906 -- Push the return placeholder. For a call returning nothing,
907 -- this is a VoidArg (tag).
908 r_sizeW = cgRepSizeW r_rep
909 d_after_r = d_after_Addr + r_sizeW
910 r_lit = mkDummyLiteral r_rep
911 push_r = (if returns_void
913 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
915 -- generate the marshalling code we're going to call
918 arg1_offW = r_sizeW + addr_sizeW
919 args_offW = map (arg1_offW +)
920 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
922 ioToBc (mkMarshalCode cconv
923 (r_offW, r_rep) addr_offW
924 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
925 recordMallocBc addr_of_marshaller `thenBc_`
927 -- Offset of the next stack frame down the stack. The CCALL
928 -- instruction needs to describe the chunk of stack containing
929 -- the ccall args to the GC, so it needs to know how large it
930 -- is. See comment in Interpreter.c with the CCALL instruction.
931 stk_offset = d_after_r - s
934 do_call = unitOL (CCALL stk_offset (castPtr addr_of_marshaller))
936 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
937 `snocOL` RETURN_UBX r_rep
939 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
942 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
946 -- Make a dummy literal, to be used as a placeholder for FFI return
947 -- values on the stack.
948 mkDummyLiteral :: CgRep -> Literal
951 NonPtrArg -> MachWord 0
952 DoubleArg -> MachDouble 0
953 FloatArg -> MachFloat 0
954 _ -> moan64 "mkDummyLiteral" (ppr pr)
958 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
959 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
962 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
964 -- Alternatively, for call-targets returning nothing, convert
966 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
967 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
971 maybe_getCCallReturnRep :: Type -> Maybe CgRep
972 maybe_getCCallReturnRep fn_ty
973 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
975 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
977 = case splitTyConApp_maybe (repType r_ty) of
978 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
980 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
981 || r_reps == [VoidArg] )
982 && isUnboxedTupleTyCon r_tycon
983 && case maybe_r_rep_to_go of
985 Just r_rep -> r_rep /= PtrArg
986 -- if it was, it would be impossible
987 -- to create a valid return value
988 -- placeholder on the stack
989 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
992 --trace (showSDoc (ppr (a_reps, r_reps))) $
993 if ok then maybe_r_rep_to_go else blargh
995 -- Compile code which expects an unboxed Int on the top of stack,
996 -- (call it i), and pushes the i'th closure in the supplied list
998 implement_tagToId :: [Name] -> BcM BCInstrList
999 implement_tagToId names
1000 = ASSERT( notNull names )
1001 getLabelsBc (length names) `thenBc` \ labels ->
1002 getLabelBc `thenBc` \ label_fail ->
1003 getLabelBc `thenBc` \ label_exit ->
1004 zip4 labels (tail labels ++ [label_fail])
1005 [0 ..] names `bind` \ infos ->
1006 map (mkStep label_exit) infos `bind` \ steps ->
1007 returnBc (concatOL steps
1009 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1011 mkStep l_exit (my_label, next_label, n, name_for_n)
1012 = toOL [LABEL my_label,
1013 TESTEQ_I n next_label,
1018 -- -----------------------------------------------------------------------------
1021 -- Push an atom onto the stack, returning suitable code & number of
1022 -- stack words used.
1024 -- The env p must map each variable to the highest- numbered stack
1025 -- slot for it. For example, if the stack has depth 4 and we
1026 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1027 -- the tag in stack[5], the stack will have depth 6, and p must map v
1028 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1029 -- depth 6 stack has valid words 0 .. 5.
1031 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1033 pushAtom d p (AnnApp f (_, AnnType _))
1034 = pushAtom d p (snd f)
1036 pushAtom d p (AnnNote note e)
1037 = pushAtom d p (snd e)
1039 pushAtom d p (AnnLam x e)
1041 = pushAtom d p (snd e)
1043 pushAtom d p (AnnVar v)
1045 | idCgRep v == VoidArg
1046 = returnBc (nilOL, 0)
1049 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1051 | Just primop <- isPrimOpId_maybe v
1052 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1054 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1055 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1056 -- d - d_v the number of words between the TOS
1057 -- and the 1st slot of the object
1059 -- d - d_v - 1 the offset from the TOS of the 1st slot
1061 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1064 -- Having found the last slot, we proceed to copy the right number of
1065 -- slots on to the top of the stack.
1067 | otherwise -- v must be a global variable
1069 returnBc (unitOL (PUSH_G (getName v)), sz)
1075 pushAtom d p (AnnLit lit)
1077 MachLabel fs _ -> code NonPtrArg
1078 MachWord w -> code NonPtrArg
1079 MachInt i -> code PtrArg
1080 MachFloat r -> code FloatArg
1081 MachDouble r -> code DoubleArg
1082 MachChar c -> code NonPtrArg
1083 MachStr s -> pushStr s
1086 = let size_host_words = cgRepSizeW rep
1087 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1091 = let getMallocvilleAddr
1093 FastString _ n _ fp _ ->
1094 -- we could grab the Ptr from the ForeignPtr,
1095 -- but then we have no way to control its lifetime.
1096 -- In reality it'll probably stay alive long enoungh
1097 -- by virtue of the global FastString table, but
1098 -- to be on the safe side we copy the string into
1099 -- a malloc'd area of memory.
1100 ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1101 recordMallocBc ptr `thenBc_`
1103 withForeignPtr fp $ \p -> do
1104 memcpy ptr p (fromIntegral n)
1105 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
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 -> Ptr b -> 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
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])