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, 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 tycs
73 = do showPass dflags "ByteCodeGen"
75 let flatBinds = [ (bndr, freeVars rhs)
76 | (bndr, rhs) <- flattenBinds binds]
78 (BcM_State final_ctr mallocd, proto_bcos)
79 <- runBc (mapM schemeTopBind flatBinds)
81 when (notNull mallocd)
82 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
84 dumpIfSet_dyn dflags Opt_D_dump_BCOs
85 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
87 assembleBCOs proto_bcos tycs
89 -- -----------------------------------------------------------------------------
90 -- Generating byte code for an expression
92 -- Returns: (the root BCO for this expression,
93 -- a list of auxilary BCOs resulting from compiling closures)
94 coreExprToBCOs :: DynFlags
97 coreExprToBCOs dflags expr
98 = do showPass dflags "ByteCodeGen"
100 -- create a totally bogus name for the top-level BCO; this
101 -- should be harmless, since it's never used for anything
102 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
103 invented_id = mkLocalId invented_name (panic "invented_id's type")
105 (BcM_State final_ctr mallocd, proto_bco)
106 <- runBc (schemeTopBind (invented_id, freeVars expr))
108 when (notNull mallocd)
109 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
111 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
113 assembleBCO proto_bco
116 -- -----------------------------------------------------------------------------
117 -- Compilation schema for the bytecode generator
119 type BCInstrList = OrdList BCInstr
121 type Sequel = Int -- back off to this depth before ENTER
123 -- Maps Ids to the offset from the stack _base_ so we don't have
124 -- to mess with it after each push/pop.
125 type BCEnv = FiniteMap Id Int -- To find vars on the stack
127 ppBCEnv :: BCEnv -> SDoc
130 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
133 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
134 cmp_snd x y = compare (snd x) (snd y)
136 -- Create a BCO and do a spot of peephole optimisation on the insns
141 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
145 -> Bool -- True <=> is a return point, rather than a function
148 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap
149 is_ret mallocd_blocks
152 protoBCOInstrs = maybe_with_stack_check,
153 protoBCOBitmap = bitmap,
154 protoBCOBitmapSize = bitmap_size,
155 protoBCOArity = arity,
156 protoBCOExpr = origin,
157 protoBCOPtrs = mallocd_blocks
160 -- Overestimate the stack usage (in words) of this BCO,
161 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
162 -- stack check. (The interpreter always does a stack check
163 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
164 -- BCO anyway, so we only need to add an explicit on in the
165 -- (hopefully rare) cases when the (overestimated) stack use
166 -- exceeds iNTERP_STACK_CHECK_THRESH.
167 maybe_with_stack_check
169 -- don't do stack checks at return points;
170 -- everything is aggregated up to the top BCO
171 -- (which must be a function)
172 | stack_overest >= 65535
173 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
175 | stack_overest >= iNTERP_STACK_CHECK_THRESH
176 = STKCHECK stack_overest : peep_d
178 = peep_d -- the supposedly common case
180 stack_overest = sum (map bciStackUse peep_d)
182 -- Merge local pushes
183 peep_d = peep (fromOL instrs_ordlist)
185 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
186 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
187 peep (PUSH_L off1 : PUSH_L off2 : rest)
188 = PUSH_LL off1 (off2-1) : peep rest
194 argBits :: [CgRep] -> [Bool]
197 | isFollowableArg rep = False : argBits args
198 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
200 -- -----------------------------------------------------------------------------
203 -- Compile code for the right-hand side of a top-level binding
205 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
208 schemeTopBind (id, rhs)
209 | Just data_con <- isDataConWorkId_maybe id,
210 isNullaryRepDataCon data_con
211 = -- Special case for the worker of a nullary data con.
212 -- It'll look like this: Nil = /\a -> Nil a
213 -- If we feed it into schemeR, we'll get
215 -- because mkConAppCode treats nullary constructor applications
216 -- by just re-using the single top-level definition. So
217 -- for the worker itself, we must allocate it directly.
218 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
219 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
222 = schemeR [{- No free variables -}] (id, rhs)
224 -- -----------------------------------------------------------------------------
227 -- Compile code for a right-hand side, to give a BCO that,
228 -- when executed with the free variables and arguments on top of the stack,
229 -- will return with a pointer to the result on top of the stack, after
230 -- removing the free variables and arguments.
232 -- Park the resulting BCO in the monad. Also requires the
233 -- variable to which this value was bound, so as to give the
234 -- resulting BCO a name.
236 schemeR :: [Id] -- Free vars of the RHS, ordered as they
237 -- will appear in the thunk. Empty for
238 -- top-level things, which have no free vars.
239 -> (Id, AnnExpr Id VarSet)
240 -> BcM (ProtoBCO Name)
241 schemeR fvs (nm, rhs)
245 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
246 $$ pprCoreExpr (deAnnotate rhs)
252 = schemeR_wrk fvs nm rhs (collect [] rhs)
254 collect xs (_, AnnNote note e) = collect xs e
255 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
256 collect xs (_, not_lambda) = (reverse xs, not_lambda)
258 schemeR_wrk fvs nm original_body (args, body)
260 all_args = reverse args ++ fvs
261 arity = length all_args
262 -- all_args are the args in reverse order. We're compiling a function
263 -- \fv1..fvn x1..xn -> e
264 -- i.e. the fvs come first
266 szsw_args = map idSizeW all_args
267 szw_args = sum szsw_args
268 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
270 -- make the arg bitmap
271 bits = argBits (reverse (map idCgRep all_args))
272 bitmap_size = length bits
273 bitmap = mkBitmap bits
275 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
276 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
277 arity bitmap_size bitmap False{-not alts-})
280 fvsToEnv :: BCEnv -> VarSet -> [Id]
281 -- Takes the free variables of a right-hand side, and
282 -- delivers an ordered list of the local variables that will
283 -- be captured in the thunk for the RHS
284 -- The BCEnv argument tells which variables are in the local
285 -- environment: these are the ones that should be captured
287 -- The code that constructs the thunk, and the code that executes
288 -- it, have to agree about this layout
289 fvsToEnv p fvs = [v | v <- varSetElems fvs,
290 isId v, -- Could be a type variable
293 -- -----------------------------------------------------------------------------
296 -- Compile code to apply the given expression to the remaining args
297 -- on the stack, returning a HNF.
298 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
300 -- Delegate tail-calls to schemeT.
301 schemeE d s p e@(AnnApp f a)
304 schemeE d s p e@(AnnVar v)
305 | not (isUnLiftedType v_type)
306 = -- Lifted-type thing; push it in the normal way
310 = -- Returning an unlifted value.
311 -- Heave it on the stack, SLIDE, and RETURN.
312 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
313 returnBc (push -- value onto stack
314 `appOL` mkSLIDE szw (d-s) -- clear to sequel
315 `snocOL` RETURN_UBX v_rep) -- go
318 v_rep = typeCgRep v_type
320 schemeE d s p (AnnLit literal)
321 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
322 let l_rep = typeCgRep (literalType literal)
323 in returnBc (push -- value onto stack
324 `appOL` mkSLIDE szw (d-s) -- clear to sequel
325 `snocOL` RETURN_UBX l_rep) -- go
328 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
329 | (AnnVar v, args_r_to_l) <- splitApp rhs,
330 Just data_con <- isDataConWorkId_maybe v,
331 dataConRepArity data_con == length args_r_to_l
332 = -- Special case for a non-recursive let whose RHS is a
333 -- saturatred constructor application.
334 -- Just allocate the constructor and carry on
335 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
336 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
337 returnBc (alloc_code `appOL` body_code)
339 -- General case for let. Generates correct, if inefficient, code in
341 schemeE d s p (AnnLet binds (_,body))
342 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
343 AnnRec xs_n_rhss -> unzip xs_n_rhss
346 fvss = map (fvsToEnv p' . fst) rhss
348 -- Sizes of free vars
349 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
351 -- the arity of each rhs
352 arities = map (length . fst . collect []) rhss
354 -- This p', d' defn is safe because all the items being pushed
355 -- are ptrs, so all have size 1. d' and p' reflect the stack
356 -- after the closures have been allocated in the heap (but not
357 -- filled in), and pointers to them parked on the stack.
358 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
360 zipE = zipEqual "schemeE"
362 -- ToDo: don't build thunks for things with no free variables
363 build_thunk dd [] size bco off
364 = returnBc (PUSH_BCO bco
365 `consOL` unitOL (MKAP (off+size) size))
366 build_thunk dd (fv:fvs) size bco off = do
367 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
368 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off
369 returnBc (push_code `appOL` more_push_code)
371 alloc_code = toOL (zipWith mkAlloc sizes arities)
372 where mkAlloc sz 0 = ALLOC_AP sz
373 mkAlloc sz arity = ALLOC_PAP arity sz
375 compile_bind d' fvs x rhs size off = do
376 bco <- schemeR fvs (x,rhs)
377 build_thunk d' fvs size bco off
380 [ compile_bind d' fvs x rhs size n
381 | (fvs, x, rhs, size, n) <-
382 zip5 fvss xs rhss sizes [n_binds, n_binds-1 .. 1]
385 body_code <- schemeE d' s p' body
386 thunk_codes <- sequence compile_binds
387 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
391 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
392 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
394 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
396 -- case .... of a { DEFAULT -> ... }
397 -- becuse the return convention for both are identical.
399 -- Note that it does not matter losing the void-rep thing from the
400 -- envt (it won't be bound now) because we never look such things up.
402 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
403 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
405 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
406 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
407 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
409 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
410 | isUnboxedTupleCon dc
411 -- Similarly, convert
412 -- case .... of x { (# a #) -> ... }
414 -- case .... of a { DEFAULT -> ... }
415 = --trace "automagic mashing of case alts (# a #)" $
416 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
418 schemeE d s p (AnnCase scrut bndr _ alts)
419 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
421 schemeE d s p (AnnNote note (_, body))
425 = pprPanic "ByteCodeGen.schemeE: unhandled case"
426 (pprCoreExpr (deAnnotate' other))
429 -- Compile code to do a tail call. Specifically, push the fn,
430 -- slide the on-stack app back down to the sequel depth,
431 -- and enter. Four cases:
434 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
435 -- The int will be on the stack. Generate a code sequence
436 -- to convert it to the relevant constructor, SLIDE and ENTER.
438 -- 1. The fn denotes a ccall. Defer to generateCCall.
440 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
441 -- it simply as b -- since the representations are identical
442 -- (the VoidArg takes up zero stack space). Also, spot
443 -- (# b #) and treat it as b.
445 -- 3. Application of a constructor, by defn saturated.
446 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
447 -- then the ptrs, and then do PACK and RETURN.
449 -- 4. Otherwise, it must be a function call. Push the args
450 -- right to left, SLIDE and ENTER.
452 schemeT :: Int -- Stack depth
453 -> Sequel -- Sequel depth
454 -> BCEnv -- stack env
455 -> AnnExpr' Id VarSet
460 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
461 -- = panic "schemeT ?!?!"
463 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
467 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
468 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
469 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
470 returnBc (push `appOL` tagToId_sequence
471 `appOL` mkSLIDE 1 (d+arg_words-s)
475 | Just (CCall ccall_spec) <- isFCallId_maybe fn
476 = generateCCall d s p ccall_spec fn args_r_to_l
478 -- Case 2: Constructor application
479 | Just con <- maybe_saturated_dcon,
480 isUnboxedTupleCon con
481 = case args_r_to_l of
482 [arg1,arg2] | isVoidArgAtom arg1 ->
483 unboxedTupleReturn d s p arg2
484 [arg1,arg2] | isVoidArgAtom arg2 ->
485 unboxedTupleReturn d s p arg1
486 _other -> unboxedTupleException
488 -- Case 3: Ordinary data constructor
489 | Just con <- maybe_saturated_dcon
490 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
491 returnBc (alloc_con `appOL`
492 mkSLIDE 1 (d - s) `snocOL`
495 -- Case 4: Tail call of function
497 = doTailCall d s p fn args_r_to_l
500 -- Detect and extract relevant info for the tagToEnum kludge.
501 maybe_is_tagToEnum_call
502 = let extract_constr_Names ty
503 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
505 = map (getName . dataConWorkId) (tyConDataCons tyc)
506 -- NOTE: use the worker name, not the source name of
507 -- the DataCon. See DataCon.lhs for details.
509 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
512 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
513 -> case isPrimOpId_maybe v of
514 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
518 -- Extract the args (R->L) and fn
519 -- The function will necessarily be a variable,
520 -- because we are compiling a tail call
521 (AnnVar fn, args_r_to_l) = splitApp app
523 -- Only consider this to be a constructor application iff it is
524 -- saturated. Otherwise, we'll call the constructor wrapper.
525 n_args = length args_r_to_l
527 = case isDataConWorkId_maybe fn of
528 Just con | dataConRepArity con == n_args -> Just con
531 -- -----------------------------------------------------------------------------
532 -- Generate code to build a constructor application,
533 -- leaving it on top of the stack
535 mkConAppCode :: Int -> Sequel -> BCEnv
536 -> DataCon -- The data constructor
537 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
540 mkConAppCode orig_d s p con [] -- Nullary constructor
541 = ASSERT( isNullaryRepDataCon con )
542 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
543 -- Instead of doing a PACK, which would allocate a fresh
544 -- copy of this constructor, use the single shared version.
546 mkConAppCode orig_d s p con args_r_to_l
547 = ASSERT( dataConRepArity con == length args_r_to_l )
548 do_pushery orig_d (non_ptr_args ++ ptr_args)
550 -- The args are already in reverse order, which is the way PACK
551 -- expects them to be. We must push the non-ptrs after the ptrs.
552 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
554 do_pushery d (arg:args)
555 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
556 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
557 returnBc (push `appOL` more_push_code)
559 = returnBc (unitOL (PACK con n_arg_words))
561 n_arg_words = d - orig_d
564 -- -----------------------------------------------------------------------------
565 -- Returning an unboxed tuple with one non-void component (the only
566 -- case we can handle).
568 -- Remember, we don't want to *evaluate* the component that is being
569 -- returned, even if it is a pointed type. We always just return.
572 :: Int -> Sequel -> BCEnv
573 -> AnnExpr' Id VarSet -> BcM BCInstrList
574 unboxedTupleReturn d s p arg = do
575 (push, sz) <- pushAtom d p arg
576 returnBc (push `appOL`
577 mkSLIDE sz (d-s) `snocOL`
578 RETURN_UBX (atomRep arg))
580 -- -----------------------------------------------------------------------------
581 -- Generate code for a tail-call
584 :: Int -> Sequel -> BCEnv
585 -> Id -> [AnnExpr' Id VarSet]
587 doTailCall init_d s p fn args
588 = do_pushes init_d args (map atomRep args)
590 do_pushes d [] reps = do
591 ASSERT( null reps ) return ()
592 (push_fn, sz) <- pushAtom d p (AnnVar fn)
593 ASSERT( sz == 1 ) return ()
594 returnBc (push_fn `appOL` (
595 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
597 do_pushes d args reps = do
598 let (push_apply, n, rest_of_reps) = findPushSeq reps
599 (these_args, rest_of_args) = splitAt n args
600 (next_d, push_code) <- push_seq d these_args
601 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
602 -- ^^^ for the PUSH_APPLY_ instruction
603 returnBc (push_code `appOL` (push_apply `consOL` instrs))
605 push_seq d [] = return (d, nilOL)
606 push_seq d (arg:args) = do
607 (push_code, sz) <- pushAtom d p arg
608 (final_d, more_push_code) <- push_seq (d+sz) args
609 return (final_d, push_code `appOL` more_push_code)
611 -- v. similar to CgStackery.findMatch, ToDo: merge
612 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
613 = (PUSH_APPLY_PPPPPP, 6, rest)
614 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
615 = (PUSH_APPLY_PPPPP, 5, rest)
616 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
617 = (PUSH_APPLY_PPPP, 4, rest)
618 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
619 = (PUSH_APPLY_PPP, 3, rest)
620 findPushSeq (PtrArg: PtrArg: rest)
621 = (PUSH_APPLY_PP, 2, rest)
622 findPushSeq (PtrArg: rest)
623 = (PUSH_APPLY_P, 1, rest)
624 findPushSeq (VoidArg: rest)
625 = (PUSH_APPLY_V, 1, rest)
626 findPushSeq (NonPtrArg: rest)
627 = (PUSH_APPLY_N, 1, rest)
628 findPushSeq (FloatArg: rest)
629 = (PUSH_APPLY_F, 1, rest)
630 findPushSeq (DoubleArg: rest)
631 = (PUSH_APPLY_D, 1, rest)
632 findPushSeq (LongArg: rest)
633 = (PUSH_APPLY_L, 1, rest)
635 = panic "ByteCodeGen.findPushSeq"
637 -- -----------------------------------------------------------------------------
640 doCase :: Int -> Sequel -> BCEnv
641 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
642 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
644 doCase d s p (_,scrut)
645 bndr alts is_unboxed_tuple
647 -- Top of stack is the return itbl, as usual.
648 -- underneath it is the pointer to the alt_code BCO.
649 -- When an alt is entered, it assumes the returned value is
650 -- on top of the itbl.
653 -- An unlifted value gets an extra info table pushed on top
654 -- when it is returned.
655 unlifted_itbl_sizeW | isAlgCase = 0
658 -- depth of stack after the return value has been pushed
659 d_bndr = d + ret_frame_sizeW + idSizeW bndr
661 -- depth of stack after the extra info table for an unboxed return
662 -- has been pushed, if any. This is the stack depth at the
664 d_alts = d_bndr + unlifted_itbl_sizeW
666 -- Env in which to compile the alts, not including
667 -- any vars bound by the alts themselves
668 p_alts = addToFM p bndr (d_bndr - 1)
670 bndr_ty = idType bndr
671 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
673 -- given an alt, return a discr and code for it.
674 codeALt alt@(DEFAULT, _, (_,rhs))
675 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
676 returnBc (NoDiscr, rhs_code)
677 codeAlt alt@(discr, bndrs, (_,rhs))
678 -- primitive or nullary constructor alt: no need to UNPACK
679 | null real_bndrs = do
680 rhs_code <- schemeE d_alts s p_alts rhs
681 returnBc (my_discr alt, rhs_code)
682 -- algebraic alt with some binders
683 | ASSERT(isAlgCase) otherwise =
685 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
686 ptr_sizes = map idSizeW ptrs
687 nptrs_sizes = map idSizeW nptrs
688 bind_sizes = ptr_sizes ++ nptrs_sizes
689 size = sum ptr_sizes + sum nptrs_sizes
690 -- the UNPACK instruction unpacks in reverse order...
691 p' = addListToFM p_alts
692 (zip (reverse (ptrs ++ nptrs))
693 (mkStackOffsets d_alts (reverse bind_sizes)))
695 rhs_code <- schemeE (d_alts+size) s p' rhs
696 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
698 real_bndrs = filter (not.isTyVar) bndrs
701 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
702 my_discr (DataAlt dc, binds, rhs)
703 | isUnboxedTupleCon dc
704 = unboxedTupleException
706 = DiscrP (dataConTag dc - fIRST_TAG)
707 my_discr (LitAlt l, binds, rhs)
708 = case l of MachInt i -> DiscrI (fromInteger i)
709 MachFloat r -> DiscrF (fromRational r)
710 MachDouble r -> DiscrD (fromRational r)
711 MachChar i -> DiscrI (ord i)
712 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
715 | not isAlgCase = Nothing
717 = case [dc | (DataAlt dc, _, _) <- alts] of
719 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
721 -- the bitmap is relative to stack depth d, i.e. before the
722 -- BCO, info table and return value are pushed on.
723 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
724 -- except that here we build the bitmap from the known bindings of
725 -- things that are pointers, whereas in CgBindery the code builds the
726 -- bitmap from the free slots and unboxed bindings.
728 bitmap = intsToReverseBitmap d{-size-} (sortLe (<=) rel_slots)
731 rel_slots = concat (map spread binds)
733 | isFollowableArg (idCgRep id) = [ rel_offset ]
735 where rel_offset = d - offset - 1
738 alt_stuff <- mapM codeAlt alts
739 alt_final <- mkMultiBranch maybe_ncons alt_stuff
741 alt_bco_name = getName bndr
742 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
743 0{-no arity-} d{-bitmap size-} bitmap True{-is alts-}
745 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
746 -- "\n bitmap = " ++ show bitmap) $ do
747 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
748 alt_bco' <- emitBc alt_bco
750 | isAlgCase = PUSH_ALTS alt_bco'
751 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
752 returnBc (push_alts `consOL` scrut_code)
755 -- -----------------------------------------------------------------------------
756 -- Deal with a CCall.
758 -- Taggedly push the args onto the stack R->L,
759 -- deferencing ForeignObj#s and adjusting addrs to point to
760 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
761 -- (machine) code for the ccall, and create bytecodes to call that and
762 -- then return in the right way.
764 generateCCall :: Int -> Sequel -- stack and sequel depths
766 -> CCallSpec -- where to call
767 -> Id -- of target, for type info
768 -> [AnnExpr' Id VarSet] -- args (atoms)
771 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
774 addr_sizeW = cgRepSizeW NonPtrArg
776 -- Get the args on the stack, with tags and suitably
777 -- dereferenced for the CCall. For each arg, return the
778 -- depth to the first word of the bits for that arg, and the
779 -- CgRep of what was actually pushed.
781 pargs d [] = returnBc []
783 = let arg_ty = repType (exprType (deAnnotate' a))
785 in case splitTyConApp_maybe arg_ty of
786 -- Don't push the FO; instead push the Addr# it
789 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
790 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
791 parg_ArrayishRep arrPtrsHdrSize d p a
793 returnBc ((code,NonPtrArg):rest)
795 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
796 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
797 parg_ArrayishRep arrWordsHdrSize d p a
799 returnBc ((code,NonPtrArg):rest)
801 -- Default case: push taggedly, but otherwise intact.
803 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
804 pargs (d+sz_a) az `thenBc` \ rest ->
805 returnBc ((code_a, atomRep a) : rest)
807 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
808 -- the stack but then advance it over the headers, so as to
809 -- point to the payload.
810 parg_ArrayishRep hdrSize d p a
811 = pushAtom d p a `thenBc` \ (push_fo, _) ->
812 -- The ptr points at the header. Advance it over the
813 -- header and then pretend this is an Addr#.
814 returnBc (push_fo `snocOL` SWIZZLE 0 hdrSize)
817 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
819 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
821 push_args = concatOL pushs_arg
822 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
824 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
825 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
827 = reverse (tail a_reps_pushed_r_to_l)
829 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
830 -- push_args is the code to do that.
831 -- d_after_args is the stack depth once the args are on.
833 -- Get the result rep.
834 (returns_void, r_rep)
835 = case maybe_getCCallReturnRep (idType fn) of
836 Nothing -> (True, VoidArg)
837 Just rr -> (False, rr)
839 Because the Haskell stack grows down, the a_reps refer to
840 lowest to highest addresses in that order. The args for the call
841 are on the stack. Now push an unboxed Addr# indicating
842 the C function to call. Then push a dummy placeholder for the
843 result. Finally, emit a CCALL insn with an offset pointing to the
844 Addr# just pushed, and a literal field holding the mallocville
845 address of the piece of marshalling code we generate.
846 So, just prior to the CCALL insn, the stack looks like this
847 (growing down, as usual):
852 Addr# address_of_C_fn
853 <placeholder-for-result#> (must be an unboxed type)
855 The interpreter then calls the marshall code mentioned
856 in the CCALL insn, passing it (& <placeholder-for-result#>),
857 that is, the addr of the topmost word in the stack.
858 When this returns, the placeholder will have been
859 filled in. The placeholder is slid down to the sequel
860 depth, and we RETURN.
862 This arrangement makes it simple to do f-i-dynamic since the Addr#
863 value is the first arg anyway.
865 The marshalling code is generated specifically for this
866 call site, and so knows exactly the (Haskell) stack
867 offsets of the args, fn address and placeholder. It
868 copies the args to the C stack, calls the stacked addr,
869 and parks the result back in the placeholder. The interpreter
870 calls it as a normal C call, assuming it has a signature
871 void marshall_code ( StgWord* ptr_to_top_of_stack )
873 -- resolve static address
877 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
879 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
882 get_target_info `thenBc` \ (is_static, static_target_addr) ->
885 -- Get the arg reps, zapping the leading Addr# in the dynamic case
886 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
887 | is_static = a_reps_pushed_RAW
888 | otherwise = if null a_reps_pushed_RAW
889 then panic "ByteCodeGen.generateCCall: dyn with no args"
890 else tail a_reps_pushed_RAW
893 (push_Addr, d_after_Addr)
895 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
896 d_after_args + addr_sizeW)
897 | otherwise -- is already on the stack
898 = (nilOL, d_after_args)
900 -- Push the return placeholder. For a call returning nothing,
901 -- this is a VoidArg (tag).
902 r_sizeW = cgRepSizeW r_rep
903 d_after_r = d_after_Addr + r_sizeW
904 r_lit = mkDummyLiteral r_rep
905 push_r = (if returns_void
907 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
909 -- generate the marshalling code we're going to call
912 arg1_offW = r_sizeW + addr_sizeW
913 args_offW = map (arg1_offW +)
914 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
916 ioToBc (mkMarshalCode cconv
917 (r_offW, r_rep) addr_offW
918 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
919 recordMallocBc addr_of_marshaller `thenBc_`
921 -- Offset of the next stack frame down the stack. The CCALL
922 -- instruction needs to describe the chunk of stack containing
923 -- the ccall args to the GC, so it needs to know how large it
924 -- is. See comment in Interpreter.c with the CCALL instruction.
925 stk_offset = d_after_r - s
928 do_call = unitOL (CCALL stk_offset (castPtr addr_of_marshaller))
930 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
931 `snocOL` RETURN_UBX r_rep
933 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
936 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
940 -- Make a dummy literal, to be used as a placeholder for FFI return
941 -- values on the stack.
942 mkDummyLiteral :: CgRep -> Literal
945 NonPtrArg -> MachWord 0
946 DoubleArg -> MachDouble 0
947 FloatArg -> MachFloat 0
948 _ -> moan64 "mkDummyLiteral" (ppr pr)
952 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
953 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
956 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
958 -- Alternatively, for call-targets returning nothing, convert
960 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
961 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
965 maybe_getCCallReturnRep :: Type -> Maybe CgRep
966 maybe_getCCallReturnRep fn_ty
967 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
969 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
971 = case splitTyConApp_maybe (repType r_ty) of
972 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
974 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
975 || r_reps == [VoidArg] )
976 && isUnboxedTupleTyCon r_tycon
977 && case maybe_r_rep_to_go of
979 Just r_rep -> r_rep /= PtrArg
980 -- if it was, it would be impossible
981 -- to create a valid return value
982 -- placeholder on the stack
983 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
986 --trace (showSDoc (ppr (a_reps, r_reps))) $
987 if ok then maybe_r_rep_to_go else blargh
989 -- Compile code which expects an unboxed Int on the top of stack,
990 -- (call it i), and pushes the i'th closure in the supplied list
992 implement_tagToId :: [Name] -> BcM BCInstrList
993 implement_tagToId names
994 = ASSERT( notNull names )
995 getLabelsBc (length names) `thenBc` \ labels ->
996 getLabelBc `thenBc` \ label_fail ->
997 getLabelBc `thenBc` \ label_exit ->
998 zip4 labels (tail labels ++ [label_fail])
999 [0 ..] names `bind` \ infos ->
1000 map (mkStep label_exit) infos `bind` \ steps ->
1001 returnBc (concatOL steps
1003 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1005 mkStep l_exit (my_label, next_label, n, name_for_n)
1006 = toOL [LABEL my_label,
1007 TESTEQ_I n next_label,
1012 -- -----------------------------------------------------------------------------
1015 -- Push an atom onto the stack, returning suitable code & number of
1016 -- stack words used.
1018 -- The env p must map each variable to the highest- numbered stack
1019 -- slot for it. For example, if the stack has depth 4 and we
1020 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1021 -- the tag in stack[5], the stack will have depth 6, and p must map v
1022 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1023 -- depth 6 stack has valid words 0 .. 5.
1025 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1027 pushAtom d p (AnnApp f (_, AnnType _))
1028 = pushAtom d p (snd f)
1030 pushAtom d p (AnnNote note e)
1031 = pushAtom d p (snd e)
1033 pushAtom d p (AnnLam x e)
1035 = pushAtom d p (snd e)
1037 pushAtom d p (AnnVar v)
1039 | idCgRep v == VoidArg
1040 = returnBc (nilOL, 0)
1043 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1045 | Just primop <- isPrimOpId_maybe v
1046 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1048 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1049 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1050 -- d - d_v the number of words between the TOS
1051 -- and the 1st slot of the object
1053 -- d - d_v - 1 the offset from the TOS of the 1st slot
1055 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1058 -- Having found the last slot, we proceed to copy the right number of
1059 -- slots on to the top of the stack.
1061 | otherwise -- v must be a global variable
1063 returnBc (unitOL (PUSH_G (getName v)), sz)
1069 pushAtom d p (AnnLit lit)
1071 MachLabel fs _ -> code NonPtrArg
1072 MachWord w -> code NonPtrArg
1073 MachInt i -> code PtrArg
1074 MachFloat r -> code FloatArg
1075 MachDouble r -> code DoubleArg
1076 MachChar c -> code NonPtrArg
1077 MachStr s -> pushStr s
1080 = let size_host_words = cgRepSizeW rep
1081 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1085 = let getMallocvilleAddr
1087 FastString _ l ba ->
1088 -- sigh, a string in the heap is no good to us.
1089 -- We need a static C pointer, since the type of
1090 -- a string literal is Addr#. So, copy the string
1091 -- into C land and remember the pointer so we can
1094 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1095 in ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1096 recordMallocBc ptr `thenBc_`
1098 do memcpy ptr ba (fromIntegral n)
1099 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1102 other -> panic "ByteCodeGen.pushAtom.pushStr"
1104 getMallocvilleAddr `thenBc` \ addr ->
1105 -- Get the addr on the stack, untaggedly
1106 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1109 = pprPanic "ByteCodeGen.pushAtom"
1110 (pprCoreExpr (deAnnotate (undefined, other)))
1112 foreign import ccall unsafe "memcpy"
1113 memcpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1116 -- -----------------------------------------------------------------------------
1117 -- Given a bunch of alts code and their discrs, do the donkey work
1118 -- of making a multiway branch using a switch tree.
1119 -- What a load of hassle!
1121 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1122 -- a hint; generates better code
1123 -- Nothing is always safe
1124 -> [(Discr, BCInstrList)]
1126 mkMultiBranch maybe_ncons raw_ways
1127 = let d_way = filter (isNoDiscr.fst) raw_ways
1129 (\w1 w2 -> leAlt (fst w1) (fst w2))
1130 (filter (not.isNoDiscr.fst) raw_ways)
1132 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1133 mkTree [] range_lo range_hi = returnBc the_default
1135 mkTree [val] range_lo range_hi
1136 | range_lo `eqAlt` range_hi
1137 = returnBc (snd val)
1139 = getLabelBc `thenBc` \ label_neq ->
1140 returnBc (mkTestEQ (fst val) label_neq
1142 `appOL` unitOL (LABEL label_neq)
1143 `appOL` the_default))
1145 mkTree vals range_lo range_hi
1146 = let n = length vals `div` 2
1147 vals_lo = take n vals
1148 vals_hi = drop n vals
1149 v_mid = fst (head vals_hi)
1151 getLabelBc `thenBc` \ label_geq ->
1152 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1153 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1154 returnBc (mkTestLT v_mid label_geq
1156 `appOL` unitOL (LABEL label_geq)
1160 = case d_way of [] -> unitOL CASEFAIL
1163 -- None of these will be needed if there are no non-default alts
1164 (mkTestLT, mkTestEQ, init_lo, init_hi)
1166 = panic "mkMultiBranch: awesome foursome"
1168 = case fst (head notd_ways) of {
1169 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1170 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1173 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1174 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1177 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1178 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1181 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1182 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1184 DiscrP algMaxBound )
1187 (algMinBound, algMaxBound)
1188 = case maybe_ncons of
1189 Just n -> (0, n - 1)
1190 Nothing -> (minBound, maxBound)
1192 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1193 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1194 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1195 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1196 NoDiscr `eqAlt` NoDiscr = True
1199 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1200 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1201 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1202 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1203 NoDiscr `leAlt` NoDiscr = True
1206 isNoDiscr NoDiscr = True
1209 dec (DiscrI i) = DiscrI (i-1)
1210 dec (DiscrP i) = DiscrP (i-1)
1211 dec other = other -- not really right, but if you
1212 -- do cases on floating values, you'll get what you deserve
1214 -- same snotty comment applies to the following
1216 minD, maxD :: Double
1222 mkTree notd_ways init_lo init_hi
1225 -- -----------------------------------------------------------------------------
1226 -- Supporting junk for the compilation schemes
1228 -- Describes case alts
1236 instance Outputable Discr where
1237 ppr (DiscrI i) = int i
1238 ppr (DiscrF f) = text (show f)
1239 ppr (DiscrD d) = text (show d)
1240 ppr (DiscrP i) = int i
1241 ppr NoDiscr = text "DEF"
1244 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1245 lookupBCEnv_maybe = lookupFM
1247 idSizeW :: Id -> Int
1248 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1250 unboxedTupleException :: a
1251 unboxedTupleException
1254 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1255 "\tto foreign import/export decls in source. Workaround:\n" ++
1256 "\tcompile this module to a .o file, then restart session."))
1259 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1262 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1263 -- The arguments are returned in *right-to-left* order
1264 splitApp (AnnApp (_,f) (_,a))
1265 | isTypeAtom a = splitApp f
1266 | otherwise = case splitApp f of
1267 (f', as) -> (f', a:as)
1268 splitApp (AnnNote n (_,e)) = splitApp e
1269 splitApp e = (e, [])
1272 isTypeAtom :: AnnExpr' id ann -> Bool
1273 isTypeAtom (AnnType _) = True
1274 isTypeAtom _ = False
1276 isVoidArgAtom :: AnnExpr' id ann -> Bool
1277 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1278 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1279 isVoidArgAtom _ = False
1281 atomRep :: AnnExpr' Id ann -> CgRep
1282 atomRep (AnnVar v) = typeCgRep (idType v)
1283 atomRep (AnnLit l) = typeCgRep (literalType l)
1284 atomRep (AnnNote n b) = atomRep (snd b)
1285 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1286 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1287 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1289 isPtrAtom :: AnnExpr' Id ann -> Bool
1290 isPtrAtom e = atomRep e == PtrArg
1292 -- Let szsw be the sizes in words of some items pushed onto the stack,
1293 -- which has initial depth d'. Return the values which the stack environment
1294 -- should map these items to.
1295 mkStackOffsets :: Int -> [Int] -> [Int]
1296 mkStackOffsets original_depth szsw
1297 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1299 -- -----------------------------------------------------------------------------
1300 -- The bytecode generator's monad
1304 nextlabel :: Int, -- for generating local labels
1305 malloced :: [Ptr ()] } -- ptrs malloced for current BCO
1306 -- Should be free()d when it is GCd
1308 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1310 ioToBc :: IO a -> BcM a
1311 ioToBc io = BcM $ \st -> do
1315 runBc :: BcM r -> IO (BcM_State, r)
1316 runBc (BcM m) = m (BcM_State 0 [])
1318 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1319 thenBc (BcM expr) cont = BcM $ \st0 -> do
1320 (st1, q) <- expr st0
1325 thenBc_ :: BcM a -> BcM b -> BcM b
1326 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1327 (st1, q) <- expr st0
1328 (st2, r) <- cont st1
1331 returnBc :: a -> BcM a
1332 returnBc result = BcM $ \st -> (return (st, result))
1334 instance Monad BcM where
1339 emitBc :: ([Ptr ()] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1341 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1343 recordMallocBc :: Ptr a -> BcM ()
1345 = BcM $ \st -> return (st{malloced = castPtr a : malloced st}, ())
1347 getLabelBc :: BcM Int
1349 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1351 getLabelsBc :: Int -> BcM [Int]
1353 = BcM $ \st -> let ctr = nextlabel st
1354 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])