2 % (c) The University of Glasgow 2000
4 \section[ByteCodeGen]{Generate bytecode from Core}
7 module ByteCodeGen ( UnlinkedBCO, UnlinkedBCOExpr, ItblEnv, ClosureEnv, HValue,
9 byteCodeGen, coreExprToBCOs
12 #include "HsVersions.h"
15 import Name ( Name, getName )
16 import Id ( Id, idType, isDataConId_maybe, isPrimOpId_maybe, isFCallId,
17 idPrimRep, mkSysLocal, idName, isFCallId_maybe )
18 import ForeignCall ( ForeignCall(..), CCallTarget(..), CCallSpec(..) )
19 import OrdList ( OrdList, consOL, snocOL, appOL, unitOL,
20 nilOL, toOL, concatOL, fromOL )
21 import FiniteMap ( FiniteMap, addListToFM, listToFM,
22 addToFM, lookupFM, fmToList )
24 import PprCore ( pprCoreExpr )
25 import Literal ( Literal(..), literalPrimRep )
26 import PrimRep ( PrimRep(..) )
27 import PrimOp ( PrimOp(..) )
28 import CoreFVs ( freeVars )
29 import Type ( typePrimRep, splitTyConApp_maybe, isTyVarTy )
30 import DataCon ( dataConTag, fIRST_TAG, dataConTyCon,
31 dataConWrapId, isUnboxedTupleCon )
32 import TyCon ( TyCon(..), tyConFamilySize, isDataTyCon, tyConDataCons,
33 isFunTyCon, isUnboxedTupleTyCon )
34 import Class ( Class, classTyCon )
35 import Type ( Type, repType, splitRepFunTys )
36 import Util ( zipEqual, zipWith4Equal, naturalMergeSortLe, nOfThem )
37 import Var ( isTyVar )
38 import VarSet ( VarSet, varSetElems )
39 import PrimRep ( isFollowableRep )
40 import CmdLineOpts ( DynFlags, DynFlag(..) )
41 import ErrUtils ( showPass, dumpIfSet_dyn )
42 import Unique ( mkPseudoUnique3 )
43 import FastString ( FastString(..) )
44 import Panic ( GhcException(..) )
45 import PprType ( pprType )
46 import ByteCodeInstr ( BCInstr(..), ProtoBCO(..), nameOfProtoBCO, bciStackUse )
47 import ByteCodeItbls ( ItblEnv, mkITbls )
48 import ByteCodeLink ( UnlinkedBCO, UnlinkedBCOExpr, assembleBCO,
49 ClosureEnv, HValue, filterNameMap,
50 iNTERP_STACK_CHECK_THRESH )
51 import ByteCodeFFI ( taggedSizeW, untaggedSizeW, mkMarshalCode )
52 import Linker ( lookupSymbol )
54 import List ( intersperse, sortBy, zip4 )
55 import Foreign ( Ptr(..), mallocBytes )
56 import Addr ( Addr(..), writeCharOffAddr )
57 import CTypes ( CInt )
58 import Exception ( throwDyn )
60 import PrelBase ( Int(..) )
61 import PrelGHC ( ByteArray# )
62 import PrelIOBase ( IO(..) )
67 %************************************************************************
69 \subsection{Functions visible from outside this module.}
71 %************************************************************************
75 byteCodeGen :: DynFlags
78 -> IO ([UnlinkedBCO], ItblEnv)
79 byteCodeGen dflags binds local_tycons local_classes
80 = do showPass dflags "ByteCodeGen"
81 let tycs = local_tycons ++ map classTyCon local_classes
82 itblenv <- mkITbls tycs
84 let flatBinds = concatMap getBind binds
85 getBind (NonRec bndr rhs) = [(bndr, freeVars rhs)]
86 getBind (Rec binds) = [(bndr, freeVars rhs) | (bndr,rhs) <- binds]
88 (BcM_State proto_bcos final_ctr mallocd, ())
89 <- runBc (BcM_State [] 0 [])
90 (mapBc (schemeR True) flatBinds `thenBc_` returnBc ())
92 when (not (null mallocd))
93 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
95 dumpIfSet_dyn dflags Opt_D_dump_BCOs
96 "Proto-bcos" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
98 bcos <- mapM assembleBCO proto_bcos
100 return (bcos, itblenv)
103 -- Returns: (the root BCO for this expression,
104 -- a list of auxilary BCOs resulting from compiling closures)
105 coreExprToBCOs :: DynFlags
107 -> IO UnlinkedBCOExpr
108 coreExprToBCOs dflags expr
109 = do showPass dflags "ByteCodeGen"
111 -- create a totally bogus name for the top-level BCO; this
112 -- should be harmless, since it's never used for anything
113 let invented_id = mkSysLocal SLIT("Expr-Top-Level") (mkPseudoUnique3 0)
114 (panic "invented_id's type")
115 let invented_name = idName invented_id
117 (BcM_State all_proto_bcos final_ctr mallocd, ())
118 <- runBc (BcM_State [] 0 [])
119 (schemeR True (invented_id, freeVars expr))
121 when (not (null mallocd))
122 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
124 dumpIfSet_dyn dflags Opt_D_dump_BCOs
125 "Proto-bcos" (vcat (intersperse (char ' ') (map ppr all_proto_bcos)))
128 = case filter ((== invented_name).nameOfProtoBCO) all_proto_bcos of
129 [root_bco] -> root_bco
131 = filter ((/= invented_name).nameOfProtoBCO) all_proto_bcos
133 auxiliary_bcos <- mapM assembleBCO auxiliary_proto_bcos
134 root_bco <- assembleBCO root_proto_bco
136 return (root_bco, auxiliary_bcos)
139 %************************************************************************
141 \subsection{Compilation schema for the bytecode generator.}
143 %************************************************************************
147 type BCInstrList = OrdList BCInstr
149 type Sequel = Int -- back off to this depth before ENTER
151 -- Maps Ids to the offset from the stack _base_ so we don't have
152 -- to mess with it after each push/pop.
153 type BCEnv = FiniteMap Id Int -- To find vars on the stack
155 ppBCEnv :: BCEnv -> SDoc
158 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
161 pp_one (var, offset) = int offset <> colon <+> ppr var
162 cmp_snd x y = compare (snd x) (snd y)
164 -- Create a BCO and do a spot of peephole optimisation on the insns
166 mkProtoBCO nm instrs_ordlist origin mallocd_blocks
167 = ProtoBCO nm maybe_with_stack_check origin mallocd_blocks
169 -- Overestimate the stack usage (in words) of this BCO,
170 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
171 -- stack check. (The interpreter always does a stack check
172 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
173 -- BCO anyway, so we only need to add an explicit on in the
174 -- (hopefully rare) cases when the (overestimated) stack use
175 -- exceeds iNTERP_STACK_CHECK_THRESH.
176 maybe_with_stack_check
177 | stack_overest >= 65535
178 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
180 | stack_overest >= iNTERP_STACK_CHECK_THRESH
181 = (STKCHECK stack_overest) : peep_d
183 = peep_d -- the supposedly common case
185 stack_overest = sum (map bciStackUse peep_d)
186 + 10 {- just to be really really sure -}
189 -- Merge local pushes
190 peep_d = peep (fromOL instrs_ordlist)
192 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
193 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
194 peep (PUSH_L off1 : PUSH_L off2 : rest)
195 = PUSH_LL off1 (off2-1) : peep rest
202 -- Compile code for the right hand side of a let binding.
203 -- Park the resulting BCO in the monad. Also requires the
204 -- variable to which this value was bound, so as to give the
205 -- resulting BCO a name. Bool indicates top-levelness.
207 schemeR :: Bool -> (Id, AnnExpr Id VarSet) -> BcM ()
208 schemeR is_top (nm, rhs)
212 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
213 $$ pprCoreExpr (deAnnotate rhs)
219 = schemeR_wrk is_top rhs nm (collect [] rhs)
222 collect xs (_, AnnNote note e)
224 collect xs (_, AnnLam x e)
225 = collect (if isTyVar x then xs else (x:xs)) e
226 collect xs not_lambda
227 = (reverse xs, not_lambda)
229 schemeR_wrk is_top original_body nm (args, body)
230 | Just dcon <- maybe_toplevel_null_con_rhs
231 = --trace ("nullary constructor! " ++ showSDocDebug (ppr nm)) (
232 emitBc (mkProtoBCO (getName nm) (toOL [PACK dcon 0, ENTER])
233 (Right original_body))
237 = let fvs = filter (not.isTyVar) (varSetElems (fst original_body))
238 all_args = reverse args ++ fvs
239 szsw_args = map taggedIdSizeW all_args
240 szw_args = sum szsw_args
241 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
242 argcheck = unitOL (ARGCHECK szw_args)
244 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
245 emitBc (mkProtoBCO (getName nm) (appOL argcheck body_code)
246 (Right original_body))
249 maybe_toplevel_null_con_rhs
250 | is_top && null args
253 -> case isDataConId_maybe v_wrk of
255 Just dc_wrk | nm == dataConWrapId dc_wrk
263 -- Let szsw be the sizes in words of some items pushed onto the stack,
264 -- which has initial depth d'. Return the values which the stack environment
265 -- should map these items to.
266 mkStackOffsets :: Int -> [Int] -> [Int]
267 mkStackOffsets original_depth szsw
268 = map (subtract 1) (tail (scanl (+) original_depth szsw))
270 -- Compile code to apply the given expression to the remaining args
271 -- on the stack, returning a HNF.
272 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr Id VarSet -> BcM BCInstrList
274 -- Delegate tail-calls to schemeT.
275 schemeE d s p e@(fvs, AnnApp f a)
276 = schemeT d s p (fvs, AnnApp f a)
278 schemeE d s p e@(fvs, AnnVar v)
279 | isFollowableRep v_rep
280 = -- Ptr-ish thing; push it in the normal way
281 schemeT d s p (fvs, AnnVar v)
284 = -- returning an unboxed value. Heave it on the stack, SLIDE, and RETURN.
285 pushAtom True d p (AnnVar v) `thenBc` \ (push, szw) ->
286 returnBc (push -- value onto stack
287 `appOL` mkSLIDE szw (d-s) -- clear to sequel
288 `snocOL` RETURN v_rep) -- go
290 v_rep = typePrimRep (idType v)
292 schemeE d s p (fvs, AnnLit literal)
293 = pushAtom True d p (AnnLit literal) `thenBc` \ (push, szw) ->
294 let l_rep = literalPrimRep literal
295 in returnBc (push -- value onto stack
296 `appOL` mkSLIDE szw (d-s) -- clear to sequel
297 `snocOL` RETURN l_rep) -- go
299 schemeE d s p (fvs, AnnLet binds b)
300 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
301 AnnRec xs_n_rhss -> unzip xs_n_rhss
303 fvss = map (filter (not.isTyVar).varSetElems.fst) rhss
305 -- Sizes of tagged free vars, + 1 for the fn
306 sizes = map (\rhs_fvs -> 1 + sum (map taggedIdSizeW rhs_fvs)) fvss
308 -- This p', d' defn is safe because all the items being pushed
309 -- are ptrs, so all have size 1. d' and p' reflect the stack
310 -- after the closures have been allocated in the heap (but not
311 -- filled in), and pointers to them parked on the stack.
312 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n 1)))
315 infos = zipE4 fvss sizes xs [n, n-1 .. 1]
316 zipE = zipEqual "schemeE"
317 zipE4 = zipWith4Equal "schemeE" (\a b c d -> (a,b,c,d))
319 -- ToDo: don't build thunks for things with no free variables
320 buildThunk dd ([], size, id, off)
321 = returnBc (PUSH_G (Left (getName id))
322 `consOL` unitOL (MKAP (off+size-1) size))
323 buildThunk dd ((fv:fvs), size, id, off)
324 = pushAtom True dd p' (AnnVar fv)
325 `thenBc` \ (push_code, pushed_szw) ->
326 buildThunk (dd+pushed_szw) (fvs, size, id, off)
327 `thenBc` \ more_push_code ->
328 returnBc (push_code `appOL` more_push_code)
330 genThunkCode = mapBc (buildThunk d') infos `thenBc` \ tcodes ->
331 returnBc (concatOL tcodes)
333 allocCode = toOL (map ALLOC sizes)
335 schemeE d' s p' b `thenBc` \ bodyCode ->
336 mapBc (schemeR False) (zip xs rhss) `thenBc_`
337 genThunkCode `thenBc` \ thunkCode ->
338 returnBc (allocCode `appOL` thunkCode `appOL` bodyCode)
344 schemeE d s p (fvs_case, AnnCase (fvs_scrut, scrut) bndr
345 [(DEFAULT, [], (fvs_rhs, rhs))])
347 | let isFunType var_type
348 = case splitTyConApp_maybe var_type of
349 Just (tycon,_) | isFunTyCon tycon -> True
351 ty_bndr = repType (idType bndr)
352 in isFunType ty_bndr || isTyVarTy ty_bndr
355 -- case scrut::suspect of bndr { DEFAULT -> rhs }
357 -- let bndr = scrut in rhs
358 -- when suspect is polymorphic or arrowtyped
359 -- So the required strictness properties are not observed.
360 -- At some point, must fix this properly.
364 (AnnNonRec bndr (fvs_scrut, scrut)) (fvs_rhs, rhs)
367 in trace ("WARNING: ignoring polymorphic case in interpreted mode.\n" ++
368 " Possibly due to strict polymorphic/functional constructor args.\n" ++
369 " Your program may leak space unexpectedly.\n")
370 (schemeE d s p new_expr)
374 {- Convert case .... of (# VoidRep'd-thing, a #) -> ...
376 case .... of a -> ...
377 Use a as the name of the binder too.
379 Also case .... of (# a #) -> ...
381 case .... of a -> ...
383 schemeE d s p (fvs, AnnCase scrut bndr [(DataAlt dc, [bind1, bind2], rhs)])
384 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind1)
385 = --trace "automagic mashing of case alts (# VoidRep, a #)" (
386 schemeE d s p (fvs, AnnCase scrut bind2 [(DEFAULT, [bind2], rhs)])
389 schemeE d s p (fvs, AnnCase scrut bndr [(DataAlt dc, [bind1], rhs)])
390 | isUnboxedTupleCon dc
391 = --trace "automagic mashing of case alts (# a #)" (
392 schemeE d s p (fvs, AnnCase scrut bind1 [(DEFAULT, [bind1], rhs)])
395 schemeE d s p (fvs, AnnCase scrut bndr alts)
397 -- Top of stack is the return itbl, as usual.
398 -- underneath it is the pointer to the alt_code BCO.
399 -- When an alt is entered, it assumes the returned value is
400 -- on top of the itbl.
403 -- Env and depth in which to compile the alts, not including
404 -- any vars bound by the alts themselves
405 d' = d + ret_frame_sizeW + taggedIdSizeW bndr
406 p' = addToFM p bndr (d' - 1)
408 scrut_primrep = typePrimRep (idType bndr)
410 | scrut_primrep == PtrRep
412 | scrut_primrep `elem`
413 [CharRep, AddrRep, WordRep, IntRep, FloatRep, DoubleRep,
414 VoidRep, Int8Rep, Int16Rep, Int32Rep, Int64Rep,
415 Word8Rep, Word16Rep, Word32Rep, Word64Rep]
418 = pprPanic "ByteCodeGen.schemeE" (ppr scrut_primrep)
420 -- given an alt, return a discr and code for it.
421 codeAlt alt@(discr, binds_f, rhs)
423 = let (unpack_code, d_after_unpack, p_after_unpack)
424 = mkUnpackCode (filter (not.isTyVar) binds_f) d' p'
425 in schemeE d_after_unpack s p_after_unpack rhs
426 `thenBc` \ rhs_code ->
427 returnBc (my_discr alt, unpack_code `appOL` rhs_code)
429 = ASSERT(null binds_f)
430 schemeE d' s p' rhs `thenBc` \ rhs_code ->
431 returnBc (my_discr alt, rhs_code)
433 my_discr (DEFAULT, binds, rhs) = NoDiscr
434 my_discr (DataAlt dc, binds, rhs)
435 | isUnboxedTupleCon dc
436 = unboxedTupleException
438 = DiscrP (dataConTag dc - fIRST_TAG)
439 my_discr (LitAlt l, binds, rhs)
440 = case l of MachInt i -> DiscrI (fromInteger i)
441 MachFloat r -> DiscrF (fromRational r)
442 MachDouble r -> DiscrD (fromRational r)
443 MachChar i -> DiscrI i
444 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
447 | not isAlgCase = Nothing
449 = case [dc | (DataAlt dc, _, _) <- alts] of
451 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
454 mapBc codeAlt alts `thenBc` \ alt_stuff ->
455 mkMultiBranch maybe_ncons alt_stuff `thenBc` \ alt_final ->
457 alt_final_ac = ARGCHECK (taggedIdSizeW bndr) `consOL` alt_final
458 alt_bco_name = getName bndr
459 alt_bco = mkProtoBCO alt_bco_name alt_final_ac (Left alts)
461 schemeE (d + ret_frame_sizeW)
462 (d + ret_frame_sizeW) p scrut `thenBc` \ scrut_code ->
464 emitBc alt_bco `thenBc_`
465 returnBc (PUSH_AS alt_bco_name scrut_primrep `consOL` scrut_code)
468 schemeE d s p (fvs, AnnNote note body)
472 = pprPanic "ByteCodeGen.schemeE: unhandled case"
473 (pprCoreExpr (deAnnotate other))
476 -- Compile code to do a tail call. Specifically, push the fn,
477 -- slide the on-stack app back down to the sequel depth,
478 -- and enter. Four cases:
481 -- An application "PrelGHC.tagToEnum# <type> unboxed-int".
482 -- The int will be on the stack. Generate a code sequence
483 -- to convert it to the relevant constructor, SLIDE and ENTER.
485 -- 1. A nullary constructor. Push its closure on the stack
486 -- and SLIDE and RETURN.
488 -- 2. (Another nasty hack). Spot (# a::VoidRep, b #) and treat
489 -- it simply as b -- since the representations are identical
490 -- (the VoidRep takes up zero stack space). Also, spot
491 -- (# b #) and treat it as b.
493 -- 3. Application of a non-nullary constructor, by defn saturated.
494 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
495 -- then the ptrs, and then do PACK and RETURN.
497 -- 4. Otherwise, it must be a function call. Push the args
498 -- right to left, SLIDE and ENTER.
500 schemeT :: Int -- Stack depth
501 -> Sequel -- Sequel depth
502 -> BCEnv -- stack env
508 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
509 -- = panic "schemeT ?!?!"
511 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate app)) ++ "\n") False
515 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
516 = pushAtom True d p arg `thenBc` \ (push, arg_words) ->
517 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
518 returnBc (push `appOL` tagToId_sequence
519 `appOL` mkSLIDE 1 (d+arg_words-s)
523 | is_con_call && null args_r_to_l
525 (PUSH_G (Left (getName con)) `consOL` mkSLIDE 1 (d-s))
530 | let isVoidRepAtom (_, AnnVar v) = VoidRep == typePrimRep (idType v)
531 isVoidRepAtom (_, AnnNote n e) = isVoidRepAtom e
532 in is_con_call && isUnboxedTupleCon con
533 && ( (length args_r_to_l == 2 && isVoidRepAtom (last (args_r_to_l)))
534 || (length args_r_to_l == 1)
536 = --trace (if length args_r_to_l == 1
537 -- then "schemeT: unboxed singleton"
538 -- else "schemeT: unboxed pair with Void first component") (
539 schemeT d s p (head args_r_to_l)
542 | Just (CCall ccall_spec) <- isFCallId_maybe fn
543 = generateCCall d s p ccall_spec fn args_r_to_l
547 = if is_con_call && isUnboxedTupleCon con
548 then unboxedTupleException
549 else do_pushery d (map snd args_final_r_to_l)
552 -- Detect and extract relevant info for the tagToEnum kludge.
553 maybe_is_tagToEnum_call
554 = let extract_constr_Names ty
555 = case splitTyConApp_maybe (repType ty) of
556 (Just (tyc, [])) | isDataTyCon tyc
557 -> map getName (tyConDataCons tyc)
558 other -> panic "maybe_is_tagToEnum_call.extract_constr_Ids"
561 (_, AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
562 -> case isPrimOpId_maybe v of
563 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
567 -- Extract the args (R->L) and fn
568 (args_r_to_l_raw, fn) = chomp app
572 AnnApp f a -> case chomp f of (az, f) -> (a:az, f)
573 AnnNote n e -> chomp e
574 other -> pprPanic "schemeT"
575 (ppr (deAnnotate (panic "schemeT.chomp", other)))
577 args_r_to_l = filter (not.isTypeAtom.snd) args_r_to_l_raw
578 isTypeAtom (AnnType _) = True
581 -- decide if this is a constructor call, and rearrange
582 -- args appropriately.
583 maybe_dcon = isDataConId_maybe fn
584 is_con_call = case maybe_dcon of Nothing -> False; Just _ -> True
585 (Just con) = maybe_dcon
591 = filter (not.isPtr.snd) args_r_to_l ++ filter (isPtr.snd) args_r_to_l
592 where isPtr = isFollowableRep . atomRep
594 -- make code to push the args and then do the SLIDE-ENTER thing
595 tag_when_push = not is_con_call
596 narg_words = sum (map (get_arg_szw . atomRep . snd) args_r_to_l)
597 get_arg_szw = if tag_when_push then taggedSizeW else untaggedSizeW
599 do_pushery d (arg:args)
600 = pushAtom tag_when_push d p arg `thenBc` \ (push, arg_words) ->
601 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
602 returnBc (push `appOL` more_push_code)
604 | Just (CCall ccall_spec) <- isFCallId_maybe fn
605 = panic "schemeT.do_pushery: unexpected ccall"
608 Just con -> returnBc (
609 (PACK con narg_words `consOL`
610 mkSLIDE 1 (d - narg_words - s)) `snocOL`
614 -> pushAtom True d p (AnnVar fn)
615 `thenBc` \ (push, arg_words) ->
616 returnBc (push `appOL` mkSLIDE (narg_words+arg_words)
622 {- Deal with a CCall. Taggedly push the args onto the stack R->L,
623 deferencing ForeignObj#s and (ToDo: adjusting addrs to point to
624 payloads in Ptr/Byte arrays). Then, generate the marshalling
625 (machine) code for the ccall, and create bytecodes to call that and
626 then return in the right way.
628 generateCCall :: Int -> Sequel -- stack and sequel depths
630 -> CCallSpec -- where to call
631 -> Id -- of target, for type info
632 -> [AnnExpr Id VarSet] -- args (atoms)
635 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
638 addr_usizeW = untaggedSizeW AddrRep
639 addr_tsizeW = taggedSizeW AddrRep
641 -- Get the args on the stack, with tags and suitably
642 -- dereferenced for the CCall. For each arg, return the
643 -- depth to the first word of the bits for that arg, and the
644 -- PrimRep of what was actually pushed.
646 pargs d [] = returnBc []
648 = let rep_arg = atomRep a
650 -- Don't push the FO; instead push the Addr# it
653 -> pushAtom False{-irrelevant-} d p a
654 `thenBc` \ (push_fo, _) ->
655 let foro_szW = taggedSizeW ForeignObjRep
656 d_now = d + addr_tsizeW
657 code = push_fo `appOL` toOL [
658 UPK_TAG addr_usizeW 0 0,
659 SLIDE addr_tsizeW foro_szW
661 in pargs d_now az `thenBc` \ rest ->
662 returnBc ((code, AddrRep) : rest)
663 -- Default case: push taggedly, but otherwise intact.
665 -> pushAtom True d p a `thenBc` \ (code_a, sz_a) ->
666 pargs (d+sz_a) az `thenBc` \ rest ->
667 returnBc ((code_a, rep_arg) : rest)
669 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
671 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
673 push_args = concatOL pushs_arg
674 d_after_args = d0 + sum (map taggedSizeW a_reps_pushed_r_to_l)
676 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
677 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
679 = reverse (tail a_reps_pushed_r_to_l)
681 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
682 -- push_args is the code to do that.
683 -- d_after_args is the stack depth once the args are on.
685 -- Get the result rep.
686 (returns_void, r_rep)
687 = case maybe_getCCallReturnRep (idType fn) of
688 Nothing -> (True, VoidRep)
689 Just rr -> (False, rr)
691 Because the Haskell stack grows down, the a_reps refer to
692 lowest to highest addresses in that order. The args for the call
693 are on the stack. Now push an unboxed, tagged Addr# indicating
694 the C function to call. Then push a dummy placeholder for the
695 result. Finally, emit a CCALL insn with an offset pointing to the
696 Addr# just pushed, and a literal field holding the mallocville
697 address of the piece of marshalling code we generate.
698 So, just prior to the CCALL insn, the stack looks like this
699 (growing down, as usual):
704 Addr# address_of_C_fn
705 <placeholder-for-result#> (must be an unboxed type)
707 The interpreter then calls the marshall code mentioned
708 in the CCALL insn, passing it (& <placeholder-for-result#>),
709 that is, the addr of the topmost word in the stack.
710 When this returns, the placeholder will have been
711 filled in. The placeholder is slid down to the sequel
712 depth, and we RETURN.
714 This arrangement makes it simple to do f-i-dynamic since the Addr#
715 value is the first arg anyway. It also has the virtue that the
716 stack is GC-understandable at all times.
718 The marshalling code is generated specifically for this
719 call site, and so knows exactly the (Haskell) stack
720 offsets of the args, fn address and placeholder. It
721 copies the args to the C stack, calls the stacked addr,
722 and parks the result back in the placeholder. The interpreter
723 calls it as a normal C call, assuming it has a signature
724 void marshall_code ( StgWord* ptr_to_top_of_stack )
726 -- resolve static address
730 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
732 -> ioToBc (lookupSymbol (_UNPK_ target)) `thenBc` \res ->
734 Just aa -> case aa of Ptr a# -> returnBc (True, A# a#)
735 Nothing -> returnBc invalid
739 invalid = pprPanic ("ByteCodeGen.generateCCall: unfindable "
740 ++ "symbol or otherwise invalid target")
743 get_target_info `thenBc` \ (is_static, static_target_addr) ->
746 -- Get the arg reps, zapping the leading Addr# in the dynamic case
747 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
748 | is_static = a_reps_pushed_RAW
749 | otherwise = if null a_reps_pushed_RAW
750 then panic "ByteCodeGen.generateCCall: dyn with no args"
751 else tail a_reps_pushed_RAW
754 (push_Addr, d_after_Addr)
756 = (toOL [PUSH_UBX (Right static_target_addr) addr_usizeW,
757 PUSH_TAG addr_usizeW],
758 d_after_args + addr_tsizeW)
759 | otherwise -- is already on the stack
760 = (nilOL, d_after_args)
762 -- Push the return placeholder. For a call returning nothing,
763 -- this is a VoidRep (tag).
764 r_usizeW = untaggedSizeW r_rep
765 r_tsizeW = taggedSizeW r_rep
766 d_after_r = d_after_Addr + r_tsizeW
767 r_lit = mkDummyLiteral r_rep
768 push_r = (if returns_void
770 else unitOL (PUSH_UBX (Left r_lit) r_usizeW))
772 unitOL (PUSH_TAG r_usizeW)
774 -- generate the marshalling code we're going to call
777 arg1_offW = r_tsizeW + addr_tsizeW
778 args_offW = map (arg1_offW +)
779 (init (scanl (+) 0 (map taggedSizeW a_reps)))
781 ioToBc (mkMarshalCode cconv
782 (r_offW, r_rep) addr_offW
783 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
784 recordMallocBc addr_of_marshaller `thenBc_`
787 do_call = unitOL (CCALL addr_of_marshaller)
789 wrapup = mkSLIDE r_tsizeW (d_after_r - r_tsizeW - s)
790 `snocOL` RETURN r_rep
792 --trace (show (arg1_offW, args_offW , (map taggedSizeW a_reps) )) (
795 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
800 -- Make a dummy literal, to be used as a placeholder for FFI return
801 -- values on the stack.
802 mkDummyLiteral :: PrimRep -> Literal
806 DoubleRep -> MachDouble 0
807 FloatRep -> MachFloat 0
808 AddrRep | taggedSizeW AddrRep == taggedSizeW WordRep -> MachWord 0
809 _ -> pprPanic "mkDummyLiteral" (ppr pr)
813 -- PrelGHC.Char# -> PrelGHC.State# PrelGHC.RealWorld
814 -- -> (# PrelGHC.State# PrelGHC.RealWorld, PrelGHC.Int# #)
817 -- and check that an unboxed pair is returned wherein the first arg is VoidRep'd.
819 -- Alternatively, for call-targets returning nothing, convert
821 -- PrelGHC.Char# -> PrelGHC.State# PrelGHC.RealWorld
822 -- -> (# PrelGHC.State# PrelGHC.RealWorld #)
826 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
827 maybe_getCCallReturnRep fn_ty
828 = let (a_tys, r_ty) = splitRepFunTys fn_ty
830 = if length r_reps == 1 then Nothing else Just (r_reps !! 1)
832 = case splitTyConApp_maybe (repType r_ty) of
833 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
835 ok = ( (length r_reps == 2 && VoidRep == head r_reps)
836 || r_reps == [VoidRep] )
837 && isUnboxedTupleTyCon r_tycon
838 && case maybe_r_rep_to_go of
840 Just r_rep -> r_rep /= PtrRep
841 -- if it was, it would be impossible
842 -- to create a valid return value
843 -- placeholder on the stack
844 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
847 --trace (showSDoc (ppr (a_reps, r_reps))) (
848 if ok then maybe_r_rep_to_go else blargh
851 atomRep (AnnVar v) = typePrimRep (idType v)
852 atomRep (AnnLit l) = literalPrimRep l
853 atomRep (AnnNote n b) = atomRep (snd b)
854 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
855 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
856 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
859 -- Compile code which expects an unboxed Int on the top of stack,
860 -- (call it i), and pushes the i'th closure in the supplied list
862 implement_tagToId :: [Name] -> BcM BCInstrList
863 implement_tagToId names
864 = ASSERT(not (null names))
865 getLabelsBc (length names) `thenBc` \ labels ->
866 getLabelBc `thenBc` \ label_fail ->
867 getLabelBc `thenBc` \ label_exit ->
868 zip4 labels (tail labels ++ [label_fail])
869 [0 ..] names `bind` \ infos ->
870 map (mkStep label_exit) infos `bind` \ steps ->
871 returnBc (concatOL steps
873 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
875 mkStep l_exit (my_label, next_label, n, name_for_n)
876 = toOL [LABEL my_label,
877 TESTEQ_I n next_label,
878 PUSH_G (Left name_for_n),
882 -- Make code to unpack the top-of-stack constructor onto the stack,
883 -- adding tags for the unboxed bits. Takes the PrimReps of the
884 -- constructor's arguments. off_h and off_s are travelling offsets
885 -- along the constructor and the stack.
887 -- Supposing a constructor in the heap has layout
889 -- Itbl p_1 ... p_i np_1 ... np_j
891 -- then we add to the stack, shown growing down, the following:
903 -- so that in the common case (ptrs only) a single UNPACK instr can
904 -- copy all the payload of the constr onto the stack with no further ado.
906 mkUnpackCode :: [Id] -- constr args
907 -> Int -- depth before unpack
908 -> BCEnv -- env before unpack
909 -> (BCInstrList, Int, BCEnv)
910 mkUnpackCode vars d p
911 = --trace ("mkUnpackCode: " ++ showSDocDebug (ppr vars)
912 -- ++ " --> " ++ show d' ++ "\n" ++ showSDocDebug (ppBCEnv p')
914 (code_p `appOL` code_np, d', p')
918 vreps = [(var, typePrimRep (idType var)) | var <- vars]
920 -- ptrs and nonptrs, forward
921 vreps_p = filter (isFollowableRep.snd) vreps
922 vreps_np = filter (not.isFollowableRep.snd) vreps
924 -- the order in which we will augment the environment
925 vreps_env = reverse vreps_p ++ reverse vreps_np
928 vreps_env_tszsw = map (taggedSizeW.snd) vreps_env
929 p' = addListToFM p (zip (map fst vreps_env)
930 (mkStackOffsets d vreps_env_tszsw))
931 d' = d + sum vreps_env_tszsw
933 -- code to unpack the ptrs
934 ptrs_szw = sum (map (untaggedSizeW.snd) vreps_p)
935 code_p | null vreps_p = nilOL
936 | otherwise = unitOL (UNPACK ptrs_szw)
938 -- code to unpack the nonptrs
939 vreps_env_uszw = sum (map (untaggedSizeW.snd) vreps_env)
940 code_np = do_nptrs vreps_env_uszw ptrs_szw (reverse (map snd vreps_np))
941 do_nptrs off_h off_s [] = nilOL
942 do_nptrs off_h off_s (npr:nprs)
943 | npr `elem` [IntRep, WordRep, FloatRep, DoubleRep, CharRep, AddrRep]
946 = pprPanic "ByteCodeGen.mkUnpackCode" (ppr npr)
948 approved = UPK_TAG usizeW (off_h-usizeW) off_s `consOL` theRest
949 theRest = do_nptrs (off_h-usizeW) (off_s + tsizeW) nprs
950 usizeW = untaggedSizeW npr
951 tsizeW = taggedSizeW npr
954 -- Push an atom onto the stack, returning suitable code & number of
955 -- stack words used. Pushes it either tagged or untagged, since
956 -- pushAtom is used to set up the stack prior to copying into the
957 -- heap for both APs (requiring tags) and constructors (which don't).
959 -- NB this means NO GC between pushing atoms for a constructor and
960 -- copying them into the heap. It probably also means that
961 -- tail calls MUST be of the form atom{atom ... atom} since if the
962 -- expression head was allowed to be arbitrary, there could be GC
963 -- in between pushing the arg atoms and completing the head.
964 -- (not sure; perhaps the allocate/doYouWantToGC interface means this
965 -- isn't a problem; but only if arbitrary graph construction for the
966 -- head doesn't leave this BCO, since GC might happen at the start of
967 -- each BCO (we consult doYouWantToGC there).
969 -- Blargh. JRS 001206
971 -- NB (further) that the env p must map each variable to the highest-
972 -- numbered stack slot for it. For example, if the stack has depth 4
973 -- and we tagged-ly push (v :: Int#) on it, the value will be in stack[4],
974 -- the tag in stack[5], the stack will have depth 6, and p must map v to
975 -- 5 and not to 4. Stack locations are numbered from zero, so a depth
976 -- 6 stack has valid words 0 .. 5.
978 pushAtom :: Bool -> Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
979 pushAtom tagged d p (AnnVar v)
981 | idPrimRep v == VoidRep
982 = if tagged then returnBc (unitOL (PUSH_TAG 0), 1)
983 else panic "ByteCodeGen.pushAtom(VoidRep,untaggedly)"
986 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
988 | Just primop <- isPrimOpId_maybe v
989 = returnBc (unitOL (PUSH_G (Right primop)), 1)
993 str = "\npushAtom " ++ showSDocDebug (ppr v)
994 ++ " :: " ++ showSDocDebug (pprType (idType v))
995 ++ ", depth = " ++ show d
996 ++ ", tagged = " ++ show tagged ++ ", env =\n" ++
997 showSDocDebug (ppBCEnv p)
998 ++ " --> words: " ++ show (snd result) ++ "\n" ++
999 showSDoc (nest 4 (vcat (map ppr (fromOL (fst result)))))
1000 ++ "\nendPushAtom " ++ showSDocDebug (ppr v)
1004 = case lookupBCEnv_maybe p v of
1005 Just d_v -> (toOL (nOfThem nwords (PUSH_L (d-d_v+sz_t-2))), nwords)
1006 Nothing -> ASSERT(sz_t == 1) (unitOL (PUSH_G (Left nm)), nwords)
1008 nm = case isDataConId_maybe v of
1010 Nothing -> getName v
1012 sz_t = taggedIdSizeW v
1013 sz_u = untaggedIdSizeW v
1014 nwords = if tagged then sz_t else sz_u
1018 pushAtom True d p (AnnLit lit)
1019 = pushAtom False d p (AnnLit lit) `thenBc` \ (ubx_code, ubx_size) ->
1020 returnBc (ubx_code `snocOL` PUSH_TAG ubx_size, 1 + ubx_size)
1022 pushAtom False d p (AnnLit lit)
1024 MachWord w -> code WordRep
1025 MachInt i -> code IntRep
1026 MachFloat r -> code FloatRep
1027 MachDouble r -> code DoubleRep
1028 MachChar c -> code CharRep
1029 MachStr s -> pushStr s
1032 = let size_host_words = untaggedSizeW rep
1033 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1037 = let getMallocvilleAddr
1039 CharStr s i -> returnBc (A# s)
1041 FastString _ l ba ->
1042 -- sigh, a string in the heap is no good to us.
1043 -- We need a static C pointer, since the type of
1044 -- a string literal is Addr#. So, copy the string
1045 -- into C land and introduce a memory leak
1046 -- at the same time.
1048 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1049 in ioToBc (mallocBytes (n+1)) `thenBc` \ (Ptr a#) ->
1050 recordMallocBc (A# a#) `thenBc_`
1052 do strncpy (Ptr a#) ba (fromIntegral n)
1053 writeCharOffAddr (A# a#) n '\0'
1056 other -> panic "ByteCodeGen.pushAtom.pushStr"
1058 getMallocvilleAddr `thenBc` \ addr ->
1059 -- Get the addr on the stack, untaggedly
1060 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1066 pushAtom tagged d p (AnnApp f (_, AnnType _))
1067 = pushAtom tagged d p (snd f)
1069 pushAtom tagged d p (AnnNote note e)
1070 = pushAtom tagged d p (snd e)
1072 pushAtom tagged d p (AnnLam x e)
1074 = pushAtom tagged d p (snd e)
1076 pushAtom tagged d p other
1077 = pprPanic "ByteCodeGen.pushAtom"
1078 (pprCoreExpr (deAnnotate (undefined, other)))
1080 foreign import "strncpy" strncpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1083 -- Given a bunch of alts code and their discrs, do the donkey work
1084 -- of making a multiway branch using a switch tree.
1085 -- What a load of hassle!
1086 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1087 -- a hint; generates better code
1088 -- Nothing is always safe
1089 -> [(Discr, BCInstrList)]
1091 mkMultiBranch maybe_ncons raw_ways
1092 = let d_way = filter (isNoDiscr.fst) raw_ways
1093 notd_ways = naturalMergeSortLe
1094 (\w1 w2 -> leAlt (fst w1) (fst w2))
1095 (filter (not.isNoDiscr.fst) raw_ways)
1097 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1098 mkTree [] range_lo range_hi = returnBc the_default
1100 mkTree [val] range_lo range_hi
1101 | range_lo `eqAlt` range_hi
1102 = returnBc (snd val)
1104 = getLabelBc `thenBc` \ label_neq ->
1105 returnBc (mkTestEQ (fst val) label_neq
1107 `appOL` unitOL (LABEL label_neq)
1108 `appOL` the_default))
1110 mkTree vals range_lo range_hi
1111 = let n = length vals `div` 2
1112 vals_lo = take n vals
1113 vals_hi = drop n vals
1114 v_mid = fst (head vals_hi)
1116 getLabelBc `thenBc` \ label_geq ->
1117 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1118 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1119 returnBc (mkTestLT v_mid label_geq
1121 `appOL` unitOL (LABEL label_geq)
1125 = case d_way of [] -> unitOL CASEFAIL
1128 -- None of these will be needed if there are no non-default alts
1129 (mkTestLT, mkTestEQ, init_lo, init_hi)
1131 = panic "mkMultiBranch: awesome foursome"
1133 = case fst (head notd_ways) of {
1134 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1135 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1138 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1139 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1142 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1143 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1146 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1147 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1149 DiscrP algMaxBound )
1152 (algMinBound, algMaxBound)
1153 = case maybe_ncons of
1154 Just n -> (0, n - 1)
1155 Nothing -> (minBound, maxBound)
1157 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1158 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1159 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1160 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1161 NoDiscr `eqAlt` NoDiscr = True
1164 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1165 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1166 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1167 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1168 NoDiscr `leAlt` NoDiscr = True
1171 isNoDiscr NoDiscr = True
1174 dec (DiscrI i) = DiscrI (i-1)
1175 dec (DiscrP i) = DiscrP (i-1)
1176 dec other = other -- not really right, but if you
1177 -- do cases on floating values, you'll get what you deserve
1179 -- same snotty comment applies to the following
1181 minD, maxD :: Double
1187 mkTree notd_ways init_lo init_hi
1191 %************************************************************************
1193 \subsection{Supporting junk for the compilation schemes}
1195 %************************************************************************
1199 -- Describes case alts
1207 instance Outputable Discr where
1208 ppr (DiscrI i) = int i
1209 ppr (DiscrF f) = text (show f)
1210 ppr (DiscrD d) = text (show d)
1211 ppr (DiscrP i) = int i
1212 ppr NoDiscr = text "DEF"
1215 -- Find things in the BCEnv (the what's-on-the-stack-env)
1216 -- See comment preceding pushAtom for precise meaning of env contents
1217 --lookupBCEnv :: BCEnv -> Id -> Int
1218 --lookupBCEnv env nm
1219 -- = case lookupFM env nm of
1220 -- Nothing -> pprPanic "lookupBCEnv"
1221 -- (ppr nm $$ char ' ' $$ vcat (map ppr (fmToList env)))
1224 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1225 lookupBCEnv_maybe = lookupFM
1228 taggedIdSizeW, untaggedIdSizeW :: Id -> Int
1229 taggedIdSizeW = taggedSizeW . typePrimRep . idType
1230 untaggedIdSizeW = untaggedSizeW . typePrimRep . idType
1232 unboxedTupleException :: a
1233 unboxedTupleException
1236 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1237 "\tto foreign import/export decls in source. Workaround:\n" ++
1238 "\tcompile this module to a .o file, then restart session."))
1241 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1246 %************************************************************************
1248 \subsection{The bytecode generator's monad}
1250 %************************************************************************
1254 = BcM_State { bcos :: [ProtoBCO Name], -- accumulates completed BCOs
1255 nextlabel :: Int, -- for generating local labels
1256 malloced :: [Addr] } -- ptrs malloced for current BCO
1257 -- Should be free()d when it is GCd
1258 type BcM r = BcM_State -> IO (BcM_State, r)
1260 ioToBc :: IO a -> BcM a
1261 ioToBc io st = do x <- io
1264 runBc :: BcM_State -> BcM r -> IO (BcM_State, r)
1265 runBc st0 m = do (st1, res) <- m st0
1268 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1269 thenBc expr cont st0
1270 = do (st1, q) <- expr st0
1271 (st2, r) <- cont q st1
1274 thenBc_ :: BcM a -> BcM b -> BcM b
1275 thenBc_ expr cont st0
1276 = do (st1, q) <- expr st0
1277 (st2, r) <- cont st1
1280 returnBc :: a -> BcM a
1281 returnBc result st = return (st, result)
1284 mapBc :: (a -> BcM b) -> [a] -> BcM [b]
1285 mapBc f [] = returnBc []
1287 = f x `thenBc` \ r ->
1288 mapBc f xs `thenBc` \ rs ->
1291 emitBc :: ([Addr] -> ProtoBCO Name) -> BcM ()
1293 = return (st{bcos = bco (malloced st) : bcos st, malloced=[]}, ())
1297 | not (null (malloced st))
1298 = panic "ByteCodeGen.newbcoBc: missed prior emitBc?"
1302 recordMallocBc :: Addr -> BcM ()
1304 = return (st{malloced = a : malloced st}, ())
1306 getLabelBc :: BcM Int
1308 = return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1310 getLabelsBc :: Int -> BcM [Int]
1312 = let ctr = nextlabel st
1313 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])