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 CStrings ( CLabelString )
29 import CoreFVs ( freeVars )
30 import Type ( typePrimRep, splitTyConApp_maybe, isTyVarTy, splitForAllTys )
31 import DataCon ( dataConTag, fIRST_TAG, dataConTyCon,
32 dataConWrapId, isUnboxedTupleCon )
33 import TyCon ( TyCon(..), tyConFamilySize, isDataTyCon, tyConDataCons,
34 isFunTyCon, isUnboxedTupleTyCon )
35 import Class ( Class, classTyCon )
36 import Type ( Type, repType, splitRepFunTys )
37 import Util ( zipEqual, zipWith4Equal, naturalMergeSortLe, nOfThem )
38 import Var ( isTyVar )
39 import VarSet ( VarSet, varSetElems )
40 import PrimRep ( getPrimRepSize, isFollowableRep )
41 import CmdLineOpts ( DynFlags, DynFlag(..) )
42 import ErrUtils ( showPass, dumpIfSet_dyn )
43 import Unique ( mkPseudoUnique3 )
44 import FastString ( FastString(..) )
45 import Panic ( GhcException(..) )
46 import SMRep ( fixedHdrSize )
47 import PprType ( pprType )
48 import ByteCodeInstr ( BCInstr(..), ProtoBCO(..), nameOfProtoBCO, bciStackUse )
49 import ByteCodeItbls ( ItblEnv, mkITbls )
50 import ByteCodeLink ( UnlinkedBCO, UnlinkedBCOExpr, assembleBCO,
51 ClosureEnv, HValue, filterNameMap,
52 iNTERP_STACK_CHECK_THRESH )
53 import ByteCodeFFI ( taggedSizeW, untaggedSizeW, mkMarshalCode )
54 import Linker ( lookupSymbol )
56 import List ( intersperse, sortBy, zip4 )
57 import Foreign ( Ptr(..), mallocBytes )
58 import Addr ( Addr(..), nullAddr, addrToInt, writeCharOffAddr )
59 import CTypes ( CInt )
60 import Exception ( throwDyn )
62 import PrelBase ( Int(..) )
63 import PrelGHC ( ByteArray# )
64 import IOExts ( unsafePerformIO )
65 import PrelIOBase ( IO(..) )
69 %************************************************************************
71 \subsection{Functions visible from outside this module.}
73 %************************************************************************
77 byteCodeGen :: DynFlags
80 -> IO ([UnlinkedBCO], ItblEnv)
81 byteCodeGen dflags binds local_tycons local_classes
82 = do showPass dflags "ByteCodeGen"
83 let tycs = local_tycons ++ map classTyCon local_classes
84 itblenv <- mkITbls tycs
86 let flatBinds = concatMap getBind binds
87 getBind (NonRec bndr rhs) = [(bndr, freeVars rhs)]
88 getBind (Rec binds) = [(bndr, freeVars rhs) | (bndr,rhs) <- binds]
90 (BcM_State proto_bcos final_ctr, ())
91 <- runBc (BcM_State [] 0)
92 (mapBc (schemeR True) flatBinds `thenBc_` returnBc ())
94 dumpIfSet_dyn dflags Opt_D_dump_BCOs
95 "Proto-bcos" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
97 bcos <- mapM assembleBCO proto_bcos
99 return (bcos, itblenv)
102 -- Returns: (the root BCO for this expression,
103 -- a list of auxilary BCOs resulting from compiling closures)
104 coreExprToBCOs :: DynFlags
106 -> IO UnlinkedBCOExpr
107 coreExprToBCOs dflags expr
108 = do showPass dflags "ByteCodeGen"
110 -- create a totally bogus name for the top-level BCO; this
111 -- should be harmless, since it's never used for anything
112 let invented_id = mkSysLocal SLIT("Expr-Top-Level") (mkPseudoUnique3 0)
113 (panic "invented_id's type")
114 let invented_name = idName invented_id
116 (BcM_State all_proto_bcos final_ctr, ())
117 <- runBc (BcM_State [] 0)
118 (schemeR True (invented_id, freeVars expr))
120 dumpIfSet_dyn dflags Opt_D_dump_BCOs
121 "Proto-bcos" (vcat (intersperse (char ' ') (map ppr all_proto_bcos)))
124 = case filter ((== invented_name).nameOfProtoBCO) all_proto_bcos of
125 [root_bco] -> root_bco
127 = filter ((/= invented_name).nameOfProtoBCO) all_proto_bcos
129 auxiliary_bcos <- mapM assembleBCO auxiliary_proto_bcos
130 root_bco <- assembleBCO root_proto_bco
132 return (root_bco, auxiliary_bcos)
135 %************************************************************************
137 \subsection{Compilation schema for the bytecode generator.}
139 %************************************************************************
143 type BCInstrList = OrdList BCInstr
145 type Sequel = Int -- back off to this depth before ENTER
147 -- Maps Ids to the offset from the stack _base_ so we don't have
148 -- to mess with it after each push/pop.
149 type BCEnv = FiniteMap Id Int -- To find vars on the stack
151 ppBCEnv :: BCEnv -> SDoc
154 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
157 pp_one (var, offset) = int offset <> colon <+> ppr var
158 cmp_snd x y = compare (snd x) (snd y)
160 -- Create a BCO and do a spot of peephole optimisation on the insns
162 mkProtoBCO nm instrs_ordlist origin
163 = ProtoBCO nm maybe_with_stack_check origin
165 -- Overestimate the stack usage (in words) of this BCO,
166 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
167 -- stack check. (The interpreter always does a stack check
168 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
169 -- BCO anyway, so we only need to add an explicit on in the
170 -- (hopefully rare) cases when the (overestimated) stack use
171 -- exceeds iNTERP_STACK_CHECK_THRESH.
172 maybe_with_stack_check
173 | stack_overest >= 65535
174 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
176 | stack_overest >= iNTERP_STACK_CHECK_THRESH
177 = (STKCHECK stack_overest) : peep_d
179 = peep_d -- the supposedly common case
181 stack_overest = sum (map bciStackUse peep_d)
182 + 10 {- just to be really really sure -}
185 -- Merge local pushes
186 peep_d = peep (fromOL instrs_ordlist)
188 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
189 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
190 peep (PUSH_L off1 : PUSH_L off2 : rest)
191 = PUSH_LL off1 (off2-1) : peep rest
198 -- Compile code for the right hand side of a let binding.
199 -- Park the resulting BCO in the monad. Also requires the
200 -- variable to which this value was bound, so as to give the
201 -- resulting BCO a name. Bool indicates top-levelness.
203 schemeR :: Bool -> (Id, AnnExpr Id VarSet) -> BcM ()
204 schemeR is_top (nm, rhs)
208 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
209 $$ pprCoreExpr (deAnnotate rhs)
215 = schemeR_wrk is_top rhs nm (collect [] rhs)
218 collect xs (_, AnnNote note e)
220 collect xs (_, AnnLam x e)
221 = collect (if isTyVar x then xs else (x:xs)) e
222 collect xs not_lambda
223 = (reverse xs, not_lambda)
225 schemeR_wrk is_top original_body nm (args, body)
226 | Just dcon <- maybe_toplevel_null_con_rhs
227 = --trace ("nullary constructor! " ++ showSDocDebug (ppr nm)) (
228 emitBc (mkProtoBCO (getName nm) (toOL [PACK dcon 0, ENTER])
229 (Right original_body))
233 = let fvs = filter (not.isTyVar) (varSetElems (fst original_body))
234 all_args = reverse args ++ fvs
235 szsw_args = map taggedIdSizeW all_args
236 szw_args = sum szsw_args
237 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
238 argcheck = unitOL (ARGCHECK szw_args)
240 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
241 emitBc (mkProtoBCO (getName nm) (appOL argcheck body_code)
242 (Right original_body))
245 maybe_toplevel_null_con_rhs
246 | is_top && null args
249 -> case isDataConId_maybe v_wrk of
251 Just dc_wrk | nm == dataConWrapId dc_wrk
259 -- Let szsw be the sizes in words of some items pushed onto the stack,
260 -- which has initial depth d'. Return the values which the stack environment
261 -- should map these items to.
262 mkStackOffsets :: Int -> [Int] -> [Int]
263 mkStackOffsets original_depth szsw
264 = map (subtract 1) (tail (scanl (+) original_depth szsw))
266 -- Compile code to apply the given expression to the remaining args
267 -- on the stack, returning a HNF.
268 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr Id VarSet -> BcM BCInstrList
270 -- Delegate tail-calls to schemeT.
271 schemeE d s p e@(fvs, AnnApp f a)
272 = schemeT d s p (fvs, AnnApp f a)
274 schemeE d s p e@(fvs, AnnVar v)
275 | isFollowableRep v_rep
276 = -- Ptr-ish thing; push it in the normal way
277 schemeT d s p (fvs, AnnVar v)
280 = -- returning an unboxed value. Heave it on the stack, SLIDE, and RETURN.
281 let (push, szw) = pushAtom True d p (AnnVar v)
282 in returnBc (push -- value onto stack
283 `appOL` mkSLIDE szw (d-s) -- clear to sequel
284 `snocOL` RETURN v_rep) -- go
286 v_rep = typePrimRep (idType v)
288 schemeE d s p (fvs, AnnLit literal)
289 = let (push, szw) = pushAtom True d p (AnnLit literal)
290 l_rep = literalPrimRep literal
291 in returnBc (push -- value onto stack
292 `appOL` mkSLIDE szw (d-s) -- clear to sequel
293 `snocOL` RETURN l_rep) -- go
295 schemeE d s p (fvs, AnnLet binds b)
296 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
297 AnnRec xs_n_rhss -> unzip xs_n_rhss
299 fvss = map (filter (not.isTyVar).varSetElems.fst) rhss
301 -- Sizes of tagged free vars, + 1 for the fn
302 sizes = map (\rhs_fvs -> 1 + sum (map taggedIdSizeW rhs_fvs)) fvss
304 -- This p', d' defn is safe because all the items being pushed
305 -- are ptrs, so all have size 1. d' and p' reflect the stack
306 -- after the closures have been allocated in the heap (but not
307 -- filled in), and pointers to them parked on the stack.
308 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n 1)))
311 infos = zipE4 fvss sizes xs [n, n-1 .. 1]
312 zipE = zipEqual "schemeE"
313 zipE4 = zipWith4Equal "schemeE" (\a b c d -> (a,b,c,d))
315 -- ToDo: don't build thunks for things with no free variables
316 buildThunk dd ([], size, id, off)
317 = PUSH_G (Left (getName id))
318 `consOL` unitOL (MKAP (off+size-1) size)
319 buildThunk dd ((fv:fvs), size, id, off)
320 = case pushAtom True dd p' (AnnVar fv) of
321 (push_code, pushed_szw)
323 buildThunk (dd+pushed_szw) (fvs, size, id, off)
325 thunkCode = concatOL (map (buildThunk d') infos)
326 allocCode = toOL (map ALLOC sizes)
328 schemeE d' s p' b `thenBc` \ bodyCode ->
329 mapBc (schemeR False) (zip xs rhss) `thenBc_`
330 returnBc (allocCode `appOL` thunkCode `appOL` bodyCode)
336 schemeE d s p (fvs_case, AnnCase (fvs_scrut, scrut) bndr
337 [(DEFAULT, [], (fvs_rhs, rhs))])
339 | let isFunType var_type
340 = case splitTyConApp_maybe var_type of
341 Just (tycon,_) | isFunTyCon tycon -> True
343 ty_bndr = repType (idType bndr)
344 in isFunType ty_bndr || isTyVarTy ty_bndr
347 -- case scrut::suspect of bndr { DEFAULT -> rhs }
349 -- let bndr = scrut in rhs
350 -- when suspect is polymorphic or arrowtyped
351 -- So the required strictness properties are not observed.
352 -- At some point, must fix this properly.
356 (AnnNonRec bndr (fvs_scrut, scrut)) (fvs_rhs, rhs)
359 in trace ("WARNING: ignoring polymorphic case in interpreted mode.\n" ++
360 " Possibly due to strict polymorphic/functional constructor args.\n" ++
361 " Your program may leak space unexpectedly.\n")
362 -- ++ showSDoc (char ' ' $$ pprCoreExpr (deAnnotate new_expr) $$ char ' '))
363 (schemeE d s p new_expr)
367 {- Convert case .... of (# VoidRep'd-thing, a #) -> ...
369 case .... of a -> ...
370 Use a as the name of the binder too.
372 Also case .... of (# a #) -> ...
374 case .... of a -> ...
376 schemeE d s p (fvs, AnnCase scrut bndr [(DataAlt dc, [bind1, bind2], rhs)])
377 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind1)
378 = --trace "automagic mashing of case alts (# VoidRep, a #)" (
379 schemeE d s p (fvs, AnnCase scrut bind2 [(DEFAULT, [bind2], rhs)])
382 schemeE d s p (fvs, AnnCase scrut bndr [(DataAlt dc, [bind1], rhs)])
383 | isUnboxedTupleCon dc
384 = --trace "automagic mashing of case alts (# a #)" (
385 schemeE d s p (fvs, AnnCase scrut bind1 [(DEFAULT, [bind1], rhs)])
388 schemeE d s p (fvs, AnnCase scrut bndr alts)
390 -- Top of stack is the return itbl, as usual.
391 -- underneath it is the pointer to the alt_code BCO.
392 -- When an alt is entered, it assumes the returned value is
393 -- on top of the itbl.
396 -- Env and depth in which to compile the alts, not including
397 -- any vars bound by the alts themselves
398 d' = d + ret_frame_sizeW + taggedIdSizeW bndr
399 p' = addToFM p bndr (d' - 1)
401 scrut_primrep = typePrimRep (idType bndr)
403 | scrut_primrep == PtrRep
405 | scrut_primrep `elem`
406 [CharRep, AddrRep, WordRep, IntRep, FloatRep, DoubleRep,
407 VoidRep, Int8Rep, Int16Rep, Int32Rep, Int64Rep,
408 Word8Rep, Word16Rep, Word32Rep, Word64Rep]
411 = pprPanic "ByteCodeGen.schemeE" (ppr scrut_primrep)
413 -- given an alt, return a discr and code for it.
414 codeAlt alt@(discr, binds_f, rhs)
416 = let (unpack_code, d_after_unpack, p_after_unpack)
417 = mkUnpackCode (filter (not.isTyVar) binds_f) d' p'
418 in schemeE d_after_unpack s p_after_unpack rhs
419 `thenBc` \ rhs_code ->
420 returnBc (my_discr alt, unpack_code `appOL` rhs_code)
422 = ASSERT(null binds_f)
423 schemeE d' s p' rhs `thenBc` \ rhs_code ->
424 returnBc (my_discr alt, rhs_code)
426 my_discr (DEFAULT, binds, rhs) = NoDiscr
427 my_discr (DataAlt dc, binds, rhs)
428 | isUnboxedTupleCon dc
429 = unboxedTupleException
431 = DiscrP (dataConTag dc - fIRST_TAG)
432 my_discr (LitAlt l, binds, rhs)
433 = case l of MachInt i -> DiscrI (fromInteger i)
434 MachFloat r -> DiscrF (fromRational r)
435 MachDouble r -> DiscrD (fromRational r)
436 MachChar i -> DiscrI i
437 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
440 | not isAlgCase = Nothing
442 = case [dc | (DataAlt dc, _, _) <- alts] of
444 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
447 mapBc codeAlt alts `thenBc` \ alt_stuff ->
448 mkMultiBranch maybe_ncons alt_stuff `thenBc` \ alt_final ->
450 alt_final_ac = ARGCHECK (taggedIdSizeW bndr) `consOL` alt_final
451 alt_bco_name = getName bndr
452 alt_bco = mkProtoBCO alt_bco_name alt_final_ac (Left alts)
454 schemeE (d + ret_frame_sizeW)
455 (d + ret_frame_sizeW) p scrut `thenBc` \ scrut_code ->
457 emitBc alt_bco `thenBc_`
458 returnBc (PUSH_AS alt_bco_name scrut_primrep `consOL` scrut_code)
461 schemeE d s p (fvs, AnnNote note body)
465 = pprPanic "ByteCodeGen.schemeE: unhandled case"
466 (pprCoreExpr (deAnnotate other))
469 -- Compile code to do a tail call. Specifically, push the fn,
470 -- slide the on-stack app back down to the sequel depth,
471 -- and enter. Four cases:
474 -- An application "PrelGHC.tagToEnum# <type> unboxed-int".
475 -- The int will be on the stack. Generate a code sequence
476 -- to convert it to the relevant constructor, SLIDE and ENTER.
478 -- 1. A nullary constructor. Push its closure on the stack
479 -- and SLIDE and RETURN.
481 -- 2. (Another nasty hack). Spot (# a::VoidRep, b #) and treat
482 -- it simply as b -- since the representations are identical
483 -- (the VoidRep takes up zero stack space). Also, spot
484 -- (# b #) and treat it as b.
486 -- 3. Application of a non-nullary constructor, by defn saturated.
487 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
488 -- then the ptrs, and then do PACK and RETURN.
490 -- 4. Otherwise, it must be a function call. Push the args
491 -- right to left, SLIDE and ENTER.
493 schemeT :: Int -- Stack depth
494 -> Sequel -- Sequel depth
495 -> BCEnv -- stack env
501 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
502 -- = panic "schemeT ?!?!"
504 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate app)) ++ "\n") False
508 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
509 = pushAtom True d p arg `bind` \ (push, arg_words) ->
510 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
511 returnBc (push `appOL` tagToId_sequence
512 `appOL` mkSLIDE 1 (d+arg_words-s)
516 | is_con_call && null args_r_to_l
518 (PUSH_G (Left (getName con)) `consOL` mkSLIDE 1 (d-s))
523 | let isVoidRepAtom (_, AnnVar v) = VoidRep == typePrimRep (idType v)
524 isVoidRepAtom (_, AnnNote n e) = isVoidRepAtom e
525 in is_con_call && isUnboxedTupleCon con
526 && ( (length args_r_to_l == 2 && isVoidRepAtom (last (args_r_to_l)))
527 || (length args_r_to_l == 1)
529 = --trace (if length args_r_to_l == 1
530 -- then "schemeT: unboxed singleton"
531 -- else "schemeT: unboxed pair with Void first component") (
532 schemeT d s p (head args_r_to_l)
535 | Just (CCall ccall_spec) <- isFCallId_maybe fn
536 = generateCCall d s p ccall_spec fn args_r_to_l
540 = if is_con_call && isUnboxedTupleCon con
541 then returnBc unboxedTupleException
542 else code `seq` returnBc code
545 -- Detect and extract relevant info for the tagToEnum kludge.
546 maybe_is_tagToEnum_call
547 = let extract_constr_Names ty
548 = case splitTyConApp_maybe (repType ty) of
549 (Just (tyc, [])) | isDataTyCon tyc
550 -> map getName (tyConDataCons tyc)
551 other -> panic "maybe_is_tagToEnum_call.extract_constr_Ids"
554 (_, AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
555 -> case isPrimOpId_maybe v of
556 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
560 -- Extract the args (R->L) and fn
561 (args_r_to_l_raw, fn) = chomp app
565 AnnApp f a -> case chomp f of (az, f) -> (a:az, f)
566 AnnNote n e -> chomp e
567 other -> pprPanic "schemeT"
568 (ppr (deAnnotate (panic "schemeT.chomp", other)))
570 args_r_to_l = filter (not.isTypeAtom.snd) args_r_to_l_raw
571 isTypeAtom (AnnType _) = True
574 -- decide if this is a constructor call, and rearrange
575 -- args appropriately.
576 maybe_dcon = isDataConId_maybe fn
577 is_con_call = case maybe_dcon of Nothing -> False; Just _ -> True
578 (Just con) = maybe_dcon
584 = filter (not.isPtr.snd) args_r_to_l ++ filter (isPtr.snd) args_r_to_l
585 where isPtr = isFollowableRep . atomRep
587 -- make code to push the args and then do the SLIDE-ENTER thing
588 code = do_pushery d (map snd args_final_r_to_l)
589 tag_when_push = not is_con_call
590 narg_words = sum (map (get_arg_szw . atomRep . snd) args_r_to_l)
591 get_arg_szw = if tag_when_push then taggedSizeW else untaggedSizeW
593 do_pushery d (arg:args)
594 = let (push, arg_words) = pushAtom tag_when_push d p arg
595 in push `appOL` do_pushery (d+arg_words) args
597 | Just (CCall ccall_spec) <- isFCallId_maybe fn
598 = panic "schemeT.do_pushery: unexpected ccall"
602 Just con -> PACK con narg_words `consOL` (
603 mkSLIDE 1 (d - narg_words - s) `snocOL` ENTER)
605 -> let (push, arg_words) = pushAtom True d p (AnnVar fn)
607 `appOL` mkSLIDE (narg_words+arg_words)
613 {- Deal with a CCall. Taggedly push the args onto the stack R->L,
614 deferencing ForeignObj#s and (ToDo: adjusting addrs to point to
615 payloads in Ptr/Byte arrays). Then, generate the marshalling
616 (machine) code for the ccall, and create bytecodes to call that and
617 then return in the right way.
619 generateCCall :: Int -> Sequel -- stack and sequel depths
621 -> CCallSpec -- where to call
622 -> Id -- of target, for type info
623 -> [AnnExpr Id VarSet] -- args (atoms)
626 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
629 addr_usizeW = untaggedSizeW AddrRep
630 addr_tsizeW = taggedSizeW AddrRep
632 -- Get the args on the stack, with tags and suitably
633 -- dereferenced for the CCall. For each arg, return the
634 -- depth to the first word of the bits for that arg, and the
635 -- PrimRep of what was actually pushed.
639 = let rep_arg = atomRep a
641 -- Don't push the FO; instead push the Addr# it
644 -> let foro_szW = taggedSizeW ForeignObjRep
645 push_fo = fst (pushAtom False{-irrelevant-} d p a)
646 d_now = d + addr_tsizeW
647 code = push_fo `appOL` toOL [
648 UPK_TAG addr_usizeW 0 0,
649 SLIDE addr_tsizeW foro_szW
651 in (code, AddrRep) : f d_now az
652 -- Default case: push taggedly, but otherwise intact.
654 -> let (code_a, sz_a) = pushAtom True d p a
655 in (code_a, rep_arg) : f (d+sz_a) az
657 (pushs_arg, a_reps_pushed_r_to_l) = unzip (f d0 args_r_to_l)
659 push_args = concatOL pushs_arg
660 d_after_args = d0 + sum (map taggedSizeW a_reps_pushed_r_to_l)
662 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
663 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
665 = reverse (tail a_reps_pushed_r_to_l)
667 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
668 -- push_args is the code to do that.
669 -- d_after_args is the stack depth once the args are on.
671 -- Get the result rep.
672 (returns_void, r_rep)
673 = case maybe_getCCallReturnRep (idType fn) of
674 Nothing -> (True, VoidRep)
675 Just rr -> (False, rr)
677 Because the Haskell stack grows down, the a_reps refer to
678 lowest to highest addresses in that order. The args for the call
679 are on the stack. Now push an unboxed, tagged Addr# indicating
680 the C function to call. Then push a dummy placeholder for the
681 result. Finally, emit a CCALL insn with an offset pointing to the
682 Addr# just pushed, and a literal field holding the mallocville
683 address of the piece of marshalling code we generate.
684 So, just prior to the CCALL insn, the stack looks like this
685 (growing down, as usual):
690 Addr# address_of_C_fn
691 <placeholder-for-result#> (must be an unboxed type)
693 The interpreter then calls the marshall code mentioned
694 in the CCALL insn, passing it (& <placeholder-for-result#>),
695 that is, the addr of the topmost word in the stack.
696 When this returns, the placeholder will have been
697 filled in. The placeholder is slid down to the sequel
698 depth, and we RETURN.
700 This arrangement makes it simple to do f-i-dynamic since the Addr#
701 value is the first arg anyway. It also has the virtue that the
702 stack is GC-understandable at all times.
704 The marshalling code is generated specifically for this
705 call site, and so knows exactly the (Haskell) stack
706 offsets of the args, fn address and placeholder. It
707 copies the args to the C stack, calls the stacked addr,
708 and parks the result back in the placeholder. The interpreter
709 calls it as a normal C call, assuming it has a signature
710 void marshall_code ( StgWord* ptr_to_top_of_stack )
712 -- resolve static address
716 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
718 -> ioToBc (lookupSymbol (_UNPK_ target)) `thenBc` \res ->
720 Just aa -> case aa of Ptr a# -> returnBc (True, A# a#)
721 Nothing -> returnBc invalid
725 invalid = pprPanic ("ByteCodeGen.generateCCall: unfindable "
726 ++ "symbol or otherwise invalid target")
729 get_target_info `thenBc` \ (is_static, static_target_addr) ->
732 -- Get the arg reps, zapping the leading Addr# in the dynamic case
733 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
734 | is_static = a_reps_pushed_RAW
735 | otherwise = if null a_reps_pushed_RAW
736 then panic "ByteCodeGen.generateCCall: dyn with no args"
737 else tail a_reps_pushed_RAW
740 (push_Addr, d_after_Addr)
742 = (toOL [PUSH_UBX (Right static_target_addr) addr_usizeW,
743 PUSH_TAG addr_usizeW],
744 d_after_args + addr_tsizeW)
745 | otherwise -- is already on the stack
746 = (nilOL, d_after_args)
748 -- Push the return placeholder. For a call returning nothing,
749 -- this is a VoidRep (tag).
750 r_usizeW = untaggedSizeW r_rep
751 r_tsizeW = taggedSizeW r_rep
752 d_after_r = d_after_Addr + r_tsizeW
753 r_lit = mkDummyLiteral r_rep
754 push_r = (if returns_void
756 else unitOL (PUSH_UBX (Left r_lit) r_usizeW))
758 unitOL (PUSH_TAG r_usizeW)
760 -- generate the marshalling code we're going to call
763 arg1_offW = r_tsizeW + addr_tsizeW
764 args_offW = map (arg1_offW +)
765 (init (scanl (+) 0 (map taggedSizeW a_reps)))
767 ioToBc (mkMarshalCode cconv
768 (r_offW, r_rep) addr_offW
769 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
772 do_call = unitOL (CCALL addr_of_marshaller)
774 wrapup = mkSLIDE r_tsizeW (d_after_r - r_tsizeW - s)
775 `snocOL` RETURN r_rep
777 --trace (show (arg1_offW, args_offW , (map taggedSizeW a_reps) )) (
780 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
785 -- Make a dummy literal, to be used as a placeholder for FFI return
786 -- values on the stack.
787 mkDummyLiteral :: PrimRep -> Literal
791 DoubleRep -> MachDouble 0
792 FloatRep -> MachFloat 0
793 AddrRep | taggedSizeW AddrRep == taggedSizeW WordRep -> MachWord 0
794 _ -> pprPanic "mkDummyLiteral" (ppr pr)
798 -- PrelGHC.Char# -> PrelGHC.State# PrelGHC.RealWorld
799 -- -> (# PrelGHC.State# PrelGHC.RealWorld, PrelGHC.Int# #)
802 -- and check that an unboxed pair is returned wherein the first arg is VoidRep'd.
804 -- Alternatively, for call-targets returning nothing, convert
806 -- PrelGHC.Char# -> PrelGHC.State# PrelGHC.RealWorld
807 -- -> (# PrelGHC.State# PrelGHC.RealWorld #)
811 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
812 maybe_getCCallReturnRep fn_ty
813 = let (a_tys, r_ty) = splitRepFunTys fn_ty
815 = if length r_reps == 1 then Nothing else Just (r_reps !! 1)
817 = case splitTyConApp_maybe (repType r_ty) of
818 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
820 ok = ( (length r_reps == 2 && VoidRep == head r_reps)
821 || r_reps == [VoidRep] )
822 && isUnboxedTupleTyCon r_tycon
823 && case maybe_r_rep_to_go of
825 Just r_rep -> r_rep /= PtrRep
826 -- if it was, it would be impossible
827 -- to create a valid return value
828 -- placeholder on the stack
829 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
832 --trace (showSDoc (ppr (a_reps, r_reps))) (
833 if ok then maybe_r_rep_to_go else blargh
836 atomRep (AnnVar v) = typePrimRep (idType v)
837 atomRep (AnnLit l) = literalPrimRep l
838 atomRep (AnnNote n b) = atomRep (snd b)
839 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
840 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
841 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
844 -- Compile code which expects an unboxed Int on the top of stack,
845 -- (call it i), and pushes the i'th closure in the supplied list
847 implement_tagToId :: [Name] -> BcM BCInstrList
848 implement_tagToId names
849 = ASSERT(not (null names))
850 getLabelsBc (length names) `thenBc` \ labels ->
851 getLabelBc `thenBc` \ label_fail ->
852 getLabelBc `thenBc` \ label_exit ->
853 zip4 labels (tail labels ++ [label_fail])
854 [0 ..] names `bind` \ infos ->
855 map (mkStep label_exit) infos `bind` \ steps ->
856 returnBc (concatOL steps
858 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
860 mkStep l_exit (my_label, next_label, n, name_for_n)
861 = toOL [LABEL my_label,
862 TESTEQ_I n next_label,
863 PUSH_G (Left name_for_n),
867 -- Make code to unpack the top-of-stack constructor onto the stack,
868 -- adding tags for the unboxed bits. Takes the PrimReps of the
869 -- constructor's arguments. off_h and off_s are travelling offsets
870 -- along the constructor and the stack.
872 -- Supposing a constructor in the heap has layout
874 -- Itbl p_1 ... p_i np_1 ... np_j
876 -- then we add to the stack, shown growing down, the following:
888 -- so that in the common case (ptrs only) a single UNPACK instr can
889 -- copy all the payload of the constr onto the stack with no further ado.
891 mkUnpackCode :: [Id] -- constr args
892 -> Int -- depth before unpack
893 -> BCEnv -- env before unpack
894 -> (BCInstrList, Int, BCEnv)
895 mkUnpackCode vars d p
896 = --trace ("mkUnpackCode: " ++ showSDocDebug (ppr vars)
897 -- ++ " --> " ++ show d' ++ "\n" ++ showSDocDebug (ppBCEnv p')
899 (code_p `appOL` code_np, d', p')
903 vreps = [(var, typePrimRep (idType var)) | var <- vars]
905 -- ptrs and nonptrs, forward
906 vreps_p = filter (isFollowableRep.snd) vreps
907 vreps_np = filter (not.isFollowableRep.snd) vreps
909 -- the order in which we will augment the environment
910 vreps_env = reverse vreps_p ++ reverse vreps_np
913 vreps_env_tszsw = map (taggedSizeW.snd) vreps_env
914 p' = addListToFM p (zip (map fst vreps_env)
915 (mkStackOffsets d vreps_env_tszsw))
916 d' = d + sum vreps_env_tszsw
918 -- code to unpack the ptrs
919 ptrs_szw = sum (map (untaggedSizeW.snd) vreps_p)
920 code_p | null vreps_p = nilOL
921 | otherwise = unitOL (UNPACK ptrs_szw)
923 -- code to unpack the nonptrs
924 vreps_env_uszw = sum (map (untaggedSizeW.snd) vreps_env)
925 code_np = do_nptrs vreps_env_uszw ptrs_szw (reverse (map snd vreps_np))
926 do_nptrs off_h off_s [] = nilOL
927 do_nptrs off_h off_s (npr:nprs)
928 | npr `elem` [IntRep, WordRep, FloatRep, DoubleRep, CharRep, AddrRep]
931 = pprPanic "ByteCodeGen.mkUnpackCode" (ppr npr)
933 approved = UPK_TAG usizeW (off_h-usizeW) off_s `consOL` theRest
934 theRest = do_nptrs (off_h-usizeW) (off_s + tsizeW) nprs
935 usizeW = untaggedSizeW npr
936 tsizeW = taggedSizeW npr
939 -- Push an atom onto the stack, returning suitable code & number of
940 -- stack words used. Pushes it either tagged or untagged, since
941 -- pushAtom is used to set up the stack prior to copying into the
942 -- heap for both APs (requiring tags) and constructors (which don't).
944 -- NB this means NO GC between pushing atoms for a constructor and
945 -- copying them into the heap. It probably also means that
946 -- tail calls MUST be of the form atom{atom ... atom} since if the
947 -- expression head was allowed to be arbitrary, there could be GC
948 -- in between pushing the arg atoms and completing the head.
949 -- (not sure; perhaps the allocate/doYouWantToGC interface means this
950 -- isn't a problem; but only if arbitrary graph construction for the
951 -- head doesn't leave this BCO, since GC might happen at the start of
952 -- each BCO (we consult doYouWantToGC there).
954 -- Blargh. JRS 001206
956 -- NB (further) that the env p must map each variable to the highest-
957 -- numbered stack slot for it. For example, if the stack has depth 4
958 -- and we tagged-ly push (v :: Int#) on it, the value will be in stack[4],
959 -- the tag in stack[5], the stack will have depth 6, and p must map v to
960 -- 5 and not to 4. Stack locations are numbered from zero, so a depth
961 -- 6 stack has valid words 0 .. 5.
963 pushAtom :: Bool -> Int -> BCEnv -> AnnExpr' Id VarSet -> (BCInstrList, Int)
964 pushAtom tagged d p (AnnVar v)
966 | idPrimRep v == VoidRep
967 = if tagged then (unitOL (PUSH_TAG 0), 1)
968 else panic "ByteCodeGen.pushAtom(VoidRep,untaggedly)"
971 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
973 | Just primop <- isPrimOpId_maybe v
974 = (unitOL (PUSH_G (Right primop)), 1)
978 str = "\npushAtom " ++ showSDocDebug (ppr v)
979 ++ " :: " ++ showSDocDebug (pprType (idType v))
980 ++ ", depth = " ++ show d
981 ++ ", tagged = " ++ show tagged ++ ", env =\n" ++
982 showSDocDebug (ppBCEnv p)
983 ++ " --> words: " ++ show (snd result) ++ "\n" ++
984 showSDoc (nest 4 (vcat (map ppr (fromOL (fst result)))))
985 ++ "\nendPushAtom " ++ showSDocDebug (ppr v)
989 = case lookupBCEnv_maybe p v of
990 Just d_v -> (toOL (nOfThem nwords (PUSH_L (d-d_v+sz_t-2))), nwords)
991 Nothing -> ASSERT(sz_t == 1) (unitOL (PUSH_G (Left nm)), nwords)
993 nm = case isDataConId_maybe v of
997 sz_t = taggedIdSizeW v
998 sz_u = untaggedIdSizeW v
999 nwords = if tagged then sz_t else sz_u
1003 pushAtom True d p (AnnLit lit)
1004 = let (ubx_code, ubx_size) = pushAtom False d p (AnnLit lit)
1005 in (ubx_code `snocOL` PUSH_TAG ubx_size, 1 + ubx_size)
1007 pushAtom False d p (AnnLit lit)
1009 MachWord w -> code WordRep
1010 MachInt i -> code IntRep
1011 MachFloat r -> code FloatRep
1012 MachDouble r -> code DoubleRep
1013 MachChar c -> code CharRep
1014 MachStr s -> pushStr s
1017 = let size_host_words = untaggedSizeW rep
1018 in (unitOL (PUSH_UBX (Left lit) size_host_words), size_host_words)
1021 = let mallocvilleAddr
1025 FastString _ l ba ->
1026 -- sigh, a string in the heap is no good to us.
1027 -- We need a static C pointer, since the type of
1028 -- a string literal is Addr#. So, copy the string
1029 -- into C land and introduce a memory leak
1030 -- at the same time.
1032 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1033 in unsafePerformIO (
1034 do (Ptr a#) <- mallocBytes (n+1)
1035 strncpy (Ptr a#) ba (fromIntegral n)
1036 writeCharOffAddr (A# a#) n '\0'
1039 _ -> panic "StgInterp.lit2expr: unhandled string constant type"
1041 -- Get the addr on the stack, untaggedly
1042 (unitOL (PUSH_UBX (Right mallocvilleAddr) 1), 1)
1048 pushAtom tagged d p (AnnApp f (_, AnnType _))
1049 = pushAtom tagged d p (snd f)
1051 pushAtom tagged d p (AnnNote note e)
1052 = pushAtom tagged d p (snd e)
1054 pushAtom tagged d p (AnnLam x e)
1056 = pushAtom tagged d p (snd e)
1058 pushAtom tagged d p other
1059 = pprPanic "ByteCodeGen.pushAtom"
1060 (pprCoreExpr (deAnnotate (undefined, other)))
1062 foreign import "strncpy" strncpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1065 -- Given a bunch of alts code and their discrs, do the donkey work
1066 -- of making a multiway branch using a switch tree.
1067 -- What a load of hassle!
1068 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1069 -- a hint; generates better code
1070 -- Nothing is always safe
1071 -> [(Discr, BCInstrList)]
1073 mkMultiBranch maybe_ncons raw_ways
1074 = let d_way = filter (isNoDiscr.fst) raw_ways
1075 notd_ways = naturalMergeSortLe
1076 (\w1 w2 -> leAlt (fst w1) (fst w2))
1077 (filter (not.isNoDiscr.fst) raw_ways)
1079 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1080 mkTree [] range_lo range_hi = returnBc the_default
1082 mkTree [val] range_lo range_hi
1083 | range_lo `eqAlt` range_hi
1084 = returnBc (snd val)
1086 = getLabelBc `thenBc` \ label_neq ->
1087 returnBc (mkTestEQ (fst val) label_neq
1089 `appOL` unitOL (LABEL label_neq)
1090 `appOL` the_default))
1092 mkTree vals range_lo range_hi
1093 = let n = length vals `div` 2
1094 vals_lo = take n vals
1095 vals_hi = drop n vals
1096 v_mid = fst (head vals_hi)
1098 getLabelBc `thenBc` \ label_geq ->
1099 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1100 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1101 returnBc (mkTestLT v_mid label_geq
1103 `appOL` unitOL (LABEL label_geq)
1107 = case d_way of [] -> unitOL CASEFAIL
1110 -- None of these will be needed if there are no non-default alts
1111 (mkTestLT, mkTestEQ, init_lo, init_hi)
1113 = panic "mkMultiBranch: awesome foursome"
1115 = case fst (head notd_ways) of {
1116 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1117 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1120 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1121 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1124 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1125 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1128 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1129 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1131 DiscrP algMaxBound )
1134 (algMinBound, algMaxBound)
1135 = case maybe_ncons of
1136 Just n -> (0, n - 1)
1137 Nothing -> (minBound, maxBound)
1139 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1140 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1141 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1142 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1143 NoDiscr `eqAlt` NoDiscr = True
1146 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1147 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1148 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1149 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1150 NoDiscr `leAlt` NoDiscr = True
1153 isNoDiscr NoDiscr = True
1156 dec (DiscrI i) = DiscrI (i-1)
1157 dec (DiscrP i) = DiscrP (i-1)
1158 dec other = other -- not really right, but if you
1159 -- do cases on floating values, you'll get what you deserve
1161 -- same snotty comment applies to the following
1163 minD, maxD :: Double
1169 mkTree notd_ways init_lo init_hi
1173 %************************************************************************
1175 \subsection{Supporting junk for the compilation schemes}
1177 %************************************************************************
1181 -- Describes case alts
1189 instance Outputable Discr where
1190 ppr (DiscrI i) = int i
1191 ppr (DiscrF f) = text (show f)
1192 ppr (DiscrD d) = text (show d)
1193 ppr (DiscrP i) = int i
1194 ppr NoDiscr = text "DEF"
1197 -- Find things in the BCEnv (the what's-on-the-stack-env)
1198 -- See comment preceding pushAtom for precise meaning of env contents
1199 --lookupBCEnv :: BCEnv -> Id -> Int
1200 --lookupBCEnv env nm
1201 -- = case lookupFM env nm of
1202 -- Nothing -> pprPanic "lookupBCEnv"
1203 -- (ppr nm $$ char ' ' $$ vcat (map ppr (fmToList env)))
1206 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1207 lookupBCEnv_maybe = lookupFM
1210 taggedIdSizeW, untaggedIdSizeW :: Id -> Int
1211 taggedIdSizeW = taggedSizeW . typePrimRep . idType
1212 untaggedIdSizeW = untaggedSizeW . typePrimRep . idType
1214 unboxedTupleException :: a
1215 unboxedTupleException
1218 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1219 "\tto foreign import/export decls in source. Workaround:\n" ++
1220 "\tcompile this module to a .o file, then restart session."))
1223 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1228 %************************************************************************
1230 \subsection{The bytecode generator's monad}
1232 %************************************************************************
1236 = BcM_State { bcos :: [ProtoBCO Name], -- accumulates completed BCOs
1237 nextlabel :: Int } -- for generating local labels
1239 type BcM r = BcM_State -> IO (BcM_State, r)
1241 ioToBc :: IO a -> BcM a
1242 ioToBc io st = do x <- io
1245 runBc :: BcM_State -> BcM r -> IO (BcM_State, r)
1246 runBc st0 m = do (st1, res) <- m st0
1249 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1250 thenBc expr cont st0
1251 = do (st1, q) <- expr st0
1252 (st2, r) <- cont q st1
1255 thenBc_ :: BcM a -> BcM b -> BcM b
1256 thenBc_ expr cont st0
1257 = do (st1, q) <- expr st0
1258 (st2, r) <- cont st1
1261 returnBc :: a -> BcM a
1262 returnBc result st = return (st, result)
1265 mapBc :: (a -> BcM b) -> [a] -> BcM [b]
1266 mapBc f [] = returnBc []
1268 = f x `thenBc` \ r ->
1269 mapBc f xs `thenBc` \ rs ->
1272 emitBc :: ProtoBCO Name -> BcM ()
1274 = return (st{bcos = bco : bcos st}, ())
1276 getLabelBc :: BcM Int
1278 = return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1280 getLabelsBc :: Int -> BcM [Int]
1282 = let ctr = nextlabel st
1283 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])