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]
89 final_state = runBc (BcM_State [] 0)
90 (mapBc (schemeR True) flatBinds
91 `thenBc_` returnBc ())
92 (BcM_State proto_bcos final_ctr) = final_state
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 let (BcM_State all_proto_bcos final_ctr)
117 = runBc (BcM_State [] 0)
118 (schemeR True (invented_id, freeVars expr))
119 dumpIfSet_dyn dflags Opt_D_dump_BCOs
120 "Proto-bcos" (vcat (intersperse (char ' ') (map ppr all_proto_bcos)))
123 = case filter ((== invented_name).nameOfProtoBCO) all_proto_bcos of
124 [root_bco] -> root_bco
126 = filter ((/= invented_name).nameOfProtoBCO) all_proto_bcos
128 auxiliary_bcos <- mapM assembleBCO auxiliary_proto_bcos
129 root_bco <- assembleBCO root_proto_bco
131 return (root_bco, auxiliary_bcos)
134 %************************************************************************
136 \subsection{Compilation schema for the bytecode generator.}
138 %************************************************************************
142 type BCInstrList = OrdList BCInstr
144 type Sequel = Int -- back off to this depth before ENTER
146 -- Maps Ids to the offset from the stack _base_ so we don't have
147 -- to mess with it after each push/pop.
148 type BCEnv = FiniteMap Id Int -- To find vars on the stack
150 ppBCEnv :: BCEnv -> SDoc
153 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
156 pp_one (var, offset) = int offset <> colon <+> ppr var
157 cmp_snd x y = compare (snd x) (snd y)
159 -- Create a BCO and do a spot of peephole optimisation on the insns
161 mkProtoBCO nm instrs_ordlist origin
162 = ProtoBCO nm maybe_with_stack_check origin
164 -- Overestimate the stack usage (in words) of this BCO,
165 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
166 -- stack check. (The interpreter always does a stack check
167 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
168 -- BCO anyway, so we only need to add an explicit on in the
169 -- (hopefully rare) cases when the (overestimated) stack use
170 -- exceeds iNTERP_STACK_CHECK_THRESH.
171 maybe_with_stack_check
172 | stack_overest >= 65535
173 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
175 | stack_overest >= iNTERP_STACK_CHECK_THRESH
176 = (STKCHECK stack_overest) : peep_d
178 = peep_d -- the supposedly common case
180 stack_overest = sum (map bciStackUse peep_d)
181 + 10 {- just to be really really sure -}
184 -- Merge local pushes
185 peep_d = peep (fromOL instrs_ordlist)
187 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
188 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
189 peep (PUSH_L off1 : PUSH_L off2 : rest)
190 = PUSH_LL off1 (off2-1) : peep rest
197 -- Compile code for the right hand side of a let binding.
198 -- Park the resulting BCO in the monad. Also requires the
199 -- variable to which this value was bound, so as to give the
200 -- resulting BCO a name. Bool indicates top-levelness.
202 schemeR :: Bool -> (Id, AnnExpr Id VarSet) -> BcM ()
203 schemeR is_top (nm, rhs)
207 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
208 $$ pprCoreExpr (deAnnotate rhs)
214 = schemeR_wrk is_top rhs nm (collect [] rhs)
217 collect xs (_, AnnNote note e)
219 collect xs (_, AnnLam x e)
220 = collect (if isTyVar x then xs else (x:xs)) e
221 collect xs not_lambda
222 = (reverse xs, not_lambda)
224 schemeR_wrk is_top original_body nm (args, body)
225 | Just dcon <- maybe_toplevel_null_con_rhs
226 = --trace ("nullary constructor! " ++ showSDocDebug (ppr nm)) (
227 emitBc (mkProtoBCO (getName nm) (toOL [PACK dcon 0, ENTER])
228 (Right original_body))
232 = let fvs = filter (not.isTyVar) (varSetElems (fst original_body))
233 all_args = reverse args ++ fvs
234 szsw_args = map taggedIdSizeW all_args
235 szw_args = sum szsw_args
236 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
237 argcheck = unitOL (ARGCHECK szw_args)
239 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
240 emitBc (mkProtoBCO (getName nm) (appOL argcheck body_code)
241 (Right original_body))
244 maybe_toplevel_null_con_rhs
245 | is_top && null args
248 -> case isDataConId_maybe v_wrk of
250 Just dc_wrk | nm == dataConWrapId dc_wrk
258 -- Let szsw be the sizes in words of some items pushed onto the stack,
259 -- which has initial depth d'. Return the values which the stack environment
260 -- should map these items to.
261 mkStackOffsets :: Int -> [Int] -> [Int]
262 mkStackOffsets original_depth szsw
263 = map (subtract 1) (tail (scanl (+) original_depth szsw))
265 -- Compile code to apply the given expression to the remaining args
266 -- on the stack, returning a HNF.
267 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr Id VarSet -> BcM BCInstrList
269 -- Delegate tail-calls to schemeT.
270 schemeE d s p e@(fvs, AnnApp f a)
271 = schemeT d s p (fvs, AnnApp f a)
273 schemeE d s p e@(fvs, AnnVar v)
274 | isFollowableRep v_rep
275 = -- Ptr-ish thing; push it in the normal way
276 schemeT d s p (fvs, AnnVar v)
279 = -- returning an unboxed value. Heave it on the stack, SLIDE, and RETURN.
280 let (push, szw) = pushAtom True d p (AnnVar v)
281 in returnBc (push -- value onto stack
282 `appOL` mkSLIDE szw (d-s) -- clear to sequel
283 `snocOL` RETURN v_rep) -- go
285 v_rep = typePrimRep (idType v)
287 schemeE d s p (fvs, AnnLit literal)
288 = let (push, szw) = pushAtom True d p (AnnLit literal)
289 l_rep = literalPrimRep literal
290 in returnBc (push -- value onto stack
291 `appOL` mkSLIDE szw (d-s) -- clear to sequel
292 `snocOL` RETURN l_rep) -- go
294 schemeE d s p (fvs, AnnLet binds b)
295 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
296 AnnRec xs_n_rhss -> unzip xs_n_rhss
298 fvss = map (filter (not.isTyVar).varSetElems.fst) rhss
300 -- Sizes of tagged free vars, + 1 for the fn
301 sizes = map (\rhs_fvs -> 1 + sum (map taggedIdSizeW rhs_fvs)) fvss
303 -- This p', d' defn is safe because all the items being pushed
304 -- are ptrs, so all have size 1. d' and p' reflect the stack
305 -- after the closures have been allocated in the heap (but not
306 -- filled in), and pointers to them parked on the stack.
307 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n 1)))
310 infos = zipE4 fvss sizes xs [n, n-1 .. 1]
311 zipE = zipEqual "schemeE"
312 zipE4 = zipWith4Equal "schemeE" (\a b c d -> (a,b,c,d))
314 -- ToDo: don't build thunks for things with no free variables
315 buildThunk dd ([], size, id, off)
316 = PUSH_G (Left (getName id))
317 `consOL` unitOL (MKAP (off+size-1) size)
318 buildThunk dd ((fv:fvs), size, id, off)
319 = case pushAtom True dd p' (AnnVar fv) of
320 (push_code, pushed_szw)
322 buildThunk (dd+pushed_szw) (fvs, size, id, off)
324 thunkCode = concatOL (map (buildThunk d') infos)
325 allocCode = toOL (map ALLOC sizes)
327 schemeE d' s p' b `thenBc` \ bodyCode ->
328 mapBc (schemeR False) (zip xs rhss) `thenBc_`
329 returnBc (allocCode `appOL` thunkCode `appOL` bodyCode)
335 schemeE d s p (fvs_case, AnnCase (fvs_scrut, scrut) bndr
336 [(DEFAULT, [], (fvs_rhs, rhs))])
338 | let isFunType var_type
339 = case splitTyConApp_maybe var_type of
340 Just (tycon,_) | isFunTyCon tycon -> True
342 ty_bndr = repType (idType bndr)
343 in isFunType ty_bndr || isTyVarTy ty_bndr
346 -- case scrut::suspect of bndr { DEFAULT -> rhs }
348 -- let bndr = scrut in rhs
349 -- when suspect is polymorphic or arrowtyped
350 -- So the required strictness properties are not observed.
351 -- At some point, must fix this properly.
355 (AnnNonRec bndr (fvs_scrut, scrut)) (fvs_rhs, rhs)
358 in trace ("WARNING: ignoring polymorphic case in interpreted mode.\n" ++
359 " Possibly due to strict polymorphic/functional constructor args.\n" ++
360 " Your program may leak space unexpectedly.\n")
361 -- ++ showSDoc (char ' ' $$ pprCoreExpr (deAnnotate new_expr) $$ char ' '))
362 (schemeE d s p new_expr)
366 {- Convert case .... of (# VoidRep'd-thing, a #) -> ...
368 case .... of a -> ...
369 Use a as the name of the binder too.
371 Also case .... of (# a #) -> ...
373 case .... of a -> ...
375 schemeE d s p (fvs, AnnCase scrut bndr [(DataAlt dc, [bind1, bind2], rhs)])
376 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind1)
377 = --trace "automagic mashing of case alts (# VoidRep, a #)" (
378 schemeE d s p (fvs, AnnCase scrut bind2 [(DEFAULT, [bind2], rhs)])
381 schemeE d s p (fvs, AnnCase scrut bndr [(DataAlt dc, [bind1], rhs)])
382 | isUnboxedTupleCon dc
383 = --trace "automagic mashing of case alts (# a #)" (
384 schemeE d s p (fvs, AnnCase scrut bind1 [(DEFAULT, [bind1], rhs)])
387 schemeE d s p (fvs, AnnCase scrut bndr alts)
389 -- Top of stack is the return itbl, as usual.
390 -- underneath it is the pointer to the alt_code BCO.
391 -- When an alt is entered, it assumes the returned value is
392 -- on top of the itbl.
395 -- Env and depth in which to compile the alts, not including
396 -- any vars bound by the alts themselves
397 d' = d + ret_frame_sizeW + taggedIdSizeW bndr
398 p' = addToFM p bndr (d' - 1)
400 scrut_primrep = typePrimRep (idType bndr)
402 | scrut_primrep == PtrRep
404 | scrut_primrep `elem`
405 [CharRep, AddrRep, WordRep, IntRep, FloatRep, DoubleRep,
406 VoidRep, Int8Rep, Int16Rep, Int32Rep, Int64Rep,
407 Word8Rep, Word16Rep, Word32Rep, Word64Rep]
410 = pprPanic "ByteCodeGen.schemeE" (ppr scrut_primrep)
412 -- given an alt, return a discr and code for it.
413 codeAlt alt@(discr, binds_f, rhs)
415 = let (unpack_code, d_after_unpack, p_after_unpack)
416 = mkUnpackCode (filter (not.isTyVar) binds_f) d' p'
417 in schemeE d_after_unpack s p_after_unpack rhs
418 `thenBc` \ rhs_code ->
419 returnBc (my_discr alt, unpack_code `appOL` rhs_code)
421 = ASSERT(null binds_f)
422 schemeE d' s p' rhs `thenBc` \ rhs_code ->
423 returnBc (my_discr alt, rhs_code)
425 my_discr (DEFAULT, binds, rhs) = NoDiscr
426 my_discr (DataAlt dc, binds, rhs)
427 | isUnboxedTupleCon dc
428 = unboxedTupleException
430 = DiscrP (dataConTag dc - fIRST_TAG)
431 my_discr (LitAlt l, binds, rhs)
432 = case l of MachInt i -> DiscrI (fromInteger i)
433 MachFloat r -> DiscrF (fromRational r)
434 MachDouble r -> DiscrD (fromRational r)
435 MachChar i -> DiscrI i
436 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
439 | not isAlgCase = Nothing
441 = case [dc | (DataAlt dc, _, _) <- alts] of
443 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
446 mapBc codeAlt alts `thenBc` \ alt_stuff ->
447 mkMultiBranch maybe_ncons alt_stuff `thenBc` \ alt_final ->
449 alt_final_ac = ARGCHECK (taggedIdSizeW bndr) `consOL` alt_final
450 alt_bco_name = getName bndr
451 alt_bco = mkProtoBCO alt_bco_name alt_final_ac (Left alts)
453 schemeE (d + ret_frame_sizeW)
454 (d + ret_frame_sizeW) p scrut `thenBc` \ scrut_code ->
456 emitBc alt_bco `thenBc_`
457 returnBc (PUSH_AS alt_bco_name scrut_primrep `consOL` scrut_code)
460 schemeE d s p (fvs, AnnNote note body)
464 = pprPanic "ByteCodeGen.schemeE: unhandled case"
465 (pprCoreExpr (deAnnotate other))
468 -- Compile code to do a tail call. Specifically, push the fn,
469 -- slide the on-stack app back down to the sequel depth,
470 -- and enter. Four cases:
473 -- An application "PrelGHC.tagToEnum# <type> unboxed-int".
474 -- The int will be on the stack. Generate a code sequence
475 -- to convert it to the relevant constructor, SLIDE and ENTER.
477 -- 1. A nullary constructor. Push its closure on the stack
478 -- and SLIDE and RETURN.
480 -- 2. (Another nasty hack). Spot (# a::VoidRep, b #) and treat
481 -- it simply as b -- since the representations are identical
482 -- (the VoidRep takes up zero stack space). Also, spot
483 -- (# b #) and treat it as b.
485 -- 3. Application of a non-nullary constructor, by defn saturated.
486 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
487 -- then the ptrs, and then do PACK and RETURN.
489 -- 4. Otherwise, it must be a function call. Push the args
490 -- right to left, SLIDE and ENTER.
492 schemeT :: Int -- Stack depth
493 -> Sequel -- Sequel depth
494 -> BCEnv -- stack env
500 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
501 -- = panic "schemeT ?!?!"
503 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate app)) ++ "\n") False
507 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
508 = pushAtom True d p arg `bind` \ (push, arg_words) ->
509 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
510 returnBc (push `appOL` tagToId_sequence
511 `appOL` mkSLIDE 1 (d+arg_words-s)
515 | is_con_call && null args_r_to_l
517 (PUSH_G (Left (getName con)) `consOL` mkSLIDE 1 (d-s))
522 | let isVoidRepAtom (_, AnnVar v) = VoidRep == typePrimRep (idType v)
523 isVoidRepAtom (_, AnnNote n e) = isVoidRepAtom e
524 in is_con_call && isUnboxedTupleCon con
525 && ( (length args_r_to_l == 2 && isVoidRepAtom (last (args_r_to_l)))
526 || (length args_r_to_l == 1)
528 = --trace (if length args_r_to_l == 1
529 -- then "schemeT: unboxed singleton"
530 -- else "schemeT: unboxed pair with Void first component") (
531 schemeT d s p (head args_r_to_l)
534 | Just (CCall ccall_spec) <- isFCallId_maybe fn
535 = generateCCall d s p ccall_spec fn args_r_to_l
539 = if is_con_call && isUnboxedTupleCon con
540 then returnBc unboxedTupleException
541 else code `seq` returnBc code
544 -- Detect and extract relevant info for the tagToEnum kludge.
545 maybe_is_tagToEnum_call
546 = let extract_constr_Names ty
547 = case splitTyConApp_maybe (repType ty) of
548 (Just (tyc, [])) | isDataTyCon tyc
549 -> map getName (tyConDataCons tyc)
550 other -> panic "maybe_is_tagToEnum_call.extract_constr_Ids"
553 (_, AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
554 -> case isPrimOpId_maybe v of
555 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
559 -- Extract the args (R->L) and fn
560 (args_r_to_l_raw, fn) = chomp app
564 AnnApp f a -> case chomp f of (az, f) -> (a:az, f)
565 AnnNote n e -> chomp e
566 other -> pprPanic "schemeT"
567 (ppr (deAnnotate (panic "schemeT.chomp", other)))
569 args_r_to_l = filter (not.isTypeAtom.snd) args_r_to_l_raw
570 isTypeAtom (AnnType _) = True
573 -- decide if this is a constructor call, and rearrange
574 -- args appropriately.
575 maybe_dcon = isDataConId_maybe fn
576 is_con_call = case maybe_dcon of Nothing -> False; Just _ -> True
577 (Just con) = maybe_dcon
583 = filter (not.isPtr.snd) args_r_to_l ++ filter (isPtr.snd) args_r_to_l
584 where isPtr = isFollowableRep . atomRep
586 -- make code to push the args and then do the SLIDE-ENTER thing
587 code = do_pushery d (map snd args_final_r_to_l)
588 tag_when_push = not is_con_call
589 narg_words = sum (map (get_arg_szw . atomRep . snd) args_r_to_l)
590 get_arg_szw = if tag_when_push then taggedSizeW else untaggedSizeW
592 do_pushery d (arg:args)
593 = let (push, arg_words) = pushAtom tag_when_push d p arg
594 in push `appOL` do_pushery (d+arg_words) args
596 | Just (CCall ccall_spec) <- isFCallId_maybe fn
597 = panic "schemeT.do_pushery: unexpected ccall"
601 Just con -> PACK con narg_words `consOL` (
602 mkSLIDE 1 (d - narg_words - s) `snocOL` ENTER)
604 -> let (push, arg_words) = pushAtom True d p (AnnVar fn)
606 `appOL` mkSLIDE (narg_words+arg_words)
612 {- Deal with a CCall. Taggedly push the args onto the stack R->L,
613 deferencing ForeignObj#s and (ToDo: adjusting addrs to point to
614 payloads in Ptr/Byte arrays). Then, generate the marshalling
615 (machine) code for the ccall, and create bytecodes to call that and
616 then return in the right way.
618 generateCCall :: Int -> Sequel -- stack and sequel depths
620 -> CCallSpec -- where to call
621 -> Id -- of target, for type info
622 -> [AnnExpr Id VarSet] -- args (atoms)
625 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
628 addr_usizeW = untaggedSizeW AddrRep
629 addr_tsizeW = taggedSizeW AddrRep
631 -- Get the args on the stack, with tags and suitably
632 -- dereferenced for the CCall. For each arg, return the
633 -- depth to the first word of the bits for that arg, and the
634 -- PrimRep of what was actually pushed.
638 = let rep_arg = atomRep a
640 -- Don't push the FO; instead push the Addr# it
643 -> let foro_szW = taggedSizeW ForeignObjRep
644 push_fo = fst (pushAtom False{-irrelevant-} d p a)
645 d_now = d + addr_tsizeW
646 code = push_fo `appOL` toOL [
647 UPK_TAG addr_usizeW 0 0,
648 SLIDE addr_tsizeW foro_szW
650 in (code, AddrRep) : f d_now az
651 -- Default case: push taggedly, but otherwise intact.
653 -> let (code_a, sz_a) = pushAtom True d p a
654 in (code_a, rep_arg) : f (d+sz_a) az
656 (pushs_arg, a_reps_pushed_r_to_l) = unzip (f d0 args_r_to_l)
658 push_args = concatOL pushs_arg
659 d_after_args = d0 + sum (map taggedSizeW a_reps_pushed_r_to_l)
661 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
662 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
664 = reverse (tail a_reps_pushed_r_to_l)
666 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
667 -- push_args is the code to do that.
668 -- d_after_args is the stack depth once the args are on.
670 -- Get the result rep.
671 (returns_void, r_rep)
672 = case maybe_getCCallReturnRep (idType fn) of
673 Nothing -> (True, VoidRep)
674 Just rr -> (False, rr)
676 Because the Haskell stack grows down, the a_reps refer to
677 lowest to highest addresses in that order. The args for the call
678 are on the stack. Now push an unboxed, tagged Addr# indicating
679 the C function to call. Then push a dummy placeholder for the
680 result. Finally, emit a CCALL insn with an offset pointing to the
681 Addr# just pushed, and a literal field holding the mallocville
682 address of the piece of marshalling code we generate.
683 So, just prior to the CCALL insn, the stack looks like this
684 (growing down, as usual):
689 Addr# address_of_C_fn
690 <placeholder-for-result#> (must be an unboxed type)
692 The interpreter then calls the marshall code mentioned
693 in the CCALL insn, passing it (& <placeholder-for-result#>),
694 that is, the addr of the topmost word in the stack.
695 When this returns, the placeholder will have been
696 filled in. The placeholder is slid down to the sequel
697 depth, and we RETURN.
699 This arrangement makes it simple to do f-i-dynamic since the Addr#
700 value is the first arg anyway. It also has the virtue that the
701 stack is GC-understandable at all times.
703 The marshalling code is generated specifically for this
704 call site, and so knows exactly the (Haskell) stack
705 offsets of the args, fn address and placeholder. It
706 copies the args to the C stack, calls the stacked addr,
707 and parks the result back in the placeholder. The interpreter
708 calls it as a normal C call, assuming it has a signature
709 void marshall_code ( StgWord* ptr_to_top_of_stack )
711 -- resolve static address
712 (is_static, static_target_addr)
715 -> (False, panic "ByteCodeGen.generateCCall(dyn)")
717 -> let unpacked = _UNPK_ target
718 in case unsafePerformIO (lookupSymbol unpacked) of
719 Just aa -> case aa of Ptr a# -> (True, A# a#)
724 invalid = pprPanic ("ByteCodeGen.generateCCall: unfindable "
725 ++ "symbol or otherwise invalid target")
728 -- Get the arg reps, zapping the leading Addr# in the dynamic case
729 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
730 | is_static = a_reps_pushed_RAW
731 | otherwise = if null a_reps_pushed_RAW
732 then panic "ByteCodeGen.generateCCall: dyn with no args"
733 else tail a_reps_pushed_RAW
736 (push_Addr, d_after_Addr)
738 = (toOL [PUSH_UBX (Right static_target_addr) addr_usizeW,
739 PUSH_TAG addr_usizeW],
740 d_after_args + addr_tsizeW)
741 | otherwise -- is already on the stack
742 = (nilOL, d_after_args)
744 -- Push the return placeholder. For a call returning nothing,
745 -- this is a VoidRep (tag).
746 r_usizeW = untaggedSizeW r_rep
747 r_tsizeW = taggedSizeW r_rep
748 d_after_r = d_after_Addr + r_tsizeW
749 r_lit = mkDummyLiteral r_rep
750 push_r = (if returns_void
752 else unitOL (PUSH_UBX (Left r_lit) r_usizeW))
754 unitOL (PUSH_TAG r_usizeW)
757 do_call = unitOL (CCALL addr_of_marshaller)
759 wrapup = mkSLIDE r_tsizeW (d_after_r - r_tsizeW - s)
760 `snocOL` RETURN r_rep
762 -- generate the marshalling code we're going to call
765 arg1_offW = r_tsizeW + addr_tsizeW
766 args_offW = map (arg1_offW +)
767 (init (scanl (+) 0 (map taggedSizeW a_reps)))
769 = mkMarshalCode cconv
770 (r_offW, r_rep) addr_offW
771 (zip args_offW a_reps)
773 --trace (show (arg1_offW, args_offW , (map taggedSizeW a_reps) )) (
776 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
781 -- Make a dummy literal, to be used as a placeholder for FFI return
782 -- values on the stack.
783 mkDummyLiteral :: PrimRep -> Literal
787 DoubleRep -> MachDouble 0
788 FloatRep -> MachFloat 0
789 AddrRep | taggedSizeW AddrRep == taggedSizeW WordRep -> MachWord 0
790 _ -> pprPanic "mkDummyLiteral" (ppr pr)
794 -- PrelGHC.Int# -> PrelGHC.State# PrelGHC.RealWorld
795 -- -> (# PrelGHC.State# PrelGHC.RealWorld, PrelGHC.Int# #)
798 -- and check that an unboxed pair isreturned wherein the first arg is VoidRep'd.
800 -- Alternatively, for call-targets returning nothing, convert
802 -- PrelGHC.Int# -> PrelGHC.State# PrelGHC.RealWorld
803 -- -> (# PrelGHC.State# PrelGHC.RealWorld, PrelGHC.Int# #)
807 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
808 maybe_getCCallReturnRep fn_ty
809 = let (a_tys, r_ty) = splitRepFunTys fn_ty
811 = if length r_reps == 1 then Nothing else Just (r_reps !! 1)
813 = case splitTyConApp_maybe (repType r_ty) of
814 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
816 ok = ( (length r_reps == 2 && VoidRep == head r_reps)
817 || r_reps == [VoidRep] )
818 && isUnboxedTupleTyCon r_tycon
819 && case maybe_r_rep_to_go of
821 Just r_rep -> r_rep /= PtrRep
822 -- if it was, it would be impossible
823 -- to create a valid return value
824 -- placeholder on the stack
825 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
828 --trace (showSDoc (ppr (a_reps, r_reps))) (
829 if ok then maybe_r_rep_to_go else blargh
832 atomRep (AnnVar v) = typePrimRep (idType v)
833 atomRep (AnnLit l) = literalPrimRep l
834 atomRep (AnnNote n b) = atomRep (snd b)
835 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
836 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
837 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
840 -- Compile code which expects an unboxed Int on the top of stack,
841 -- (call it i), and pushes the i'th closure in the supplied list
843 implement_tagToId :: [Name] -> BcM BCInstrList
844 implement_tagToId names
845 = ASSERT(not (null names))
846 getLabelsBc (length names) `thenBc` \ labels ->
847 getLabelBc `thenBc` \ label_fail ->
848 getLabelBc `thenBc` \ label_exit ->
849 zip4 labels (tail labels ++ [label_fail])
850 [0 ..] names `bind` \ infos ->
851 map (mkStep label_exit) infos `bind` \ steps ->
852 returnBc (concatOL steps
854 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
856 mkStep l_exit (my_label, next_label, n, name_for_n)
857 = toOL [LABEL my_label,
858 TESTEQ_I n next_label,
859 PUSH_G (Left name_for_n),
863 -- Make code to unpack the top-of-stack constructor onto the stack,
864 -- adding tags for the unboxed bits. Takes the PrimReps of the
865 -- constructor's arguments. off_h and off_s are travelling offsets
866 -- along the constructor and the stack.
868 -- Supposing a constructor in the heap has layout
870 -- Itbl p_1 ... p_i np_1 ... np_j
872 -- then we add to the stack, shown growing down, the following:
884 -- so that in the common case (ptrs only) a single UNPACK instr can
885 -- copy all the payload of the constr onto the stack with no further ado.
887 mkUnpackCode :: [Id] -- constr args
888 -> Int -- depth before unpack
889 -> BCEnv -- env before unpack
890 -> (BCInstrList, Int, BCEnv)
891 mkUnpackCode vars d p
892 = --trace ("mkUnpackCode: " ++ showSDocDebug (ppr vars)
893 -- ++ " --> " ++ show d' ++ "\n" ++ showSDocDebug (ppBCEnv p')
895 (code_p `appOL` code_np, d', p')
899 vreps = [(var, typePrimRep (idType var)) | var <- vars]
901 -- ptrs and nonptrs, forward
902 vreps_p = filter (isFollowableRep.snd) vreps
903 vreps_np = filter (not.isFollowableRep.snd) vreps
905 -- the order in which we will augment the environment
906 vreps_env = reverse vreps_p ++ reverse vreps_np
909 vreps_env_tszsw = map (taggedSizeW.snd) vreps_env
910 p' = addListToFM p (zip (map fst vreps_env)
911 (mkStackOffsets d vreps_env_tszsw))
912 d' = d + sum vreps_env_tszsw
914 -- code to unpack the ptrs
915 ptrs_szw = sum (map (untaggedSizeW.snd) vreps_p)
916 code_p | null vreps_p = nilOL
917 | otherwise = unitOL (UNPACK ptrs_szw)
919 -- code to unpack the nonptrs
920 vreps_env_uszw = sum (map (untaggedSizeW.snd) vreps_env)
921 code_np = do_nptrs vreps_env_uszw ptrs_szw (reverse (map snd vreps_np))
922 do_nptrs off_h off_s [] = nilOL
923 do_nptrs off_h off_s (npr:nprs)
924 | npr `elem` [IntRep, WordRep, FloatRep, DoubleRep, CharRep, AddrRep]
927 = pprPanic "ByteCodeGen.mkUnpackCode" (ppr npr)
929 approved = UPK_TAG usizeW (off_h-usizeW) off_s `consOL` theRest
930 theRest = do_nptrs (off_h-usizeW) (off_s + tsizeW) nprs
931 usizeW = untaggedSizeW npr
932 tsizeW = taggedSizeW npr
935 -- Push an atom onto the stack, returning suitable code & number of
936 -- stack words used. Pushes it either tagged or untagged, since
937 -- pushAtom is used to set up the stack prior to copying into the
938 -- heap for both APs (requiring tags) and constructors (which don't).
940 -- NB this means NO GC between pushing atoms for a constructor and
941 -- copying them into the heap. It probably also means that
942 -- tail calls MUST be of the form atom{atom ... atom} since if the
943 -- expression head was allowed to be arbitrary, there could be GC
944 -- in between pushing the arg atoms and completing the head.
945 -- (not sure; perhaps the allocate/doYouWantToGC interface means this
946 -- isn't a problem; but only if arbitrary graph construction for the
947 -- head doesn't leave this BCO, since GC might happen at the start of
948 -- each BCO (we consult doYouWantToGC there).
950 -- Blargh. JRS 001206
952 -- NB (further) that the env p must map each variable to the highest-
953 -- numbered stack slot for it. For example, if the stack has depth 4
954 -- and we tagged-ly push (v :: Int#) on it, the value will be in stack[4],
955 -- the tag in stack[5], the stack will have depth 6, and p must map v to
956 -- 5 and not to 4. Stack locations are numbered from zero, so a depth
957 -- 6 stack has valid words 0 .. 5.
959 pushAtom :: Bool -> Int -> BCEnv -> AnnExpr' Id VarSet -> (BCInstrList, Int)
960 pushAtom tagged d p (AnnVar v)
962 | idPrimRep v == VoidRep
963 = if tagged then (unitOL (PUSH_TAG 0), 1)
964 else panic "ByteCodeGen.pushAtom(VoidRep,untaggedly)"
967 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
969 | Just primop <- isPrimOpId_maybe v
970 = (unitOL (PUSH_G (Right primop)), 1)
974 str = "\npushAtom " ++ showSDocDebug (ppr v)
975 ++ " :: " ++ showSDocDebug (pprType (idType v))
976 ++ ", depth = " ++ show d
977 ++ ", tagged = " ++ show tagged ++ ", env =\n" ++
978 showSDocDebug (ppBCEnv p)
979 ++ " --> words: " ++ show (snd result) ++ "\n" ++
980 showSDoc (nest 4 (vcat (map ppr (fromOL (fst result)))))
981 ++ "\nendPushAtom " ++ showSDocDebug (ppr v)
985 = case lookupBCEnv_maybe p v of
986 Just d_v -> (toOL (nOfThem nwords (PUSH_L (d-d_v+sz_t-2))), nwords)
987 Nothing -> ASSERT(sz_t == 1) (unitOL (PUSH_G (Left nm)), nwords)
989 nm = case isDataConId_maybe v of
993 sz_t = taggedIdSizeW v
994 sz_u = untaggedIdSizeW v
995 nwords = if tagged then sz_t else sz_u
999 pushAtom True d p (AnnLit lit)
1000 = let (ubx_code, ubx_size) = pushAtom False d p (AnnLit lit)
1001 in (ubx_code `snocOL` PUSH_TAG ubx_size, 1 + ubx_size)
1003 pushAtom False d p (AnnLit lit)
1005 MachWord w -> code WordRep
1006 MachInt i -> code IntRep
1007 MachFloat r -> code FloatRep
1008 MachDouble r -> code DoubleRep
1009 MachChar c -> code CharRep
1010 MachStr s -> pushStr s
1013 = let size_host_words = untaggedSizeW rep
1014 in (unitOL (PUSH_UBX (Left lit) size_host_words), size_host_words)
1017 = let mallocvilleAddr
1021 FastString _ l ba ->
1022 -- sigh, a string in the heap is no good to us.
1023 -- We need a static C pointer, since the type of
1024 -- a string literal is Addr#. So, copy the string
1025 -- into C land and introduce a memory leak
1026 -- at the same time.
1028 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1029 in unsafePerformIO (
1030 do (Ptr a#) <- mallocBytes (n+1)
1031 strncpy (Ptr a#) ba (fromIntegral n)
1032 writeCharOffAddr (A# a#) n '\0'
1035 _ -> panic "StgInterp.lit2expr: unhandled string constant type"
1037 -- Get the addr on the stack, untaggedly
1038 (unitOL (PUSH_UBX (Right mallocvilleAddr) 1), 1)
1044 pushAtom tagged d p (AnnApp f (_, AnnType _))
1045 = pushAtom tagged d p (snd f)
1047 pushAtom tagged d p (AnnNote note e)
1048 = pushAtom tagged d p (snd e)
1050 pushAtom tagged d p (AnnLam x e)
1052 = pushAtom tagged d p (snd e)
1054 pushAtom tagged d p other
1055 = pprPanic "ByteCodeGen.pushAtom"
1056 (pprCoreExpr (deAnnotate (undefined, other)))
1058 foreign import "strncpy" strncpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1061 -- Given a bunch of alts code and their discrs, do the donkey work
1062 -- of making a multiway branch using a switch tree.
1063 -- What a load of hassle!
1064 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1065 -- a hint; generates better code
1066 -- Nothing is always safe
1067 -> [(Discr, BCInstrList)]
1069 mkMultiBranch maybe_ncons raw_ways
1070 = let d_way = filter (isNoDiscr.fst) raw_ways
1071 notd_ways = naturalMergeSortLe
1072 (\w1 w2 -> leAlt (fst w1) (fst w2))
1073 (filter (not.isNoDiscr.fst) raw_ways)
1075 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1076 mkTree [] range_lo range_hi = returnBc the_default
1078 mkTree [val] range_lo range_hi
1079 | range_lo `eqAlt` range_hi
1080 = returnBc (snd val)
1082 = getLabelBc `thenBc` \ label_neq ->
1083 returnBc (mkTestEQ (fst val) label_neq
1085 `appOL` unitOL (LABEL label_neq)
1086 `appOL` the_default))
1088 mkTree vals range_lo range_hi
1089 = let n = length vals `div` 2
1090 vals_lo = take n vals
1091 vals_hi = drop n vals
1092 v_mid = fst (head vals_hi)
1094 getLabelBc `thenBc` \ label_geq ->
1095 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1096 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1097 returnBc (mkTestLT v_mid label_geq
1099 `appOL` unitOL (LABEL label_geq)
1103 = case d_way of [] -> unitOL CASEFAIL
1106 -- None of these will be needed if there are no non-default alts
1107 (mkTestLT, mkTestEQ, init_lo, init_hi)
1109 = panic "mkMultiBranch: awesome foursome"
1111 = case fst (head notd_ways) of {
1112 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1113 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1116 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1117 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1120 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1121 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1124 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1125 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1127 DiscrP algMaxBound )
1130 (algMinBound, algMaxBound)
1131 = case maybe_ncons of
1132 Just n -> (0, n - 1)
1133 Nothing -> (minBound, maxBound)
1135 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1136 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1137 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1138 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1139 NoDiscr `eqAlt` NoDiscr = True
1142 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1143 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1144 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1145 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1146 NoDiscr `leAlt` NoDiscr = True
1149 isNoDiscr NoDiscr = True
1152 dec (DiscrI i) = DiscrI (i-1)
1153 dec (DiscrP i) = DiscrP (i-1)
1154 dec other = other -- not really right, but if you
1155 -- do cases on floating values, you'll get what you deserve
1157 -- same snotty comment applies to the following
1159 minD, maxD :: Double
1165 mkTree notd_ways init_lo init_hi
1169 %************************************************************************
1171 \subsection{Supporting junk for the compilation schemes}
1173 %************************************************************************
1177 -- Describes case alts
1185 instance Outputable Discr where
1186 ppr (DiscrI i) = int i
1187 ppr (DiscrF f) = text (show f)
1188 ppr (DiscrD d) = text (show d)
1189 ppr (DiscrP i) = int i
1190 ppr NoDiscr = text "DEF"
1193 -- Find things in the BCEnv (the what's-on-the-stack-env)
1194 -- See comment preceding pushAtom for precise meaning of env contents
1195 --lookupBCEnv :: BCEnv -> Id -> Int
1196 --lookupBCEnv env nm
1197 -- = case lookupFM env nm of
1198 -- Nothing -> pprPanic "lookupBCEnv"
1199 -- (ppr nm $$ char ' ' $$ vcat (map ppr (fmToList env)))
1202 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1203 lookupBCEnv_maybe = lookupFM
1206 taggedIdSizeW, untaggedIdSizeW :: Id -> Int
1207 taggedIdSizeW = taggedSizeW . typePrimRep . idType
1208 untaggedIdSizeW = untaggedSizeW . typePrimRep . idType
1210 unboxedTupleException :: a
1211 unboxedTupleException
1214 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1215 "\tto foreign import/export decls in source. Workaround:\n" ++
1216 "\tcompile this module to a .o file, then restart session."))
1219 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1224 %************************************************************************
1226 \subsection{The bytecode generator's monad}
1228 %************************************************************************
1232 = BcM_State { bcos :: [ProtoBCO Name], -- accumulates completed BCOs
1233 nextlabel :: Int } -- for generating local labels
1235 type BcM result = BcM_State -> (result, BcM_State)
1237 runBc :: BcM_State -> BcM () -> BcM_State
1238 runBc init_st m = case m init_st of { (r,st) -> st }
1240 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1242 = case expr st of { (result, st') -> cont result st' }
1244 thenBc_ :: BcM a -> BcM b -> BcM b
1245 thenBc_ expr cont st
1246 = case expr st of { (result, st') -> cont st' }
1248 returnBc :: a -> BcM a
1249 returnBc result st = (result, st)
1251 mapBc :: (a -> BcM b) -> [a] -> BcM [b]
1252 mapBc f [] = returnBc []
1254 = f x `thenBc` \ r ->
1255 mapBc f xs `thenBc` \ rs ->
1258 emitBc :: ProtoBCO Name -> BcM ()
1260 = ((), st{bcos = bco : bcos st})
1262 getLabelBc :: BcM Int
1264 = (nextlabel st, st{nextlabel = 1 + nextlabel st})
1266 getLabelsBc :: Int -> BcM [Int]
1268 = let ctr = nextlabel st
1269 in ([ctr .. ctr+n-1], st{nextlabel = ctr+n})