2 % (c) The University of Glasgow 2002-2006
5 ByteCodeGen: Generate bytecode from Core
9 -- The above warning supression flag is a temporary kludge.
10 -- While working on this module you are encouraged to remove it and fix
11 -- any warnings in the module. See
12 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
15 module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where
17 #include "HsVersions.h"
57 import Data.List ( intersperse, sortBy, zip4, zip6, partition )
58 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8,
59 withForeignPtr, castFunPtrToPtr, nullPtr, plusPtr )
61 import Control.Exception ( throwDyn )
63 import GHC.Exts ( Int(..), ByteArray# )
65 import Control.Monad ( when )
66 import Data.Char ( ord, chr )
74 -- -----------------------------------------------------------------------------
75 -- Generating byte code for a complete module
77 byteCodeGen :: DynFlags
81 -> IO CompiledByteCode
82 byteCodeGen dflags binds tycs modBreaks
83 = do showPass dflags "ByteCodeGen"
85 let flatBinds = [ (bndr, freeVars rhs)
86 | (bndr, rhs) <- flattenBinds binds]
88 us <- mkSplitUniqSupply 'y'
89 (BcM_State _us final_ctr mallocd _, proto_bcos)
90 <- runBc us modBreaks (mapM schemeTopBind flatBinds)
92 when (notNull 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 assembleBCOs proto_bcos tycs
100 -- -----------------------------------------------------------------------------
101 -- Generating byte code for an expression
103 -- Returns: (the root BCO for this expression,
104 -- a list of auxilary BCOs resulting from compiling closures)
105 coreExprToBCOs :: DynFlags
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_name = mkSystemVarName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
114 invented_id = Id.mkLocalId invented_name (panic "invented_id's type")
116 -- the uniques are needed to generate fresh variables when we introduce new
117 -- let bindings for ticked expressions
118 us <- mkSplitUniqSupply 'y'
119 (BcM_State _us final_ctr mallocd _ , proto_bco)
120 <- runBc us emptyModBreaks (schemeTopBind (invented_id, freeVars expr))
122 when (notNull mallocd)
123 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
125 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
127 assembleBCO proto_bco
130 -- -----------------------------------------------------------------------------
131 -- Compilation schema for the bytecode generator
133 type BCInstrList = OrdList BCInstr
135 type Sequel = Int -- back off to this depth before ENTER
137 -- Maps Ids to the offset from the stack _base_ so we don't have
138 -- to mess with it after each push/pop.
139 type BCEnv = FiniteMap Id Int -- To find vars on the stack
141 ppBCEnv :: BCEnv -> SDoc
144 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
147 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
148 cmp_snd x y = compare (snd x) (snd y)
150 -- Create a BCO and do a spot of peephole optimisation on the insns
155 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
159 -> Bool -- True <=> is a return point, rather than a function
162 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap is_ret mallocd_blocks
165 protoBCOInstrs = maybe_with_stack_check,
166 protoBCOBitmap = bitmap,
167 protoBCOBitmapSize = bitmap_size,
168 protoBCOArity = arity,
169 protoBCOExpr = origin,
170 protoBCOPtrs = mallocd_blocks
173 -- Overestimate the stack usage (in words) of this BCO,
174 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
175 -- stack check. (The interpreter always does a stack check
176 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
177 -- BCO anyway, so we only need to add an explicit on in the
178 -- (hopefully rare) cases when the (overestimated) stack use
179 -- exceeds iNTERP_STACK_CHECK_THRESH.
180 maybe_with_stack_check
181 | is_ret && stack_usage < aP_STACK_SPLIM = peep_d
182 -- don't do stack checks at return points,
183 -- everything is aggregated up to the top BCO
184 -- (which must be a function).
185 -- That is, unless the stack usage is >= AP_STACK_SPLIM,
187 | stack_usage >= iNTERP_STACK_CHECK_THRESH
188 = STKCHECK stack_usage : peep_d
190 = peep_d -- the supposedly common case
192 -- We assume that this sum doesn't wrap
193 stack_usage = sum (map bciStackUse peep_d)
195 -- Merge local pushes
196 peep_d = peep (fromOL instrs_ordlist)
198 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
199 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
200 peep (PUSH_L off1 : PUSH_L off2 : rest)
201 = PUSH_LL off1 (off2-1) : peep rest
207 argBits :: [CgRep] -> [Bool]
210 | isFollowableArg rep = False : argBits args
211 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
213 -- -----------------------------------------------------------------------------
216 -- Compile code for the right-hand side of a top-level binding
218 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
221 schemeTopBind (id, rhs)
222 | Just data_con <- isDataConWorkId_maybe id,
223 isNullaryRepDataCon data_con = do
224 -- Special case for the worker of a nullary data con.
225 -- It'll look like this: Nil = /\a -> Nil a
226 -- If we feed it into schemeR, we'll get
228 -- because mkConAppCode treats nullary constructor applications
229 -- by just re-using the single top-level definition. So
230 -- for the worker itself, we must allocate it directly.
231 -- ioToBc (putStrLn $ "top level BCO")
232 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
233 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
236 = schemeR [{- No free variables -}] (id, rhs)
239 -- -----------------------------------------------------------------------------
242 -- Compile code for a right-hand side, to give a BCO that,
243 -- when executed with the free variables and arguments on top of the stack,
244 -- will return with a pointer to the result on top of the stack, after
245 -- removing the free variables and arguments.
247 -- Park the resulting BCO in the monad. Also requires the
248 -- variable to which this value was bound, so as to give the
249 -- resulting BCO a name.
251 schemeR :: [Id] -- Free vars of the RHS, ordered as they
252 -- will appear in the thunk. Empty for
253 -- top-level things, which have no free vars.
254 -> (Id, AnnExpr Id VarSet)
255 -> BcM (ProtoBCO Name)
256 schemeR fvs (nm, rhs)
260 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
261 $$ pprCoreExpr (deAnnotate rhs)
267 = schemeR_wrk fvs nm rhs (collect [] rhs)
269 collect :: [Var] -> AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
270 collect xs (_, AnnNote note e) = collect xs e
271 collect xs (_, AnnCast e _) = collect xs e
272 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
273 collect xs (_, not_lambda) = (reverse xs, not_lambda)
275 schemeR_wrk :: [Id] -> Id -> AnnExpr Id VarSet -> ([Var], AnnExpr' Var VarSet) -> BcM (ProtoBCO Name)
276 schemeR_wrk fvs nm original_body (args, body)
278 all_args = reverse args ++ fvs
279 arity = length all_args
280 -- all_args are the args in reverse order. We're compiling a function
281 -- \fv1..fvn x1..xn -> e
282 -- i.e. the fvs come first
284 szsw_args = map idSizeW all_args
285 szw_args = sum szsw_args
286 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
288 -- make the arg bitmap
289 bits = argBits (reverse (map idCgRep all_args))
290 bitmap_size = length bits
291 bitmap = mkBitmap bits
293 body_code <- schemeER_wrk szw_args p_init body
295 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
296 arity bitmap_size bitmap False{-not alts-})
298 -- introduce break instructions for ticked expressions
299 schemeER_wrk :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
301 | Just (tickInfo, (_annot, newRhs)) <- isTickedExp' rhs = do
302 code <- schemeE d 0 p newRhs
304 let idOffSets = getVarOffSets d p tickInfo
305 let tickNumber = tickInfo_number tickInfo
306 let breakInfo = BreakInfo
307 { breakInfo_module = tickInfo_module tickInfo
308 , breakInfo_number = tickNumber
309 , breakInfo_vars = idOffSets
310 , breakInfo_resty = exprType (deAnnotate' newRhs)
312 let breakInstr = case arr of (BA arr#) -> BRK_FUN arr# tickNumber breakInfo
313 return $ breakInstr `consOL` code
314 | otherwise = schemeE d 0 p rhs
316 getVarOffSets :: Int -> BCEnv -> TickInfo -> [(Id, Int)]
317 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
319 getOffSet :: Int -> BCEnv -> Id -> Maybe (Id, Int)
321 = case lookupBCEnv_maybe env id of
323 Just offset -> Just (id, d - offset)
325 fvsToEnv :: BCEnv -> VarSet -> [Id]
326 -- Takes the free variables of a right-hand side, and
327 -- delivers an ordered list of the local variables that will
328 -- be captured in the thunk for the RHS
329 -- The BCEnv argument tells which variables are in the local
330 -- environment: these are the ones that should be captured
332 -- The code that constructs the thunk, and the code that executes
333 -- it, have to agree about this layout
334 fvsToEnv p fvs = [v | v <- varSetElems fvs,
335 isId v, -- Could be a type variable
338 -- -----------------------------------------------------------------------------
343 { tickInfo_number :: Int -- the (module) unique number of the tick
344 , tickInfo_module :: Module -- the origin of the ticked expression
345 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
348 instance Outputable TickInfo where
349 ppr info = text "TickInfo" <+>
350 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
351 ppr (tickInfo_locals info))
353 -- Compile code to apply the given expression to the remaining args
354 -- on the stack, returning a HNF.
355 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
357 -- Delegate tail-calls to schemeT.
358 schemeE d s p e@(AnnApp f a)
361 schemeE d s p e@(AnnVar v)
362 | not (isUnLiftedType v_type)
363 = -- Lifted-type thing; push it in the normal way
367 = do -- Returning an unlifted value.
368 -- Heave it on the stack, SLIDE, and RETURN.
369 (push, szw) <- pushAtom d p (AnnVar v)
370 return (push -- value onto stack
371 `appOL` mkSLIDE szw (d-s) -- clear to sequel
372 `snocOL` RETURN_UBX v_rep) -- go
375 v_rep = typeCgRep v_type
377 schemeE d s p (AnnLit literal)
378 = do (push, szw) <- pushAtom d p (AnnLit literal)
379 let l_rep = typeCgRep (literalType literal)
380 return (push -- value onto stack
381 `appOL` mkSLIDE szw (d-s) -- clear to sequel
382 `snocOL` RETURN_UBX l_rep) -- go
384 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
385 | (AnnVar v, args_r_to_l) <- splitApp rhs,
386 Just data_con <- isDataConWorkId_maybe v,
387 dataConRepArity data_con == length args_r_to_l
388 = do -- Special case for a non-recursive let whose RHS is a
389 -- saturatred constructor application.
390 -- Just allocate the constructor and carry on
391 alloc_code <- mkConAppCode d s p data_con args_r_to_l
392 body_code <- schemeE (d+1) s (addToFM p x d) body
393 return (alloc_code `appOL` body_code)
395 -- General case for let. Generates correct, if inefficient, code in
397 schemeE d s p (AnnLet binds (_,body))
398 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
399 AnnRec xs_n_rhss -> unzip xs_n_rhss
402 fvss = map (fvsToEnv p' . fst) rhss
404 -- Sizes of free vars
405 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
407 -- the arity of each rhs
408 arities = map (length . fst . collect []) rhss
410 -- This p', d' defn is safe because all the items being pushed
411 -- are ptrs, so all have size 1. d' and p' reflect the stack
412 -- after the closures have been allocated in the heap (but not
413 -- filled in), and pointers to them parked on the stack.
414 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
416 zipE = zipEqual "schemeE"
418 -- ToDo: don't build thunks for things with no free variables
419 build_thunk dd [] size bco off arity
420 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
422 mkap | arity == 0 = MKAP
424 build_thunk dd (fv:fvs) size bco off arity = do
425 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
426 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
427 return (push_code `appOL` more_push_code)
429 alloc_code = toOL (zipWith mkAlloc sizes arities)
431 | is_tick = ALLOC_AP_NOUPD sz
432 | otherwise = ALLOC_AP sz
433 mkAlloc sz arity = ALLOC_PAP arity sz
435 is_tick = case binds of
436 AnnNonRec id _ -> occNameFS (getOccName id) == tickFS
439 compile_bind d' fvs x rhs size arity off = do
440 bco <- schemeR fvs (x,rhs)
441 build_thunk d' fvs size bco off arity
444 [ compile_bind d' fvs x rhs size arity n
445 | (fvs, x, rhs, size, arity, n) <-
446 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
449 body_code <- schemeE d' s p' body
450 thunk_codes <- sequence compile_binds
451 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
453 -- introduce a let binding for a ticked case expression. This rule
454 -- *should* only fire when the expression was not already let-bound
455 -- (the code gen for let bindings should take care of that). Todo: we
456 -- call exprFreeVars on a deAnnotated expression, this may not be the
457 -- best way to calculate the free vars but it seemed like the least
458 -- intrusive thing to do
459 schemeE d s p exp@(AnnCase {})
460 | Just (tickInfo,rhs) <- isTickedExp' exp
461 = if isUnLiftedType ty
462 then schemeE d s p (snd rhs)
465 -- Todo: is emptyVarSet correct on the next line?
466 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
468 where exp' = deAnnotate' exp
469 fvs = exprFreeVars exp'
472 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
473 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
475 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
477 -- case .... of a { DEFAULT -> ... }
478 -- becuse the return convention for both are identical.
480 -- Note that it does not matter losing the void-rep thing from the
481 -- envt (it won't be bound now) because we never look such things up.
483 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
484 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
486 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
487 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
488 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
490 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
491 | isUnboxedTupleCon dc
492 -- Similarly, convert
493 -- case .... of x { (# a #) -> ... }
495 -- case .... of a { DEFAULT -> ... }
496 = --trace "automagic mashing of case alts (# a #)" $
497 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
499 schemeE d s p (AnnCase scrut bndr _ alts)
500 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
502 schemeE d s p (AnnNote note (_, body))
505 schemeE d s p (AnnCast (_, body) _)
509 = pprPanic "ByteCodeGen.schemeE: unhandled case"
510 (pprCoreExpr (deAnnotate' other))
516 A ticked expression looks like this:
518 case tick<n> var1 ... varN of DEFAULT -> e
520 (*) <n> is the number of the tick, which is unique within a module
521 (*) var1 ... varN are the local variables in scope at the tick site
523 If we find a ticked expression we return:
525 Just ((n, [var1 ... varN]), e)
527 otherwise we return Nothing.
529 The idea is that the "case tick<n> ..." is really just an annotation on
530 the code. When we find such a thing, we pull out the useful information,
531 and then compile the code as if it was just the expression "e".
535 isTickedExp :: AnnExpr Id a -> Maybe (TickInfo, AnnExpr Id a)
536 isTickedExp (annot, expr) = isTickedExp' expr
538 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
539 isTickedExp' (AnnCase scrut _bndr _type alts)
540 | Just tickInfo <- isTickedScrut scrut,
541 [(DEFAULT, _bndr, rhs)] <- alts
542 = Just (tickInfo, rhs)
544 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
547 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
548 = Just $ TickInfo { tickInfo_number = tickNumber
549 , tickInfo_module = modName
550 , tickInfo_locals = idsOfArgs args
552 | otherwise = Nothing
554 (f, args) = collectArgs $ deAnnotate expr
555 idsOfArgs :: [Expr Id] -> [Id]
556 idsOfArgs = catMaybes . map exprId
557 exprId :: Expr Id -> Maybe Id
558 exprId (Var id) = Just id
559 exprId other = Nothing
561 isTickedExp' other = Nothing
563 -- Compile code to do a tail call. Specifically, push the fn,
564 -- slide the on-stack app back down to the sequel depth,
565 -- and enter. Four cases:
568 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
569 -- The int will be on the stack. Generate a code sequence
570 -- to convert it to the relevant constructor, SLIDE and ENTER.
572 -- 1. The fn denotes a ccall. Defer to generateCCall.
574 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
575 -- it simply as b -- since the representations are identical
576 -- (the VoidArg takes up zero stack space). Also, spot
577 -- (# b #) and treat it as b.
579 -- 3. Application of a constructor, by defn saturated.
580 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
581 -- then the ptrs, and then do PACK and RETURN.
583 -- 4. Otherwise, it must be a function call. Push the args
584 -- right to left, SLIDE and ENTER.
586 schemeT :: Int -- Stack depth
587 -> Sequel -- Sequel depth
588 -> BCEnv -- stack env
589 -> AnnExpr' Id VarSet
594 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
595 -- = panic "schemeT ?!?!"
597 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
601 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
602 = do (push, arg_words) <- pushAtom d p arg
603 tagToId_sequence <- implement_tagToId constr_names
604 return (push `appOL` tagToId_sequence
605 `appOL` mkSLIDE 1 (d+arg_words-s)
609 | Just (CCall ccall_spec) <- isFCallId_maybe fn
610 = generateCCall d s p ccall_spec fn args_r_to_l
612 -- Case 2: Constructor application
613 | Just con <- maybe_saturated_dcon,
614 isUnboxedTupleCon con
615 = case args_r_to_l of
616 [arg1,arg2] | isVoidArgAtom arg1 ->
617 unboxedTupleReturn d s p arg2
618 [arg1,arg2] | isVoidArgAtom arg2 ->
619 unboxedTupleReturn d s p arg1
620 _other -> unboxedTupleException
622 -- Case 3: Ordinary data constructor
623 | Just con <- maybe_saturated_dcon
624 = do alloc_con <- mkConAppCode d s p con args_r_to_l
625 return (alloc_con `appOL`
626 mkSLIDE 1 (d - s) `snocOL`
629 -- Case 4: Tail call of function
631 = doTailCall d s p fn args_r_to_l
634 -- Detect and extract relevant info for the tagToEnum kludge.
635 maybe_is_tagToEnum_call
636 = let extract_constr_Names ty
637 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
639 = map (getName . dataConWorkId) (tyConDataCons tyc)
640 -- NOTE: use the worker name, not the source name of
641 -- the DataCon. See DataCon.lhs for details.
643 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
646 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
647 -> case isPrimOpId_maybe v of
648 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
652 -- Extract the args (R->L) and fn
653 -- The function will necessarily be a variable,
654 -- because we are compiling a tail call
655 (AnnVar fn, args_r_to_l) = splitApp app
657 -- Only consider this to be a constructor application iff it is
658 -- saturated. Otherwise, we'll call the constructor wrapper.
659 n_args = length args_r_to_l
661 = case isDataConWorkId_maybe fn of
662 Just con | dataConRepArity con == n_args -> Just con
665 -- -----------------------------------------------------------------------------
666 -- Generate code to build a constructor application,
667 -- leaving it on top of the stack
669 mkConAppCode :: Int -> Sequel -> BCEnv
670 -> DataCon -- The data constructor
671 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
674 mkConAppCode orig_d s p con [] -- Nullary constructor
675 = ASSERT( isNullaryRepDataCon con )
676 return (unitOL (PUSH_G (getName (dataConWorkId con))))
677 -- Instead of doing a PACK, which would allocate a fresh
678 -- copy of this constructor, use the single shared version.
680 mkConAppCode orig_d s p con args_r_to_l
681 = ASSERT( dataConRepArity con == length args_r_to_l )
682 do_pushery orig_d (non_ptr_args ++ ptr_args)
684 -- The args are already in reverse order, which is the way PACK
685 -- expects them to be. We must push the non-ptrs after the ptrs.
686 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
688 do_pushery d (arg:args)
689 = do (push, arg_words) <- pushAtom d p arg
690 more_push_code <- do_pushery (d+arg_words) args
691 return (push `appOL` more_push_code)
693 = return (unitOL (PACK con n_arg_words))
695 n_arg_words = d - orig_d
698 -- -----------------------------------------------------------------------------
699 -- Returning an unboxed tuple with one non-void component (the only
700 -- case we can handle).
702 -- Remember, we don't want to *evaluate* the component that is being
703 -- returned, even if it is a pointed type. We always just return.
706 :: Int -> Sequel -> BCEnv
707 -> AnnExpr' Id VarSet -> BcM BCInstrList
708 unboxedTupleReturn d s p arg = do
709 (push, sz) <- pushAtom d p arg
711 mkSLIDE sz (d-s) `snocOL`
712 RETURN_UBX (atomRep arg))
714 -- -----------------------------------------------------------------------------
715 -- Generate code for a tail-call
718 :: Int -> Sequel -> BCEnv
719 -> Id -> [AnnExpr' Id VarSet]
721 doTailCall init_d s p fn args
722 = do_pushes init_d args (map atomRep args)
724 do_pushes d [] reps = do
725 ASSERT( null reps ) return ()
726 (push_fn, sz) <- pushAtom d p (AnnVar fn)
727 ASSERT( sz == 1 ) return ()
728 return (push_fn `appOL` (
729 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
731 do_pushes d args reps = do
732 let (push_apply, n, rest_of_reps) = findPushSeq reps
733 (these_args, rest_of_args) = splitAt n args
734 (next_d, push_code) <- push_seq d these_args
735 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
736 -- ^^^ for the PUSH_APPLY_ instruction
737 return (push_code `appOL` (push_apply `consOL` instrs))
739 push_seq d [] = return (d, nilOL)
740 push_seq d (arg:args) = do
741 (push_code, sz) <- pushAtom d p arg
742 (final_d, more_push_code) <- push_seq (d+sz) args
743 return (final_d, push_code `appOL` more_push_code)
745 -- v. similar to CgStackery.findMatch, ToDo: merge
746 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
747 = (PUSH_APPLY_PPPPPP, 6, rest)
748 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
749 = (PUSH_APPLY_PPPPP, 5, rest)
750 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
751 = (PUSH_APPLY_PPPP, 4, rest)
752 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
753 = (PUSH_APPLY_PPP, 3, rest)
754 findPushSeq (PtrArg: PtrArg: rest)
755 = (PUSH_APPLY_PP, 2, rest)
756 findPushSeq (PtrArg: rest)
757 = (PUSH_APPLY_P, 1, rest)
758 findPushSeq (VoidArg: rest)
759 = (PUSH_APPLY_V, 1, rest)
760 findPushSeq (NonPtrArg: rest)
761 = (PUSH_APPLY_N, 1, rest)
762 findPushSeq (FloatArg: rest)
763 = (PUSH_APPLY_F, 1, rest)
764 findPushSeq (DoubleArg: rest)
765 = (PUSH_APPLY_D, 1, rest)
766 findPushSeq (LongArg: rest)
767 = (PUSH_APPLY_L, 1, rest)
769 = panic "ByteCodeGen.findPushSeq"
771 -- -----------------------------------------------------------------------------
774 doCase :: Int -> Sequel -> BCEnv
775 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
776 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
778 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
780 -- Top of stack is the return itbl, as usual.
781 -- underneath it is the pointer to the alt_code BCO.
782 -- When an alt is entered, it assumes the returned value is
783 -- on top of the itbl.
786 -- An unlifted value gets an extra info table pushed on top
787 -- when it is returned.
788 unlifted_itbl_sizeW | isAlgCase = 0
791 -- depth of stack after the return value has been pushed
792 d_bndr = d + ret_frame_sizeW + idSizeW bndr
794 -- depth of stack after the extra info table for an unboxed return
795 -- has been pushed, if any. This is the stack depth at the
797 d_alts = d_bndr + unlifted_itbl_sizeW
799 -- Env in which to compile the alts, not including
800 -- any vars bound by the alts themselves
801 p_alts = addToFM p bndr (d_bndr - 1)
803 bndr_ty = idType bndr
804 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
806 -- given an alt, return a discr and code for it.
807 codeAlt alt@(DEFAULT, _, (_,rhs))
808 = do rhs_code <- schemeE d_alts s p_alts rhs
809 return (NoDiscr, rhs_code)
811 codeAlt alt@(discr, bndrs, (_,rhs))
812 -- primitive or nullary constructor alt: no need to UNPACK
813 | null real_bndrs = do
814 rhs_code <- schemeE d_alts s p_alts rhs
815 return (my_discr alt, rhs_code)
816 -- algebraic alt with some binders
817 | ASSERT(isAlgCase) otherwise =
819 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
820 ptr_sizes = map idSizeW ptrs
821 nptrs_sizes = map idSizeW nptrs
822 bind_sizes = ptr_sizes ++ nptrs_sizes
823 size = sum ptr_sizes + sum nptrs_sizes
824 -- the UNPACK instruction unpacks in reverse order...
825 p' = addListToFM p_alts
826 (zip (reverse (ptrs ++ nptrs))
827 (mkStackOffsets d_alts (reverse bind_sizes)))
829 rhs_code <- schemeE (d_alts+size) s p' rhs
830 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
832 real_bndrs = filter (not.isTyVar) bndrs
834 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
835 my_discr (DataAlt dc, binds, rhs)
836 | isUnboxedTupleCon dc
837 = unboxedTupleException
839 = DiscrP (dataConTag dc - fIRST_TAG)
840 my_discr (LitAlt l, binds, rhs)
841 = case l of MachInt i -> DiscrI (fromInteger i)
842 MachFloat r -> DiscrF (fromRational r)
843 MachDouble r -> DiscrD (fromRational r)
844 MachChar i -> DiscrI (ord i)
845 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
848 | not isAlgCase = Nothing
850 = case [dc | (DataAlt dc, _, _) <- alts] of
852 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
854 -- the bitmap is relative to stack depth d, i.e. before the
855 -- BCO, info table and return value are pushed on.
856 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
857 -- except that here we build the bitmap from the known bindings of
858 -- things that are pointers, whereas in CgBindery the code builds the
859 -- bitmap from the free slots and unboxed bindings.
862 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
863 -- The bitmap must cover the portion of the stack up to the sequel only.
864 -- Previously we were building a bitmap for the whole depth (d), but we
865 -- really want a bitmap up to depth (d-s). This affects compilation of
866 -- case-of-case expressions, which is the only time we can be compiling a
867 -- case expression with s /= 0.
869 bitmap = intsToReverseBitmap bitmap_size{-size-}
870 (sortLe (<=) (filter (< bitmap_size) rel_slots))
873 rel_slots = concat (map spread binds)
875 | isFollowableArg (idCgRep id) = [ rel_offset ]
877 where rel_offset = d - offset - 1
880 alt_stuff <- mapM codeAlt alts
881 alt_final <- mkMultiBranch maybe_ncons alt_stuff
884 alt_bco_name = getName bndr
885 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
886 0{-no arity-} bitmap_size bitmap True{-is alts-}
888 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
889 -- "\n bitmap = " ++ show bitmap) $ do
890 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
891 alt_bco' <- emitBc alt_bco
893 | isAlgCase = PUSH_ALTS alt_bco'
894 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
895 return (push_alts `consOL` scrut_code)
898 -- -----------------------------------------------------------------------------
899 -- Deal with a CCall.
901 -- Taggedly push the args onto the stack R->L,
902 -- deferencing ForeignObj#s and adjusting addrs to point to
903 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
904 -- (machine) code for the ccall, and create bytecodes to call that and
905 -- then return in the right way.
907 generateCCall :: Int -> Sequel -- stack and sequel depths
909 -> CCallSpec -- where to call
910 -> Id -- of target, for type info
911 -> [AnnExpr' Id VarSet] -- args (atoms)
914 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
917 addr_sizeW = cgRepSizeW NonPtrArg
919 -- Get the args on the stack, with tags and suitably
920 -- dereferenced for the CCall. For each arg, return the
921 -- depth to the first word of the bits for that arg, and the
922 -- CgRep of what was actually pushed.
924 pargs d [] = return []
926 = let arg_ty = repType (exprType (deAnnotate' a))
928 in case splitTyConApp_maybe arg_ty of
929 -- Don't push the FO; instead push the Addr# it
932 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
933 -> do rest <- pargs (d + addr_sizeW) az
934 code <- parg_ArrayishRep arrPtrsHdrSize d p a
935 return ((code,NonPtrArg):rest)
937 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
938 -> do rest <- pargs (d + addr_sizeW) az
939 code <- parg_ArrayishRep arrWordsHdrSize d p a
940 return ((code,NonPtrArg):rest)
942 -- Default case: push taggedly, but otherwise intact.
944 -> do (code_a, sz_a) <- pushAtom d p a
945 rest <- pargs (d+sz_a) az
946 return ((code_a, atomRep a) : rest)
948 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
949 -- the stack but then advance it over the headers, so as to
950 -- point to the payload.
951 parg_ArrayishRep hdrSize d p a
952 = do (push_fo, _) <- pushAtom d p a
953 -- The ptr points at the header. Advance it over the
954 -- header and then pretend this is an Addr#.
955 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
958 code_n_reps <- pargs d0 args_r_to_l
960 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
962 push_args = concatOL pushs_arg
963 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
965 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
966 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
968 = reverse (tail a_reps_pushed_r_to_l)
970 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
971 -- push_args is the code to do that.
972 -- d_after_args is the stack depth once the args are on.
974 -- Get the result rep.
975 (returns_void, r_rep)
976 = case maybe_getCCallReturnRep (idType fn) of
977 Nothing -> (True, VoidArg)
978 Just rr -> (False, rr)
980 Because the Haskell stack grows down, the a_reps refer to
981 lowest to highest addresses in that order. The args for the call
982 are on the stack. Now push an unboxed Addr# indicating
983 the C function to call. Then push a dummy placeholder for the
984 result. Finally, emit a CCALL insn with an offset pointing to the
985 Addr# just pushed, and a literal field holding the mallocville
986 address of the piece of marshalling code we generate.
987 So, just prior to the CCALL insn, the stack looks like this
988 (growing down, as usual):
993 Addr# address_of_C_fn
994 <placeholder-for-result#> (must be an unboxed type)
996 The interpreter then calls the marshall code mentioned
997 in the CCALL insn, passing it (& <placeholder-for-result#>),
998 that is, the addr of the topmost word in the stack.
999 When this returns, the placeholder will have been
1000 filled in. The placeholder is slid down to the sequel
1001 depth, and we RETURN.
1003 This arrangement makes it simple to do f-i-dynamic since the Addr#
1004 value is the first arg anyway.
1006 The marshalling code is generated specifically for this
1007 call site, and so knows exactly the (Haskell) stack
1008 offsets of the args, fn address and placeholder. It
1009 copies the args to the C stack, calls the stacked addr,
1010 and parks the result back in the placeholder. The interpreter
1011 calls it as a normal C call, assuming it has a signature
1012 void marshall_code ( StgWord* ptr_to_top_of_stack )
1014 -- resolve static address
1018 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
1020 -> do res <- ioToBc (lookupStaticPtr target)
1023 (is_static, static_target_addr) <- get_target_info
1026 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1027 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1028 | is_static = a_reps_pushed_RAW
1029 | otherwise = if null a_reps_pushed_RAW
1030 then panic "ByteCodeGen.generateCCall: dyn with no args"
1031 else tail a_reps_pushed_RAW
1034 (push_Addr, d_after_Addr)
1036 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1037 d_after_args + addr_sizeW)
1038 | otherwise -- is already on the stack
1039 = (nilOL, d_after_args)
1041 -- Push the return placeholder. For a call returning nothing,
1042 -- this is a VoidArg (tag).
1043 r_sizeW = cgRepSizeW r_rep
1044 d_after_r = d_after_Addr + r_sizeW
1045 r_lit = mkDummyLiteral r_rep
1046 push_r = (if returns_void
1048 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1050 -- generate the marshalling code we're going to call
1053 arg1_offW = r_sizeW + addr_sizeW
1054 args_offW = map (arg1_offW +)
1055 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
1057 addr_of_marshaller <- ioToBc (mkMarshalCode cconv
1058 (r_offW, r_rep) addr_offW
1059 (zip args_offW a_reps))
1060 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1062 -- Offset of the next stack frame down the stack. The CCALL
1063 -- instruction needs to describe the chunk of stack containing
1064 -- the ccall args to the GC, so it needs to know how large it
1065 -- is. See comment in Interpreter.c with the CCALL instruction.
1066 stk_offset = d_after_r - s
1069 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1071 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1072 `snocOL` RETURN_UBX r_rep
1074 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1077 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1081 -- Make a dummy literal, to be used as a placeholder for FFI return
1082 -- values on the stack.
1083 mkDummyLiteral :: CgRep -> Literal
1086 NonPtrArg -> MachWord 0
1087 DoubleArg -> MachDouble 0
1088 FloatArg -> MachFloat 0
1089 LongArg -> MachWord64 0
1090 _ -> moan64 "mkDummyLiteral" (ppr pr)
1094 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1095 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1098 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1100 -- Alternatively, for call-targets returning nothing, convert
1102 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1103 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1107 maybe_getCCallReturnRep :: Type -> Maybe CgRep
1108 maybe_getCCallReturnRep fn_ty
1109 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1111 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1113 = case splitTyConApp_maybe (repType r_ty) of
1114 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
1116 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
1117 || r_reps == [VoidArg] )
1118 && isUnboxedTupleTyCon r_tycon
1119 && case maybe_r_rep_to_go of
1121 Just r_rep -> r_rep /= PtrArg
1122 -- if it was, it would be impossible
1123 -- to create a valid return value
1124 -- placeholder on the stack
1125 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1128 --trace (showSDoc (ppr (a_reps, r_reps))) $
1129 if ok then maybe_r_rep_to_go else blargh
1131 -- Compile code which expects an unboxed Int on the top of stack,
1132 -- (call it i), and pushes the i'th closure in the supplied list
1133 -- as a consequence.
1134 implement_tagToId :: [Name] -> BcM BCInstrList
1135 implement_tagToId names
1136 = ASSERT( notNull names )
1137 do labels <- getLabelsBc (length names)
1138 label_fail <- getLabelBc
1139 label_exit <- getLabelBc
1140 let infos = zip4 labels (tail labels ++ [label_fail])
1142 steps = map (mkStep label_exit) infos
1143 return (concatOL steps
1145 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1147 mkStep l_exit (my_label, next_label, n, name_for_n)
1148 = toOL [LABEL my_label,
1149 TESTEQ_I n next_label,
1154 -- -----------------------------------------------------------------------------
1157 -- Push an atom onto the stack, returning suitable code & number of
1158 -- stack words used.
1160 -- The env p must map each variable to the highest- numbered stack
1161 -- slot for it. For example, if the stack has depth 4 and we
1162 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1163 -- the tag in stack[5], the stack will have depth 6, and p must map v
1164 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1165 -- depth 6 stack has valid words 0 .. 5.
1167 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1169 pushAtom d p (AnnApp f (_, AnnType _))
1170 = pushAtom d p (snd f)
1172 pushAtom d p (AnnNote note e)
1173 = pushAtom d p (snd e)
1175 pushAtom d p (AnnLam x e)
1177 = pushAtom d p (snd e)
1179 pushAtom d p (AnnVar v)
1181 | idCgRep v == VoidArg
1185 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1187 | Just primop <- isPrimOpId_maybe v
1188 = return (unitOL (PUSH_PRIMOP primop), 1)
1190 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1191 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1192 -- d - d_v the number of words between the TOS
1193 -- and the 1st slot of the object
1195 -- d - d_v - 1 the offset from the TOS of the 1st slot
1197 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1200 -- Having found the last slot, we proceed to copy the right number of
1201 -- slots on to the top of the stack.
1203 | otherwise -- v must be a global variable
1205 return (unitOL (PUSH_G (getName v)), sz)
1211 pushAtom d p (AnnLit lit)
1213 MachLabel fs _ -> code NonPtrArg
1214 MachWord w -> code NonPtrArg
1215 MachInt i -> code PtrArg
1216 MachFloat r -> code FloatArg
1217 MachDouble r -> code DoubleArg
1218 MachChar c -> code NonPtrArg
1219 MachStr s -> pushStr s
1222 = let size_host_words = cgRepSizeW rep
1223 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1227 = let getMallocvilleAddr
1229 FastString _ n _ fp _ ->
1230 -- we could grab the Ptr from the ForeignPtr,
1231 -- but then we have no way to control its lifetime.
1232 -- In reality it'll probably stay alive long enoungh
1233 -- by virtue of the global FastString table, but
1234 -- to be on the safe side we copy the string into
1235 -- a malloc'd area of memory.
1236 do ptr <- ioToBc (mallocBytes (n+1))
1239 withForeignPtr fp $ \p -> do
1240 memcpy ptr p (fromIntegral n)
1241 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1245 addr <- getMallocvilleAddr
1246 -- Get the addr on the stack, untaggedly
1247 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1249 pushAtom d p (AnnCast e _)
1250 = pushAtom d p (snd e)
1253 = pprPanic "ByteCodeGen.pushAtom"
1254 (pprCoreExpr (deAnnotate (undefined, other)))
1256 foreign import ccall unsafe "memcpy"
1257 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1260 -- -----------------------------------------------------------------------------
1261 -- Given a bunch of alts code and their discrs, do the donkey work
1262 -- of making a multiway branch using a switch tree.
1263 -- What a load of hassle!
1265 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1266 -- a hint; generates better code
1267 -- Nothing is always safe
1268 -> [(Discr, BCInstrList)]
1270 mkMultiBranch maybe_ncons raw_ways
1271 = let d_way = filter (isNoDiscr.fst) raw_ways
1273 (\w1 w2 -> leAlt (fst w1) (fst w2))
1274 (filter (not.isNoDiscr.fst) raw_ways)
1276 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1277 mkTree [] range_lo range_hi = return the_default
1279 mkTree [val] range_lo range_hi
1280 | range_lo `eqAlt` range_hi
1283 = do label_neq <- getLabelBc
1284 return (mkTestEQ (fst val) label_neq
1286 `appOL` unitOL (LABEL label_neq)
1287 `appOL` the_default))
1289 mkTree vals range_lo range_hi
1290 = let n = length vals `div` 2
1291 vals_lo = take n vals
1292 vals_hi = drop n vals
1293 v_mid = fst (head vals_hi)
1295 label_geq <- getLabelBc
1296 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1297 code_hi <- mkTree vals_hi v_mid range_hi
1298 return (mkTestLT v_mid label_geq
1300 `appOL` unitOL (LABEL label_geq)
1304 = case d_way of [] -> unitOL CASEFAIL
1307 -- None of these will be needed if there are no non-default alts
1308 (mkTestLT, mkTestEQ, init_lo, init_hi)
1310 = panic "mkMultiBranch: awesome foursome"
1312 = case fst (head notd_ways) of {
1313 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1314 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1317 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1318 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1321 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1322 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1325 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1326 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1328 DiscrP algMaxBound )
1331 (algMinBound, algMaxBound)
1332 = case maybe_ncons of
1333 Just n -> (0, n - 1)
1334 Nothing -> (minBound, maxBound)
1336 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1337 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1338 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1339 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1340 NoDiscr `eqAlt` NoDiscr = True
1343 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1344 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1345 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1346 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1347 NoDiscr `leAlt` NoDiscr = True
1350 isNoDiscr NoDiscr = True
1353 dec (DiscrI i) = DiscrI (i-1)
1354 dec (DiscrP i) = DiscrP (i-1)
1355 dec other = other -- not really right, but if you
1356 -- do cases on floating values, you'll get what you deserve
1358 -- same snotty comment applies to the following
1360 minD, maxD :: Double
1366 mkTree notd_ways init_lo init_hi
1369 -- -----------------------------------------------------------------------------
1370 -- Supporting junk for the compilation schemes
1372 -- Describes case alts
1380 instance Outputable Discr where
1381 ppr (DiscrI i) = int i
1382 ppr (DiscrF f) = text (show f)
1383 ppr (DiscrD d) = text (show d)
1384 ppr (DiscrP i) = int i
1385 ppr NoDiscr = text "DEF"
1388 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1389 lookupBCEnv_maybe = lookupFM
1391 idSizeW :: Id -> Int
1392 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1395 unboxedTupleException :: a
1396 unboxedTupleException
1399 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1400 " Possibly due to foreign import/export decls in source.\n"++
1401 " Workaround: use -fobject-code, or compile this module to .o separately."))
1404 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1407 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1408 -- The arguments are returned in *right-to-left* order
1409 splitApp (AnnApp (_,f) (_,a))
1410 | isTypeAtom a = splitApp f
1411 | otherwise = case splitApp f of
1412 (f', as) -> (f', a:as)
1413 splitApp (AnnNote n (_,e)) = splitApp e
1414 splitApp (AnnCast (_,e) _) = splitApp e
1415 splitApp e = (e, [])
1418 isTypeAtom :: AnnExpr' id ann -> Bool
1419 isTypeAtom (AnnType _) = True
1420 isTypeAtom _ = False
1422 isVoidArgAtom :: AnnExpr' id ann -> Bool
1423 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1424 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1425 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1426 isVoidArgAtom _ = False
1428 atomRep :: AnnExpr' Id ann -> CgRep
1429 atomRep (AnnVar v) = typeCgRep (idType v)
1430 atomRep (AnnLit l) = typeCgRep (literalType l)
1431 atomRep (AnnNote n b) = atomRep (snd b)
1432 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1433 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1434 atomRep (AnnCast b _) = atomRep (snd b)
1435 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1437 isPtrAtom :: AnnExpr' Id ann -> Bool
1438 isPtrAtom e = atomRep e == PtrArg
1440 -- Let szsw be the sizes in words of some items pushed onto the stack,
1441 -- which has initial depth d'. Return the values which the stack environment
1442 -- should map these items to.
1443 mkStackOffsets :: Int -> [Int] -> [Int]
1444 mkStackOffsets original_depth szsw
1445 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1447 -- -----------------------------------------------------------------------------
1448 -- The bytecode generator's monad
1450 type BcPtr = Either ItblPtr (Ptr ())
1454 uniqSupply :: UniqSupply, -- for generating fresh variable names
1455 nextlabel :: Int, -- for generating local labels
1456 malloced :: [BcPtr], -- thunks malloced for current BCO
1457 -- Should be free()d when it is GCd
1458 breakArray :: BreakArray -- array of breakpoint flags
1461 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1463 ioToBc :: IO a -> BcM a
1464 ioToBc io = BcM $ \st -> do
1468 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1469 runBc us modBreaks (BcM m)
1470 = m (BcM_State us 0 [] breakArray)
1472 breakArray = modBreaks_flags modBreaks
1474 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1475 thenBc (BcM expr) cont = BcM $ \st0 -> do
1476 (st1, q) <- expr st0
1481 thenBc_ :: BcM a -> BcM b -> BcM b
1482 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1483 (st1, q) <- expr st0
1484 (st2, r) <- cont st1
1487 returnBc :: a -> BcM a
1488 returnBc result = BcM $ \st -> (return (st, result))
1490 instance Monad BcM where
1495 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1497 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1499 recordMallocBc :: Ptr a -> BcM ()
1501 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1503 recordItblMallocBc :: ItblPtr -> BcM ()
1504 recordItblMallocBc a
1505 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1507 getLabelBc :: BcM Int
1509 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1511 getLabelsBc :: Int -> BcM [Int]
1513 = BcM $ \st -> let ctr = nextlabel st
1514 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1516 getBreakArray :: BcM BreakArray
1517 getBreakArray = BcM $ \st -> return (st, breakArray st)
1519 newUnique :: BcM Unique
1521 \st -> case splitUniqSupply (uniqSupply st) of
1522 (us1, us2) -> let newState = st { uniqSupply = us2 }
1523 in return (newState, uniqFromSupply us1)
1525 newId :: Type -> BcM Id
1528 return $ mkSysLocal tickFS uniq ty
1530 tickFS = FSLIT("ticked")