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"
60 import Data.List ( intersperse, sortBy, zip4, zip6, partition )
63 import Control.Exception ( throwDyn )
65 import GHC.Exts ( Int(..), ByteArray# )
67 import Control.Monad ( when )
68 import Data.Char ( ord, chr )
76 -- -----------------------------------------------------------------------------
77 -- Generating byte code for a complete module
79 byteCodeGen :: DynFlags
83 -> IO CompiledByteCode
84 byteCodeGen dflags binds tycs modBreaks
85 = do showPass dflags "ByteCodeGen"
87 let flatBinds = [ (bndr, freeVars rhs)
88 | (bndr, rhs) <- flattenBinds binds]
90 us <- mkSplitUniqSupply 'y'
91 (BcM_State _us final_ctr mallocd _, proto_bcos)
92 <- runBc us modBreaks (mapM schemeTopBind flatBinds)
94 when (notNull mallocd)
95 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
97 dumpIfSet_dyn dflags Opt_D_dump_BCOs
98 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
100 assembleBCOs proto_bcos tycs
102 -- -----------------------------------------------------------------------------
103 -- Generating byte code for an expression
105 -- Returns: (the root BCO for this expression,
106 -- a list of auxilary BCOs resulting from compiling closures)
107 coreExprToBCOs :: DynFlags
110 coreExprToBCOs dflags expr
111 = do showPass dflags "ByteCodeGen"
113 -- create a totally bogus name for the top-level BCO; this
114 -- should be harmless, since it's never used for anything
115 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
116 invented_id = Id.mkLocalId invented_name (panic "invented_id's type")
118 -- the uniques are needed to generate fresh variables when we introduce new
119 -- let bindings for ticked expressions
120 us <- mkSplitUniqSupply 'y'
121 (BcM_State _us final_ctr mallocd _ , proto_bco)
122 <- runBc us emptyModBreaks (schemeTopBind (invented_id, freeVars expr))
124 when (notNull mallocd)
125 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
127 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
129 assembleBCO proto_bco
132 -- -----------------------------------------------------------------------------
133 -- Compilation schema for the bytecode generator
135 type BCInstrList = OrdList BCInstr
137 type Sequel = Int -- back off to this depth before ENTER
139 -- Maps Ids to the offset from the stack _base_ so we don't have
140 -- to mess with it after each push/pop.
141 type BCEnv = FiniteMap Id Int -- To find vars on the stack
143 ppBCEnv :: BCEnv -> SDoc
146 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
149 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
150 cmp_snd x y = compare (snd x) (snd y)
152 -- Create a BCO and do a spot of peephole optimisation on the insns
157 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
161 -> Bool -- True <=> is a return point, rather than a function
164 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap is_ret mallocd_blocks
167 protoBCOInstrs = maybe_with_stack_check,
168 protoBCOBitmap = bitmap,
169 protoBCOBitmapSize = bitmap_size,
170 protoBCOArity = arity,
171 protoBCOExpr = origin,
172 protoBCOPtrs = mallocd_blocks
175 -- Overestimate the stack usage (in words) of this BCO,
176 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
177 -- stack check. (The interpreter always does a stack check
178 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
179 -- BCO anyway, so we only need to add an explicit on in the
180 -- (hopefully rare) cases when the (overestimated) stack use
181 -- exceeds iNTERP_STACK_CHECK_THRESH.
182 maybe_with_stack_check
183 | is_ret && stack_usage < aP_STACK_SPLIM = peep_d
184 -- don't do stack checks at return points,
185 -- everything is aggregated up to the top BCO
186 -- (which must be a function).
187 -- That is, unless the stack usage is >= AP_STACK_SPLIM,
189 | stack_usage >= iNTERP_STACK_CHECK_THRESH
190 = STKCHECK stack_usage : peep_d
192 = peep_d -- the supposedly common case
194 -- We assume that this sum doesn't wrap
195 stack_usage = sum (map bciStackUse peep_d)
197 -- Merge local pushes
198 peep_d = peep (fromOL instrs_ordlist)
200 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
201 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
202 peep (PUSH_L off1 : PUSH_L off2 : rest)
203 = PUSH_LL off1 (off2-1) : peep rest
209 argBits :: [CgRep] -> [Bool]
212 | isFollowableArg rep = False : argBits args
213 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
215 -- -----------------------------------------------------------------------------
218 -- Compile code for the right-hand side of a top-level binding
220 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
223 schemeTopBind (id, rhs)
224 | Just data_con <- isDataConWorkId_maybe id,
225 isNullaryRepDataCon data_con = do
226 -- Special case for the worker of a nullary data con.
227 -- It'll look like this: Nil = /\a -> Nil a
228 -- If we feed it into schemeR, we'll get
230 -- because mkConAppCode treats nullary constructor applications
231 -- by just re-using the single top-level definition. So
232 -- for the worker itself, we must allocate it directly.
233 -- ioToBc (putStrLn $ "top level BCO")
234 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
235 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
238 = schemeR [{- No free variables -}] (id, rhs)
241 -- -----------------------------------------------------------------------------
244 -- Compile code for a right-hand side, to give a BCO that,
245 -- when executed with the free variables and arguments on top of the stack,
246 -- will return with a pointer to the result on top of the stack, after
247 -- removing the free variables and arguments.
249 -- Park the resulting BCO in the monad. Also requires the
250 -- variable to which this value was bound, so as to give the
251 -- resulting BCO a name.
253 schemeR :: [Id] -- Free vars of the RHS, ordered as they
254 -- will appear in the thunk. Empty for
255 -- top-level things, which have no free vars.
256 -> (Id, AnnExpr Id VarSet)
257 -> BcM (ProtoBCO Name)
258 schemeR fvs (nm, rhs)
262 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
263 $$ pprCoreExpr (deAnnotate rhs)
269 = schemeR_wrk fvs nm rhs (collect [] rhs)
271 collect :: [Var] -> AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
272 collect xs (_, AnnNote note e) = collect xs e
273 collect xs (_, AnnCast e _) = collect xs e
274 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
275 collect xs (_, not_lambda) = (reverse xs, not_lambda)
277 schemeR_wrk :: [Id] -> Id -> AnnExpr Id VarSet -> ([Var], AnnExpr' Var VarSet) -> BcM (ProtoBCO Name)
278 schemeR_wrk fvs nm original_body (args, body)
280 all_args = reverse args ++ fvs
281 arity = length all_args
282 -- all_args are the args in reverse order. We're compiling a function
283 -- \fv1..fvn x1..xn -> e
284 -- i.e. the fvs come first
286 szsw_args = map idSizeW all_args
287 szw_args = sum szsw_args
288 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
290 -- make the arg bitmap
291 bits = argBits (reverse (map idCgRep all_args))
292 bitmap_size = length bits
293 bitmap = mkBitmap bits
295 body_code <- schemeER_wrk szw_args p_init body
297 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
298 arity bitmap_size bitmap False{-not alts-})
300 -- introduce break instructions for ticked expressions
301 schemeER_wrk :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
303 | Just (tickInfo, (_annot, newRhs)) <- isTickedExp' rhs = do
304 code <- schemeE d 0 p newRhs
306 let idOffSets = getVarOffSets d p tickInfo
307 let tickNumber = tickInfo_number tickInfo
308 let breakInfo = BreakInfo
309 { breakInfo_module = tickInfo_module tickInfo
310 , breakInfo_number = tickNumber
311 , breakInfo_vars = idOffSets
312 , breakInfo_resty = exprType (deAnnotate' newRhs)
314 let breakInstr = case arr of (BA arr#) -> BRK_FUN arr# tickNumber breakInfo
315 return $ breakInstr `consOL` code
316 | otherwise = schemeE d 0 p rhs
318 getVarOffSets :: Int -> BCEnv -> TickInfo -> [(Id, Int)]
319 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
321 getOffSet :: Int -> BCEnv -> Id -> Maybe (Id, Int)
323 = case lookupBCEnv_maybe env id of
325 Just offset -> Just (id, d - offset)
327 fvsToEnv :: BCEnv -> VarSet -> [Id]
328 -- Takes the free variables of a right-hand side, and
329 -- delivers an ordered list of the local variables that will
330 -- be captured in the thunk for the RHS
331 -- The BCEnv argument tells which variables are in the local
332 -- environment: these are the ones that should be captured
334 -- The code that constructs the thunk, and the code that executes
335 -- it, have to agree about this layout
336 fvsToEnv p fvs = [v | v <- varSetElems fvs,
337 isId v, -- Could be a type variable
340 -- -----------------------------------------------------------------------------
345 { tickInfo_number :: Int -- the (module) unique number of the tick
346 , tickInfo_module :: Module -- the origin of the ticked expression
347 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
350 instance Outputable TickInfo where
351 ppr info = text "TickInfo" <+>
352 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
353 ppr (tickInfo_locals info))
355 -- Compile code to apply the given expression to the remaining args
356 -- on the stack, returning a HNF.
357 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
359 -- Delegate tail-calls to schemeT.
360 schemeE d s p e@(AnnApp f a)
363 schemeE d s p e@(AnnVar v)
364 | not (isUnLiftedType v_type)
365 = -- Lifted-type thing; push it in the normal way
369 = do -- Returning an unlifted value.
370 -- Heave it on the stack, SLIDE, and RETURN.
371 (push, szw) <- pushAtom d p (AnnVar v)
372 return (push -- value onto stack
373 `appOL` mkSLIDE szw (d-s) -- clear to sequel
374 `snocOL` RETURN_UBX v_rep) -- go
377 v_rep = typeCgRep v_type
379 schemeE d s p (AnnLit literal)
380 = do (push, szw) <- pushAtom d p (AnnLit literal)
381 let l_rep = typeCgRep (literalType literal)
382 return (push -- value onto stack
383 `appOL` mkSLIDE szw (d-s) -- clear to sequel
384 `snocOL` RETURN_UBX l_rep) -- go
386 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
387 | (AnnVar v, args_r_to_l) <- splitApp rhs,
388 Just data_con <- isDataConWorkId_maybe v,
389 dataConRepArity data_con == length args_r_to_l
390 = do -- Special case for a non-recursive let whose RHS is a
391 -- saturatred constructor application.
392 -- Just allocate the constructor and carry on
393 alloc_code <- mkConAppCode d s p data_con args_r_to_l
394 body_code <- schemeE (d+1) s (addToFM p x d) body
395 return (alloc_code `appOL` body_code)
397 -- General case for let. Generates correct, if inefficient, code in
399 schemeE d s p (AnnLet binds (_,body))
400 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
401 AnnRec xs_n_rhss -> unzip xs_n_rhss
404 fvss = map (fvsToEnv p' . fst) rhss
406 -- Sizes of free vars
407 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
409 -- the arity of each rhs
410 arities = map (length . fst . collect []) rhss
412 -- This p', d' defn is safe because all the items being pushed
413 -- are ptrs, so all have size 1. d' and p' reflect the stack
414 -- after the closures have been allocated in the heap (but not
415 -- filled in), and pointers to them parked on the stack.
416 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
418 zipE = zipEqual "schemeE"
420 -- ToDo: don't build thunks for things with no free variables
421 build_thunk dd [] size bco off arity
422 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
424 mkap | arity == 0 = MKAP
426 build_thunk dd (fv:fvs) size bco off arity = do
427 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
428 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
429 return (push_code `appOL` more_push_code)
431 alloc_code = toOL (zipWith mkAlloc sizes arities)
433 | is_tick = ALLOC_AP_NOUPD sz
434 | otherwise = ALLOC_AP sz
435 mkAlloc sz arity = ALLOC_PAP arity sz
437 is_tick = case binds of
438 AnnNonRec id _ -> occNameFS (getOccName id) == tickFS
441 compile_bind d' fvs x rhs size arity off = do
442 bco <- schemeR fvs (x,rhs)
443 build_thunk d' fvs size bco off arity
446 [ compile_bind d' fvs x rhs size arity n
447 | (fvs, x, rhs, size, arity, n) <-
448 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
451 body_code <- schemeE d' s p' body
452 thunk_codes <- sequence compile_binds
453 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
455 -- introduce a let binding for a ticked case expression. This rule
456 -- *should* only fire when the expression was not already let-bound
457 -- (the code gen for let bindings should take care of that). Todo: we
458 -- call exprFreeVars on a deAnnotated expression, this may not be the
459 -- best way to calculate the free vars but it seemed like the least
460 -- intrusive thing to do
461 schemeE d s p exp@(AnnCase {})
462 | Just (tickInfo,rhs) <- isTickedExp' exp
463 = if isUnLiftedType ty
464 then schemeE d s p (snd rhs)
467 -- Todo: is emptyVarSet correct on the next line?
468 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
470 where exp' = deAnnotate' exp
471 fvs = exprFreeVars exp'
474 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
475 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
477 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
479 -- case .... of a { DEFAULT -> ... }
480 -- becuse the return convention for both are identical.
482 -- Note that it does not matter losing the void-rep thing from the
483 -- envt (it won't be bound now) because we never look such things up.
485 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
486 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
488 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
489 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
490 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
492 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
493 | isUnboxedTupleCon dc
494 -- Similarly, convert
495 -- case .... of x { (# a #) -> ... }
497 -- case .... of a { DEFAULT -> ... }
498 = --trace "automagic mashing of case alts (# a #)" $
499 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
501 schemeE d s p (AnnCase scrut bndr _ alts)
502 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
504 schemeE d s p (AnnNote note (_, body))
507 schemeE d s p (AnnCast (_, body) _)
511 = pprPanic "ByteCodeGen.schemeE: unhandled case"
512 (pprCoreExpr (deAnnotate' other))
518 A ticked expression looks like this:
520 case tick<n> var1 ... varN of DEFAULT -> e
522 (*) <n> is the number of the tick, which is unique within a module
523 (*) var1 ... varN are the local variables in scope at the tick site
525 If we find a ticked expression we return:
527 Just ((n, [var1 ... varN]), e)
529 otherwise we return Nothing.
531 The idea is that the "case tick<n> ..." is really just an annotation on
532 the code. When we find such a thing, we pull out the useful information,
533 and then compile the code as if it was just the expression "e".
537 isTickedExp :: AnnExpr Id a -> Maybe (TickInfo, AnnExpr Id a)
538 isTickedExp (annot, expr) = isTickedExp' expr
540 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
541 isTickedExp' (AnnCase scrut _bndr _type alts)
542 | Just tickInfo <- isTickedScrut scrut,
543 [(DEFAULT, _bndr, rhs)] <- alts
544 = Just (tickInfo, rhs)
546 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
549 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
550 = Just $ TickInfo { tickInfo_number = tickNumber
551 , tickInfo_module = modName
552 , tickInfo_locals = idsOfArgs args
554 | otherwise = Nothing
556 (f, args) = collectArgs $ deAnnotate expr
557 idsOfArgs :: [Expr Id] -> [Id]
558 idsOfArgs = catMaybes . map exprId
559 exprId :: Expr Id -> Maybe Id
560 exprId (Var id) = Just id
561 exprId other = Nothing
563 isTickedExp' other = Nothing
565 -- Compile code to do a tail call. Specifically, push the fn,
566 -- slide the on-stack app back down to the sequel depth,
567 -- and enter. Four cases:
570 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
571 -- The int will be on the stack. Generate a code sequence
572 -- to convert it to the relevant constructor, SLIDE and ENTER.
574 -- 1. The fn denotes a ccall. Defer to generateCCall.
576 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
577 -- it simply as b -- since the representations are identical
578 -- (the VoidArg takes up zero stack space). Also, spot
579 -- (# b #) and treat it as b.
581 -- 3. Application of a constructor, by defn saturated.
582 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
583 -- then the ptrs, and then do PACK and RETURN.
585 -- 4. Otherwise, it must be a function call. Push the args
586 -- right to left, SLIDE and ENTER.
588 schemeT :: Int -- Stack depth
589 -> Sequel -- Sequel depth
590 -> BCEnv -- stack env
591 -> AnnExpr' Id VarSet
596 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
597 -- = panic "schemeT ?!?!"
599 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
603 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
604 = do (push, arg_words) <- pushAtom d p arg
605 tagToId_sequence <- implement_tagToId constr_names
606 return (push `appOL` tagToId_sequence
607 `appOL` mkSLIDE 1 (d+arg_words-s)
611 | Just (CCall ccall_spec) <- isFCallId_maybe fn
612 = generateCCall d s p ccall_spec fn args_r_to_l
614 -- Case 2: Constructor application
615 | Just con <- maybe_saturated_dcon,
616 isUnboxedTupleCon con
617 = case args_r_to_l of
618 [arg1,arg2] | isVoidArgAtom arg1 ->
619 unboxedTupleReturn d s p arg2
620 [arg1,arg2] | isVoidArgAtom arg2 ->
621 unboxedTupleReturn d s p arg1
622 _other -> unboxedTupleException
624 -- Case 3: Ordinary data constructor
625 | Just con <- maybe_saturated_dcon
626 = do alloc_con <- mkConAppCode d s p con args_r_to_l
627 return (alloc_con `appOL`
628 mkSLIDE 1 (d - s) `snocOL`
631 -- Case 4: Tail call of function
633 = doTailCall d s p fn args_r_to_l
636 -- Detect and extract relevant info for the tagToEnum kludge.
637 maybe_is_tagToEnum_call
638 = let extract_constr_Names ty
639 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
641 = map (getName . dataConWorkId) (tyConDataCons tyc)
642 -- NOTE: use the worker name, not the source name of
643 -- the DataCon. See DataCon.lhs for details.
645 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
648 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
649 -> case isPrimOpId_maybe v of
650 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
654 -- Extract the args (R->L) and fn
655 -- The function will necessarily be a variable,
656 -- because we are compiling a tail call
657 (AnnVar fn, args_r_to_l) = splitApp app
659 -- Only consider this to be a constructor application iff it is
660 -- saturated. Otherwise, we'll call the constructor wrapper.
661 n_args = length args_r_to_l
663 = case isDataConWorkId_maybe fn of
664 Just con | dataConRepArity con == n_args -> Just con
667 -- -----------------------------------------------------------------------------
668 -- Generate code to build a constructor application,
669 -- leaving it on top of the stack
671 mkConAppCode :: Int -> Sequel -> BCEnv
672 -> DataCon -- The data constructor
673 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
676 mkConAppCode orig_d s p con [] -- Nullary constructor
677 = ASSERT( isNullaryRepDataCon con )
678 return (unitOL (PUSH_G (getName (dataConWorkId con))))
679 -- Instead of doing a PACK, which would allocate a fresh
680 -- copy of this constructor, use the single shared version.
682 mkConAppCode orig_d s p con args_r_to_l
683 = ASSERT( dataConRepArity con == length args_r_to_l )
684 do_pushery orig_d (non_ptr_args ++ ptr_args)
686 -- The args are already in reverse order, which is the way PACK
687 -- expects them to be. We must push the non-ptrs after the ptrs.
688 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
690 do_pushery d (arg:args)
691 = do (push, arg_words) <- pushAtom d p arg
692 more_push_code <- do_pushery (d+arg_words) args
693 return (push `appOL` more_push_code)
695 = return (unitOL (PACK con n_arg_words))
697 n_arg_words = d - orig_d
700 -- -----------------------------------------------------------------------------
701 -- Returning an unboxed tuple with one non-void component (the only
702 -- case we can handle).
704 -- Remember, we don't want to *evaluate* the component that is being
705 -- returned, even if it is a pointed type. We always just return.
708 :: Int -> Sequel -> BCEnv
709 -> AnnExpr' Id VarSet -> BcM BCInstrList
710 unboxedTupleReturn d s p arg = do
711 (push, sz) <- pushAtom d p arg
713 mkSLIDE sz (d-s) `snocOL`
714 RETURN_UBX (atomRep arg))
716 -- -----------------------------------------------------------------------------
717 -- Generate code for a tail-call
720 :: Int -> Sequel -> BCEnv
721 -> Id -> [AnnExpr' Id VarSet]
723 doTailCall init_d s p fn args
724 = do_pushes init_d args (map atomRep args)
726 do_pushes d [] reps = do
727 ASSERT( null reps ) return ()
728 (push_fn, sz) <- pushAtom d p (AnnVar fn)
729 ASSERT( sz == 1 ) return ()
730 return (push_fn `appOL` (
731 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
733 do_pushes d args reps = do
734 let (push_apply, n, rest_of_reps) = findPushSeq reps
735 (these_args, rest_of_args) = splitAt n args
736 (next_d, push_code) <- push_seq d these_args
737 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
738 -- ^^^ for the PUSH_APPLY_ instruction
739 return (push_code `appOL` (push_apply `consOL` instrs))
741 push_seq d [] = return (d, nilOL)
742 push_seq d (arg:args) = do
743 (push_code, sz) <- pushAtom d p arg
744 (final_d, more_push_code) <- push_seq (d+sz) args
745 return (final_d, push_code `appOL` more_push_code)
747 -- v. similar to CgStackery.findMatch, ToDo: merge
748 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
749 = (PUSH_APPLY_PPPPPP, 6, rest)
750 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
751 = (PUSH_APPLY_PPPPP, 5, rest)
752 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
753 = (PUSH_APPLY_PPPP, 4, rest)
754 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
755 = (PUSH_APPLY_PPP, 3, rest)
756 findPushSeq (PtrArg: PtrArg: rest)
757 = (PUSH_APPLY_PP, 2, rest)
758 findPushSeq (PtrArg: rest)
759 = (PUSH_APPLY_P, 1, rest)
760 findPushSeq (VoidArg: rest)
761 = (PUSH_APPLY_V, 1, rest)
762 findPushSeq (NonPtrArg: rest)
763 = (PUSH_APPLY_N, 1, rest)
764 findPushSeq (FloatArg: rest)
765 = (PUSH_APPLY_F, 1, rest)
766 findPushSeq (DoubleArg: rest)
767 = (PUSH_APPLY_D, 1, rest)
768 findPushSeq (LongArg: rest)
769 = (PUSH_APPLY_L, 1, rest)
771 = panic "ByteCodeGen.findPushSeq"
773 -- -----------------------------------------------------------------------------
776 doCase :: Int -> Sequel -> BCEnv
777 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
778 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
780 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
782 -- Top of stack is the return itbl, as usual.
783 -- underneath it is the pointer to the alt_code BCO.
784 -- When an alt is entered, it assumes the returned value is
785 -- on top of the itbl.
788 -- An unlifted value gets an extra info table pushed on top
789 -- when it is returned.
790 unlifted_itbl_sizeW | isAlgCase = 0
793 -- depth of stack after the return value has been pushed
794 d_bndr = d + ret_frame_sizeW + idSizeW bndr
796 -- depth of stack after the extra info table for an unboxed return
797 -- has been pushed, if any. This is the stack depth at the
799 d_alts = d_bndr + unlifted_itbl_sizeW
801 -- Env in which to compile the alts, not including
802 -- any vars bound by the alts themselves
803 p_alts = addToFM p bndr (d_bndr - 1)
805 bndr_ty = idType bndr
806 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
808 -- given an alt, return a discr and code for it.
809 codeAlt alt@(DEFAULT, _, (_,rhs))
810 = do rhs_code <- schemeE d_alts s p_alts rhs
811 return (NoDiscr, rhs_code)
813 codeAlt alt@(discr, bndrs, (_,rhs))
814 -- primitive or nullary constructor alt: no need to UNPACK
815 | null real_bndrs = do
816 rhs_code <- schemeE d_alts s p_alts rhs
817 return (my_discr alt, rhs_code)
818 -- algebraic alt with some binders
819 | ASSERT(isAlgCase) otherwise =
821 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
822 ptr_sizes = map idSizeW ptrs
823 nptrs_sizes = map idSizeW nptrs
824 bind_sizes = ptr_sizes ++ nptrs_sizes
825 size = sum ptr_sizes + sum nptrs_sizes
826 -- the UNPACK instruction unpacks in reverse order...
827 p' = addListToFM p_alts
828 (zip (reverse (ptrs ++ nptrs))
829 (mkStackOffsets d_alts (reverse bind_sizes)))
831 rhs_code <- schemeE (d_alts+size) s p' rhs
832 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
834 real_bndrs = filter (not.isTyVar) bndrs
836 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
837 my_discr (DataAlt dc, binds, rhs)
838 | isUnboxedTupleCon dc
839 = unboxedTupleException
841 = DiscrP (dataConTag dc - fIRST_TAG)
842 my_discr (LitAlt l, binds, rhs)
843 = case l of MachInt i -> DiscrI (fromInteger i)
844 MachFloat r -> DiscrF (fromRational r)
845 MachDouble r -> DiscrD (fromRational r)
846 MachChar i -> DiscrI (ord i)
847 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
850 | not isAlgCase = Nothing
852 = case [dc | (DataAlt dc, _, _) <- alts] of
854 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
856 -- the bitmap is relative to stack depth d, i.e. before the
857 -- BCO, info table and return value are pushed on.
858 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
859 -- except that here we build the bitmap from the known bindings of
860 -- things that are pointers, whereas in CgBindery the code builds the
861 -- bitmap from the free slots and unboxed bindings.
864 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
865 -- The bitmap must cover the portion of the stack up to the sequel only.
866 -- Previously we were building a bitmap for the whole depth (d), but we
867 -- really want a bitmap up to depth (d-s). This affects compilation of
868 -- case-of-case expressions, which is the only time we can be compiling a
869 -- case expression with s /= 0.
871 bitmap = intsToReverseBitmap bitmap_size{-size-}
872 (sortLe (<=) (filter (< bitmap_size) rel_slots))
875 rel_slots = concat (map spread binds)
877 | isFollowableArg (idCgRep id) = [ rel_offset ]
879 where rel_offset = d - offset - 1
882 alt_stuff <- mapM codeAlt alts
883 alt_final <- mkMultiBranch maybe_ncons alt_stuff
886 alt_bco_name = getName bndr
887 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
888 0{-no arity-} bitmap_size bitmap True{-is alts-}
890 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
891 -- "\n bitmap = " ++ show bitmap) $ do
892 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
893 alt_bco' <- emitBc alt_bco
895 | isAlgCase = PUSH_ALTS alt_bco'
896 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
897 return (push_alts `consOL` scrut_code)
900 -- -----------------------------------------------------------------------------
901 -- Deal with a CCall.
903 -- Taggedly push the args onto the stack R->L,
904 -- deferencing ForeignObj#s and adjusting addrs to point to
905 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
906 -- (machine) code for the ccall, and create bytecodes to call that and
907 -- then return in the right way.
909 generateCCall :: Int -> Sequel -- stack and sequel depths
911 -> CCallSpec -- where to call
912 -> Id -- of target, for type info
913 -> [AnnExpr' Id VarSet] -- args (atoms)
916 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
919 addr_sizeW = cgRepSizeW NonPtrArg
921 -- Get the args on the stack, with tags and suitably
922 -- dereferenced for the CCall. For each arg, return the
923 -- depth to the first word of the bits for that arg, and the
924 -- CgRep of what was actually pushed.
926 pargs d [] = return []
928 = let arg_ty = repType (exprType (deAnnotate' a))
930 in case splitTyConApp_maybe arg_ty of
931 -- Don't push the FO; instead push the Addr# it
934 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
935 -> do rest <- pargs (d + addr_sizeW) az
936 code <- parg_ArrayishRep arrPtrsHdrSize d p a
937 return ((code,AddrRep):rest)
939 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
940 -> do rest <- pargs (d + addr_sizeW) az
941 code <- parg_ArrayishRep arrWordsHdrSize d p a
942 return ((code,AddrRep):rest)
944 -- Default case: push taggedly, but otherwise intact.
946 -> do (code_a, sz_a) <- pushAtom d p a
947 rest <- pargs (d+sz_a) az
948 return ((code_a, atomPrimRep a) : rest)
950 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
951 -- the stack but then advance it over the headers, so as to
952 -- point to the payload.
953 parg_ArrayishRep hdrSize d p a
954 = do (push_fo, _) <- pushAtom d p a
955 -- The ptr points at the header. Advance it over the
956 -- header and then pretend this is an Addr#.
957 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
960 code_n_reps <- pargs d0 args_r_to_l
962 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
964 push_args = concatOL pushs_arg
965 d_after_args = d0 + sum (map primRepSizeW a_reps_pushed_r_to_l)
967 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
968 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
970 = reverse (tail a_reps_pushed_r_to_l)
972 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
973 -- push_args is the code to do that.
974 -- d_after_args is the stack depth once the args are on.
976 -- Get the result rep.
977 (returns_void, r_rep)
978 = case maybe_getCCallReturnRep (idType fn) of
979 Nothing -> (True, VoidRep)
980 Just rr -> (False, rr)
982 Because the Haskell stack grows down, the a_reps refer to
983 lowest to highest addresses in that order. The args for the call
984 are on the stack. Now push an unboxed Addr# indicating
985 the C function to call. Then push a dummy placeholder for the
986 result. Finally, emit a CCALL insn with an offset pointing to the
987 Addr# just pushed, and a literal field holding the mallocville
988 address of the piece of marshalling code we generate.
989 So, just prior to the CCALL insn, the stack looks like this
990 (growing down, as usual):
995 Addr# address_of_C_fn
996 <placeholder-for-result#> (must be an unboxed type)
998 The interpreter then calls the marshall code mentioned
999 in the CCALL insn, passing it (& <placeholder-for-result#>),
1000 that is, the addr of the topmost word in the stack.
1001 When this returns, the placeholder will have been
1002 filled in. The placeholder is slid down to the sequel
1003 depth, and we RETURN.
1005 This arrangement makes it simple to do f-i-dynamic since the Addr#
1006 value is the first arg anyway.
1008 The marshalling code is generated specifically for this
1009 call site, and so knows exactly the (Haskell) stack
1010 offsets of the args, fn address and placeholder. It
1011 copies the args to the C stack, calls the stacked addr,
1012 and parks the result back in the placeholder. The interpreter
1013 calls it as a normal C call, assuming it has a signature
1014 void marshall_code ( StgWord* ptr_to_top_of_stack )
1016 -- resolve static address
1020 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
1022 -> do res <- ioToBc (lookupStaticPtr target)
1025 (is_static, static_target_addr) <- get_target_info
1028 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1029 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1030 | is_static = a_reps_pushed_RAW
1031 | otherwise = if null a_reps_pushed_RAW
1032 then panic "ByteCodeGen.generateCCall: dyn with no args"
1033 else tail a_reps_pushed_RAW
1036 (push_Addr, d_after_Addr)
1038 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1039 d_after_args + addr_sizeW)
1040 | otherwise -- is already on the stack
1041 = (nilOL, d_after_args)
1043 -- Push the return placeholder. For a call returning nothing,
1044 -- this is a VoidArg (tag).
1045 r_sizeW = primRepSizeW r_rep
1046 d_after_r = d_after_Addr + r_sizeW
1047 r_lit = mkDummyLiteral r_rep
1048 push_r = (if returns_void
1050 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1052 -- generate the marshalling code we're going to call
1055 arg1_offW = r_sizeW + addr_sizeW
1056 args_offW = map (arg1_offW +)
1057 (init (scanl (+) 0 (map primRepSizeW a_reps)))
1059 -- Offset of the next stack frame down the stack. The CCALL
1060 -- instruction needs to describe the chunk of stack containing
1061 -- the ccall args to the GC, so it needs to know how large it
1062 -- is. See comment in Interpreter.c with the CCALL instruction.
1063 stk_offset = d_after_r - s
1066 #if !defined(USE_LIBFFI)
1067 -- In the native case, we build marshalling code and attach the
1068 -- address of that to the CCALL instruction
1069 addr_of_marshaller <- ioToBc (mkMarshalCode cconv
1070 (r_offW, r_rep) addr_offW
1071 (zip args_offW a_reps))
1073 -- the only difference in libffi mode is that we prepare a cif
1074 -- describing the call type by calling libffi, and we attach the
1075 -- address of this to the CCALL instruction.
1076 token <- ioToBc $ prepForeignCall cconv a_reps r_rep
1077 let addr_of_marshaller = castPtrToFunPtr token
1080 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1083 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1085 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1086 `snocOL` RETURN_UBX (primRepToCgRep r_rep)
1088 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1091 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1094 -- Make a dummy literal, to be used as a placeholder for FFI return
1095 -- values on the stack.
1096 mkDummyLiteral :: PrimRep -> Literal
1100 WordRep -> MachWord 0
1101 AddrRep -> MachNullAddr
1102 DoubleRep -> MachDouble 0
1103 FloatRep -> MachFloat 0
1104 Int64Rep -> MachInt64 0
1105 Word64Rep -> MachWord64 0
1106 _ -> panic "mkDummyLiteral"
1110 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1111 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1114 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1116 -- Alternatively, for call-targets returning nothing, convert
1118 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1119 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1123 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
1124 maybe_getCCallReturnRep fn_ty
1125 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1127 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1129 = case splitTyConApp_maybe (repType r_ty) of
1130 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
1132 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
1133 || r_reps == [VoidRep] )
1134 && isUnboxedTupleTyCon r_tycon
1135 && case maybe_r_rep_to_go of
1137 Just r_rep -> r_rep /= PtrRep
1138 -- if it was, it would be impossible
1139 -- to create a valid return value
1140 -- placeholder on the stack
1141 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1144 --trace (showSDoc (ppr (a_reps, r_reps))) $
1145 if ok then maybe_r_rep_to_go else blargh
1147 -- Compile code which expects an unboxed Int on the top of stack,
1148 -- (call it i), and pushes the i'th closure in the supplied list
1149 -- as a consequence.
1150 implement_tagToId :: [Name] -> BcM BCInstrList
1151 implement_tagToId names
1152 = ASSERT( notNull names )
1153 do labels <- getLabelsBc (length names)
1154 label_fail <- getLabelBc
1155 label_exit <- getLabelBc
1156 let infos = zip4 labels (tail labels ++ [label_fail])
1158 steps = map (mkStep label_exit) infos
1159 return (concatOL steps
1161 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1163 mkStep l_exit (my_label, next_label, n, name_for_n)
1164 = toOL [LABEL my_label,
1165 TESTEQ_I n next_label,
1170 -- -----------------------------------------------------------------------------
1173 -- Push an atom onto the stack, returning suitable code & number of
1174 -- stack words used.
1176 -- The env p must map each variable to the highest- numbered stack
1177 -- slot for it. For example, if the stack has depth 4 and we
1178 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1179 -- the tag in stack[5], the stack will have depth 6, and p must map v
1180 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1181 -- depth 6 stack has valid words 0 .. 5.
1183 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1185 pushAtom d p (AnnApp f (_, AnnType _))
1186 = pushAtom d p (snd f)
1188 pushAtom d p (AnnNote note e)
1189 = pushAtom d p (snd e)
1191 pushAtom d p (AnnLam x e)
1193 = pushAtom d p (snd e)
1195 pushAtom d p (AnnVar v)
1197 | idCgRep v == VoidArg
1201 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1203 | Just primop <- isPrimOpId_maybe v
1204 = return (unitOL (PUSH_PRIMOP primop), 1)
1206 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1207 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1208 -- d - d_v the number of words between the TOS
1209 -- and the 1st slot of the object
1211 -- d - d_v - 1 the offset from the TOS of the 1st slot
1213 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1216 -- Having found the last slot, we proceed to copy the right number of
1217 -- slots on to the top of the stack.
1219 | otherwise -- v must be a global variable
1221 return (unitOL (PUSH_G (getName v)), sz)
1227 pushAtom d p (AnnLit lit)
1229 MachLabel fs _ -> code NonPtrArg
1230 MachWord w -> code NonPtrArg
1231 MachInt i -> code PtrArg
1232 MachFloat r -> code FloatArg
1233 MachDouble r -> code DoubleArg
1234 MachChar c -> code NonPtrArg
1235 MachStr s -> pushStr s
1238 = let size_host_words = cgRepSizeW rep
1239 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1243 = let getMallocvilleAddr
1245 FastString _ n _ fp _ ->
1246 -- we could grab the Ptr from the ForeignPtr,
1247 -- but then we have no way to control its lifetime.
1248 -- In reality it'll probably stay alive long enoungh
1249 -- by virtue of the global FastString table, but
1250 -- to be on the safe side we copy the string into
1251 -- a malloc'd area of memory.
1252 do ptr <- ioToBc (mallocBytes (n+1))
1255 withForeignPtr fp $ \p -> do
1256 memcpy ptr p (fromIntegral n)
1257 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1261 addr <- getMallocvilleAddr
1262 -- Get the addr on the stack, untaggedly
1263 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1265 pushAtom d p (AnnCast e _)
1266 = pushAtom d p (snd e)
1269 = pprPanic "ByteCodeGen.pushAtom"
1270 (pprCoreExpr (deAnnotate (undefined, other)))
1272 foreign import ccall unsafe "memcpy"
1273 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1276 -- -----------------------------------------------------------------------------
1277 -- Given a bunch of alts code and their discrs, do the donkey work
1278 -- of making a multiway branch using a switch tree.
1279 -- What a load of hassle!
1281 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1282 -- a hint; generates better code
1283 -- Nothing is always safe
1284 -> [(Discr, BCInstrList)]
1286 mkMultiBranch maybe_ncons raw_ways
1287 = let d_way = filter (isNoDiscr.fst) raw_ways
1289 (\w1 w2 -> leAlt (fst w1) (fst w2))
1290 (filter (not.isNoDiscr.fst) raw_ways)
1292 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1293 mkTree [] range_lo range_hi = return the_default
1295 mkTree [val] range_lo range_hi
1296 | range_lo `eqAlt` range_hi
1299 = do label_neq <- getLabelBc
1300 return (mkTestEQ (fst val) label_neq
1302 `appOL` unitOL (LABEL label_neq)
1303 `appOL` the_default))
1305 mkTree vals range_lo range_hi
1306 = let n = length vals `div` 2
1307 vals_lo = take n vals
1308 vals_hi = drop n vals
1309 v_mid = fst (head vals_hi)
1311 label_geq <- getLabelBc
1312 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1313 code_hi <- mkTree vals_hi v_mid range_hi
1314 return (mkTestLT v_mid label_geq
1316 `appOL` unitOL (LABEL label_geq)
1320 = case d_way of [] -> unitOL CASEFAIL
1323 -- None of these will be needed if there are no non-default alts
1324 (mkTestLT, mkTestEQ, init_lo, init_hi)
1326 = panic "mkMultiBranch: awesome foursome"
1328 = case fst (head notd_ways) of {
1329 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1330 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1333 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1334 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1337 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1338 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1341 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1342 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1344 DiscrP algMaxBound )
1347 (algMinBound, algMaxBound)
1348 = case maybe_ncons of
1349 Just n -> (0, n - 1)
1350 Nothing -> (minBound, maxBound)
1352 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1353 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1354 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1355 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1356 NoDiscr `eqAlt` NoDiscr = True
1359 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1360 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1361 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1362 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1363 NoDiscr `leAlt` NoDiscr = True
1366 isNoDiscr NoDiscr = True
1369 dec (DiscrI i) = DiscrI (i-1)
1370 dec (DiscrP i) = DiscrP (i-1)
1371 dec other = other -- not really right, but if you
1372 -- do cases on floating values, you'll get what you deserve
1374 -- same snotty comment applies to the following
1376 minD, maxD :: Double
1382 mkTree notd_ways init_lo init_hi
1385 -- -----------------------------------------------------------------------------
1386 -- Supporting junk for the compilation schemes
1388 -- Describes case alts
1396 instance Outputable Discr where
1397 ppr (DiscrI i) = int i
1398 ppr (DiscrF f) = text (show f)
1399 ppr (DiscrD d) = text (show d)
1400 ppr (DiscrP i) = int i
1401 ppr NoDiscr = text "DEF"
1404 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1405 lookupBCEnv_maybe = lookupFM
1407 idSizeW :: Id -> Int
1408 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1411 unboxedTupleException :: a
1412 unboxedTupleException
1415 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1416 " Possibly due to foreign import/export decls in source.\n"++
1417 " Workaround: use -fobject-code, or compile this module to .o separately."))
1420 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1423 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1424 -- The arguments are returned in *right-to-left* order
1425 splitApp (AnnApp (_,f) (_,a))
1426 | isTypeAtom a = splitApp f
1427 | otherwise = case splitApp f of
1428 (f', as) -> (f', a:as)
1429 splitApp (AnnNote n (_,e)) = splitApp e
1430 splitApp (AnnCast (_,e) _) = splitApp e
1431 splitApp e = (e, [])
1434 isTypeAtom :: AnnExpr' id ann -> Bool
1435 isTypeAtom (AnnType _) = True
1436 isTypeAtom _ = False
1438 isVoidArgAtom :: AnnExpr' id ann -> Bool
1439 isVoidArgAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1440 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1441 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1442 isVoidArgAtom _ = False
1444 atomPrimRep :: AnnExpr' Id ann -> PrimRep
1445 atomPrimRep (AnnVar v) = typePrimRep (idType v)
1446 atomPrimRep (AnnLit l) = typePrimRep (literalType l)
1447 atomPrimRep (AnnNote n b) = atomPrimRep (snd b)
1448 atomPrimRep (AnnApp f (_, AnnType _)) = atomPrimRep (snd f)
1449 atomPrimRep (AnnLam x e) | isTyVar x = atomPrimRep (snd e)
1450 atomPrimRep (AnnCast b _) = atomPrimRep (snd b)
1451 atomPrimRep other = pprPanic "atomPrimRep" (ppr (deAnnotate (undefined,other)))
1453 atomRep :: AnnExpr' Id ann -> CgRep
1454 atomRep e = primRepToCgRep (atomPrimRep e)
1456 isPtrAtom :: AnnExpr' Id ann -> Bool
1457 isPtrAtom e = atomRep e == PtrArg
1459 -- Let szsw be the sizes in words of some items pushed onto the stack,
1460 -- which has initial depth d'. Return the values which the stack environment
1461 -- should map these items to.
1462 mkStackOffsets :: Int -> [Int] -> [Int]
1463 mkStackOffsets original_depth szsw
1464 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1466 -- -----------------------------------------------------------------------------
1467 -- The bytecode generator's monad
1469 type BcPtr = Either ItblPtr (Ptr ())
1473 uniqSupply :: UniqSupply, -- for generating fresh variable names
1474 nextlabel :: Int, -- for generating local labels
1475 malloced :: [BcPtr], -- thunks malloced for current BCO
1476 -- Should be free()d when it is GCd
1477 breakArray :: BreakArray -- array of breakpoint flags
1480 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1482 ioToBc :: IO a -> BcM a
1483 ioToBc io = BcM $ \st -> do
1487 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1488 runBc us modBreaks (BcM m)
1489 = m (BcM_State us 0 [] breakArray)
1491 breakArray = modBreaks_flags modBreaks
1493 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1494 thenBc (BcM expr) cont = BcM $ \st0 -> do
1495 (st1, q) <- expr st0
1500 thenBc_ :: BcM a -> BcM b -> BcM b
1501 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1502 (st1, q) <- expr st0
1503 (st2, r) <- cont st1
1506 returnBc :: a -> BcM a
1507 returnBc result = BcM $ \st -> (return (st, result))
1509 instance Monad BcM where
1514 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1516 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1518 recordMallocBc :: Ptr a -> BcM ()
1520 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1522 recordItblMallocBc :: ItblPtr -> BcM ()
1523 recordItblMallocBc a
1524 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1526 getLabelBc :: BcM Int
1528 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1530 getLabelsBc :: Int -> BcM [Int]
1532 = BcM $ \st -> let ctr = nextlabel st
1533 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1535 getBreakArray :: BcM BreakArray
1536 getBreakArray = BcM $ \st -> return (st, breakArray st)
1538 newUnique :: BcM Unique
1540 \st -> case splitUniqSupply (uniqSupply st) of
1541 (us1, us2) -> let newState = st { uniqSupply = us2 }
1542 in return (newState, uniqFromSupply us1)
1544 newId :: Type -> BcM Id
1547 return $ mkSysLocal tickFS uniq ty
1549 tickFS = FSLIT("ticked")