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)
430 where mkAlloc sz 0 = ALLOC_AP sz
431 mkAlloc sz arity = ALLOC_PAP arity sz
433 compile_bind d' fvs x rhs size arity off = do
434 bco <- schemeR fvs (x,rhs)
435 build_thunk d' fvs size bco off arity
438 [ compile_bind d' fvs x rhs size arity n
439 | (fvs, x, rhs, size, arity, n) <-
440 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
443 body_code <- schemeE d' s p' body
444 thunk_codes <- sequence compile_binds
445 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
447 -- introduce a let binding for a ticked case expression. This rule
448 -- *should* only fire when the expression was not already let-bound
449 -- (the code gen for let bindings should take care of that). Todo: we
450 -- call exprFreeVars on a deAnnotated expression, this may not be the
451 -- best way to calculate the free vars but it seemed like the least
452 -- intrusive thing to do
453 schemeE d s p exp@(AnnCase {})
454 | Just (tickInfo,rhs) <- isTickedExp' exp
455 = if isUnLiftedType ty
456 then schemeE d s p (snd rhs)
459 -- Todo: is emptyVarSet correct on the next line?
460 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
462 where exp' = deAnnotate' exp
463 fvs = exprFreeVars exp'
466 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
467 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
469 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
471 -- case .... of a { DEFAULT -> ... }
472 -- becuse the return convention for both are identical.
474 -- Note that it does not matter losing the void-rep thing from the
475 -- envt (it won't be bound now) because we never look such things up.
477 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
478 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
480 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
481 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
482 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
484 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
485 | isUnboxedTupleCon dc
486 -- Similarly, convert
487 -- case .... of x { (# a #) -> ... }
489 -- case .... of a { DEFAULT -> ... }
490 = --trace "automagic mashing of case alts (# a #)" $
491 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
493 schemeE d s p (AnnCase scrut bndr _ alts)
494 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
496 schemeE d s p (AnnNote note (_, body))
499 schemeE d s p (AnnCast (_, body) _)
503 = pprPanic "ByteCodeGen.schemeE: unhandled case"
504 (pprCoreExpr (deAnnotate' other))
510 A ticked expression looks like this:
512 case tick<n> var1 ... varN of DEFAULT -> e
514 (*) <n> is the number of the tick, which is unique within a module
515 (*) var1 ... varN are the local variables in scope at the tick site
517 If we find a ticked expression we return:
519 Just ((n, [var1 ... varN]), e)
521 otherwise we return Nothing.
523 The idea is that the "case tick<n> ..." is really just an annotation on
524 the code. When we find such a thing, we pull out the useful information,
525 and then compile the code as if it was just the expression "e".
529 isTickedExp :: AnnExpr Id a -> Maybe (TickInfo, AnnExpr Id a)
530 isTickedExp (annot, expr) = isTickedExp' expr
532 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
533 isTickedExp' (AnnCase scrut _bndr _type alts)
534 | Just tickInfo <- isTickedScrut scrut,
535 [(DEFAULT, _bndr, rhs)] <- alts
536 = Just (tickInfo, rhs)
538 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
541 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
542 = Just $ TickInfo { tickInfo_number = tickNumber
543 , tickInfo_module = modName
544 , tickInfo_locals = idsOfArgs args
546 | otherwise = Nothing
548 (f, args) = collectArgs $ deAnnotate expr
549 idsOfArgs :: [Expr Id] -> [Id]
550 idsOfArgs = catMaybes . map exprId
551 exprId :: Expr Id -> Maybe Id
552 exprId (Var id) = Just id
553 exprId other = Nothing
555 isTickedExp' other = Nothing
557 -- Compile code to do a tail call. Specifically, push the fn,
558 -- slide the on-stack app back down to the sequel depth,
559 -- and enter. Four cases:
562 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
563 -- The int will be on the stack. Generate a code sequence
564 -- to convert it to the relevant constructor, SLIDE and ENTER.
566 -- 1. The fn denotes a ccall. Defer to generateCCall.
568 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
569 -- it simply as b -- since the representations are identical
570 -- (the VoidArg takes up zero stack space). Also, spot
571 -- (# b #) and treat it as b.
573 -- 3. Application of a constructor, by defn saturated.
574 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
575 -- then the ptrs, and then do PACK and RETURN.
577 -- 4. Otherwise, it must be a function call. Push the args
578 -- right to left, SLIDE and ENTER.
580 schemeT :: Int -- Stack depth
581 -> Sequel -- Sequel depth
582 -> BCEnv -- stack env
583 -> AnnExpr' Id VarSet
588 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
589 -- = panic "schemeT ?!?!"
591 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
595 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
596 = do (push, arg_words) <- pushAtom d p arg
597 tagToId_sequence <- implement_tagToId constr_names
598 return (push `appOL` tagToId_sequence
599 `appOL` mkSLIDE 1 (d+arg_words-s)
603 | Just (CCall ccall_spec) <- isFCallId_maybe fn
604 = generateCCall d s p ccall_spec fn args_r_to_l
606 -- Case 2: Constructor application
607 | Just con <- maybe_saturated_dcon,
608 isUnboxedTupleCon con
609 = case args_r_to_l of
610 [arg1,arg2] | isVoidArgAtom arg1 ->
611 unboxedTupleReturn d s p arg2
612 [arg1,arg2] | isVoidArgAtom arg2 ->
613 unboxedTupleReturn d s p arg1
614 _other -> unboxedTupleException
616 -- Case 3: Ordinary data constructor
617 | Just con <- maybe_saturated_dcon
618 = do alloc_con <- mkConAppCode d s p con args_r_to_l
619 return (alloc_con `appOL`
620 mkSLIDE 1 (d - s) `snocOL`
623 -- Case 4: Tail call of function
625 = doTailCall d s p fn args_r_to_l
628 -- Detect and extract relevant info for the tagToEnum kludge.
629 maybe_is_tagToEnum_call
630 = let extract_constr_Names ty
631 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
633 = map (getName . dataConWorkId) (tyConDataCons tyc)
634 -- NOTE: use the worker name, not the source name of
635 -- the DataCon. See DataCon.lhs for details.
637 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
640 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
641 -> case isPrimOpId_maybe v of
642 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
646 -- Extract the args (R->L) and fn
647 -- The function will necessarily be a variable,
648 -- because we are compiling a tail call
649 (AnnVar fn, args_r_to_l) = splitApp app
651 -- Only consider this to be a constructor application iff it is
652 -- saturated. Otherwise, we'll call the constructor wrapper.
653 n_args = length args_r_to_l
655 = case isDataConWorkId_maybe fn of
656 Just con | dataConRepArity con == n_args -> Just con
659 -- -----------------------------------------------------------------------------
660 -- Generate code to build a constructor application,
661 -- leaving it on top of the stack
663 mkConAppCode :: Int -> Sequel -> BCEnv
664 -> DataCon -- The data constructor
665 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
668 mkConAppCode orig_d s p con [] -- Nullary constructor
669 = ASSERT( isNullaryRepDataCon con )
670 return (unitOL (PUSH_G (getName (dataConWorkId con))))
671 -- Instead of doing a PACK, which would allocate a fresh
672 -- copy of this constructor, use the single shared version.
674 mkConAppCode orig_d s p con args_r_to_l
675 = ASSERT( dataConRepArity con == length args_r_to_l )
676 do_pushery orig_d (non_ptr_args ++ ptr_args)
678 -- The args are already in reverse order, which is the way PACK
679 -- expects them to be. We must push the non-ptrs after the ptrs.
680 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
682 do_pushery d (arg:args)
683 = do (push, arg_words) <- pushAtom d p arg
684 more_push_code <- do_pushery (d+arg_words) args
685 return (push `appOL` more_push_code)
687 = return (unitOL (PACK con n_arg_words))
689 n_arg_words = d - orig_d
692 -- -----------------------------------------------------------------------------
693 -- Returning an unboxed tuple with one non-void component (the only
694 -- case we can handle).
696 -- Remember, we don't want to *evaluate* the component that is being
697 -- returned, even if it is a pointed type. We always just return.
700 :: Int -> Sequel -> BCEnv
701 -> AnnExpr' Id VarSet -> BcM BCInstrList
702 unboxedTupleReturn d s p arg = do
703 (push, sz) <- pushAtom d p arg
705 mkSLIDE sz (d-s) `snocOL`
706 RETURN_UBX (atomRep arg))
708 -- -----------------------------------------------------------------------------
709 -- Generate code for a tail-call
712 :: Int -> Sequel -> BCEnv
713 -> Id -> [AnnExpr' Id VarSet]
715 doTailCall init_d s p fn args
716 = do_pushes init_d args (map atomRep args)
718 do_pushes d [] reps = do
719 ASSERT( null reps ) return ()
720 (push_fn, sz) <- pushAtom d p (AnnVar fn)
721 ASSERT( sz == 1 ) return ()
722 return (push_fn `appOL` (
723 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
725 do_pushes d args reps = do
726 let (push_apply, n, rest_of_reps) = findPushSeq reps
727 (these_args, rest_of_args) = splitAt n args
728 (next_d, push_code) <- push_seq d these_args
729 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
730 -- ^^^ for the PUSH_APPLY_ instruction
731 return (push_code `appOL` (push_apply `consOL` instrs))
733 push_seq d [] = return (d, nilOL)
734 push_seq d (arg:args) = do
735 (push_code, sz) <- pushAtom d p arg
736 (final_d, more_push_code) <- push_seq (d+sz) args
737 return (final_d, push_code `appOL` more_push_code)
739 -- v. similar to CgStackery.findMatch, ToDo: merge
740 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
741 = (PUSH_APPLY_PPPPPP, 6, rest)
742 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
743 = (PUSH_APPLY_PPPPP, 5, rest)
744 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
745 = (PUSH_APPLY_PPPP, 4, rest)
746 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
747 = (PUSH_APPLY_PPP, 3, rest)
748 findPushSeq (PtrArg: PtrArg: rest)
749 = (PUSH_APPLY_PP, 2, rest)
750 findPushSeq (PtrArg: rest)
751 = (PUSH_APPLY_P, 1, rest)
752 findPushSeq (VoidArg: rest)
753 = (PUSH_APPLY_V, 1, rest)
754 findPushSeq (NonPtrArg: rest)
755 = (PUSH_APPLY_N, 1, rest)
756 findPushSeq (FloatArg: rest)
757 = (PUSH_APPLY_F, 1, rest)
758 findPushSeq (DoubleArg: rest)
759 = (PUSH_APPLY_D, 1, rest)
760 findPushSeq (LongArg: rest)
761 = (PUSH_APPLY_L, 1, rest)
763 = panic "ByteCodeGen.findPushSeq"
765 -- -----------------------------------------------------------------------------
768 doCase :: Int -> Sequel -> BCEnv
769 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
770 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
772 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
774 -- Top of stack is the return itbl, as usual.
775 -- underneath it is the pointer to the alt_code BCO.
776 -- When an alt is entered, it assumes the returned value is
777 -- on top of the itbl.
780 -- An unlifted value gets an extra info table pushed on top
781 -- when it is returned.
782 unlifted_itbl_sizeW | isAlgCase = 0
785 -- depth of stack after the return value has been pushed
786 d_bndr = d + ret_frame_sizeW + idSizeW bndr
788 -- depth of stack after the extra info table for an unboxed return
789 -- has been pushed, if any. This is the stack depth at the
791 d_alts = d_bndr + unlifted_itbl_sizeW
793 -- Env in which to compile the alts, not including
794 -- any vars bound by the alts themselves
795 p_alts = addToFM p bndr (d_bndr - 1)
797 bndr_ty = idType bndr
798 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
800 -- given an alt, return a discr and code for it.
801 codeAlt alt@(DEFAULT, _, (_,rhs))
802 = do rhs_code <- schemeE d_alts s p_alts rhs
803 return (NoDiscr, rhs_code)
805 codeAlt alt@(discr, bndrs, (_,rhs))
806 -- primitive or nullary constructor alt: no need to UNPACK
807 | null real_bndrs = do
808 rhs_code <- schemeE d_alts s p_alts rhs
809 return (my_discr alt, rhs_code)
810 -- algebraic alt with some binders
811 | ASSERT(isAlgCase) otherwise =
813 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
814 ptr_sizes = map idSizeW ptrs
815 nptrs_sizes = map idSizeW nptrs
816 bind_sizes = ptr_sizes ++ nptrs_sizes
817 size = sum ptr_sizes + sum nptrs_sizes
818 -- the UNPACK instruction unpacks in reverse order...
819 p' = addListToFM p_alts
820 (zip (reverse (ptrs ++ nptrs))
821 (mkStackOffsets d_alts (reverse bind_sizes)))
823 rhs_code <- schemeE (d_alts+size) s p' rhs
824 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
826 real_bndrs = filter (not.isTyVar) bndrs
828 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
829 my_discr (DataAlt dc, binds, rhs)
830 | isUnboxedTupleCon dc
831 = unboxedTupleException
833 = DiscrP (dataConTag dc - fIRST_TAG)
834 my_discr (LitAlt l, binds, rhs)
835 = case l of MachInt i -> DiscrI (fromInteger i)
836 MachFloat r -> DiscrF (fromRational r)
837 MachDouble r -> DiscrD (fromRational r)
838 MachChar i -> DiscrI (ord i)
839 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
842 | not isAlgCase = Nothing
844 = case [dc | (DataAlt dc, _, _) <- alts] of
846 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
848 -- the bitmap is relative to stack depth d, i.e. before the
849 -- BCO, info table and return value are pushed on.
850 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
851 -- except that here we build the bitmap from the known bindings of
852 -- things that are pointers, whereas in CgBindery the code builds the
853 -- bitmap from the free slots and unboxed bindings.
856 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
857 -- The bitmap must cover the portion of the stack up to the sequel only.
858 -- Previously we were building a bitmap for the whole depth (d), but we
859 -- really want a bitmap up to depth (d-s). This affects compilation of
860 -- case-of-case expressions, which is the only time we can be compiling a
861 -- case expression with s /= 0.
863 bitmap = intsToReverseBitmap bitmap_size{-size-}
864 (sortLe (<=) (filter (< bitmap_size) rel_slots))
867 rel_slots = concat (map spread binds)
869 | isFollowableArg (idCgRep id) = [ rel_offset ]
871 where rel_offset = d - offset - 1
874 alt_stuff <- mapM codeAlt alts
875 alt_final <- mkMultiBranch maybe_ncons alt_stuff
878 alt_bco_name = getName bndr
879 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
880 0{-no arity-} bitmap_size bitmap True{-is alts-}
882 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
883 -- "\n bitmap = " ++ show bitmap) $ do
884 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
885 alt_bco' <- emitBc alt_bco
887 | isAlgCase = PUSH_ALTS alt_bco'
888 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
889 return (push_alts `consOL` scrut_code)
892 -- -----------------------------------------------------------------------------
893 -- Deal with a CCall.
895 -- Taggedly push the args onto the stack R->L,
896 -- deferencing ForeignObj#s and adjusting addrs to point to
897 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
898 -- (machine) code for the ccall, and create bytecodes to call that and
899 -- then return in the right way.
901 generateCCall :: Int -> Sequel -- stack and sequel depths
903 -> CCallSpec -- where to call
904 -> Id -- of target, for type info
905 -> [AnnExpr' Id VarSet] -- args (atoms)
908 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
911 addr_sizeW = cgRepSizeW NonPtrArg
913 -- Get the args on the stack, with tags and suitably
914 -- dereferenced for the CCall. For each arg, return the
915 -- depth to the first word of the bits for that arg, and the
916 -- CgRep of what was actually pushed.
918 pargs d [] = return []
920 = let arg_ty = repType (exprType (deAnnotate' a))
922 in case splitTyConApp_maybe arg_ty of
923 -- Don't push the FO; instead push the Addr# it
926 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
927 -> do rest <- pargs (d + addr_sizeW) az
928 code <- parg_ArrayishRep arrPtrsHdrSize d p a
929 return ((code,NonPtrArg):rest)
931 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
932 -> do rest <- pargs (d + addr_sizeW) az
933 code <- parg_ArrayishRep arrWordsHdrSize d p a
934 return ((code,NonPtrArg):rest)
936 -- Default case: push taggedly, but otherwise intact.
938 -> do (code_a, sz_a) <- pushAtom d p a
939 rest <- pargs (d+sz_a) az
940 return ((code_a, atomRep a) : rest)
942 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
943 -- the stack but then advance it over the headers, so as to
944 -- point to the payload.
945 parg_ArrayishRep hdrSize d p a
946 = do (push_fo, _) <- pushAtom d p a
947 -- The ptr points at the header. Advance it over the
948 -- header and then pretend this is an Addr#.
949 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
952 code_n_reps <- pargs d0 args_r_to_l
954 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
956 push_args = concatOL pushs_arg
957 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
959 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
960 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
962 = reverse (tail a_reps_pushed_r_to_l)
964 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
965 -- push_args is the code to do that.
966 -- d_after_args is the stack depth once the args are on.
968 -- Get the result rep.
969 (returns_void, r_rep)
970 = case maybe_getCCallReturnRep (idType fn) of
971 Nothing -> (True, VoidArg)
972 Just rr -> (False, rr)
974 Because the Haskell stack grows down, the a_reps refer to
975 lowest to highest addresses in that order. The args for the call
976 are on the stack. Now push an unboxed Addr# indicating
977 the C function to call. Then push a dummy placeholder for the
978 result. Finally, emit a CCALL insn with an offset pointing to the
979 Addr# just pushed, and a literal field holding the mallocville
980 address of the piece of marshalling code we generate.
981 So, just prior to the CCALL insn, the stack looks like this
982 (growing down, as usual):
987 Addr# address_of_C_fn
988 <placeholder-for-result#> (must be an unboxed type)
990 The interpreter then calls the marshall code mentioned
991 in the CCALL insn, passing it (& <placeholder-for-result#>),
992 that is, the addr of the topmost word in the stack.
993 When this returns, the placeholder will have been
994 filled in. The placeholder is slid down to the sequel
995 depth, and we RETURN.
997 This arrangement makes it simple to do f-i-dynamic since the Addr#
998 value is the first arg anyway.
1000 The marshalling code is generated specifically for this
1001 call site, and so knows exactly the (Haskell) stack
1002 offsets of the args, fn address and placeholder. It
1003 copies the args to the C stack, calls the stacked addr,
1004 and parks the result back in the placeholder. The interpreter
1005 calls it as a normal C call, assuming it has a signature
1006 void marshall_code ( StgWord* ptr_to_top_of_stack )
1008 -- resolve static address
1012 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
1014 -> do res <- ioToBc (lookupStaticPtr target)
1017 (is_static, static_target_addr) <- get_target_info
1020 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1021 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1022 | is_static = a_reps_pushed_RAW
1023 | otherwise = if null a_reps_pushed_RAW
1024 then panic "ByteCodeGen.generateCCall: dyn with no args"
1025 else tail a_reps_pushed_RAW
1028 (push_Addr, d_after_Addr)
1030 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1031 d_after_args + addr_sizeW)
1032 | otherwise -- is already on the stack
1033 = (nilOL, d_after_args)
1035 -- Push the return placeholder. For a call returning nothing,
1036 -- this is a VoidArg (tag).
1037 r_sizeW = cgRepSizeW r_rep
1038 d_after_r = d_after_Addr + r_sizeW
1039 r_lit = mkDummyLiteral r_rep
1040 push_r = (if returns_void
1042 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1044 -- generate the marshalling code we're going to call
1047 arg1_offW = r_sizeW + addr_sizeW
1048 args_offW = map (arg1_offW +)
1049 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
1051 addr_of_marshaller <- ioToBc (mkMarshalCode cconv
1052 (r_offW, r_rep) addr_offW
1053 (zip args_offW a_reps))
1054 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1056 -- Offset of the next stack frame down the stack. The CCALL
1057 -- instruction needs to describe the chunk of stack containing
1058 -- the ccall args to the GC, so it needs to know how large it
1059 -- is. See comment in Interpreter.c with the CCALL instruction.
1060 stk_offset = d_after_r - s
1063 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1065 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1066 `snocOL` RETURN_UBX r_rep
1068 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1071 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1075 -- Make a dummy literal, to be used as a placeholder for FFI return
1076 -- values on the stack.
1077 mkDummyLiteral :: CgRep -> Literal
1080 NonPtrArg -> MachWord 0
1081 DoubleArg -> MachDouble 0
1082 FloatArg -> MachFloat 0
1083 LongArg -> MachWord64 0
1084 _ -> moan64 "mkDummyLiteral" (ppr pr)
1088 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1089 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1092 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1094 -- Alternatively, for call-targets returning nothing, convert
1096 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1097 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1101 maybe_getCCallReturnRep :: Type -> Maybe CgRep
1102 maybe_getCCallReturnRep fn_ty
1103 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1105 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1107 = case splitTyConApp_maybe (repType r_ty) of
1108 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
1110 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
1111 || r_reps == [VoidArg] )
1112 && isUnboxedTupleTyCon r_tycon
1113 && case maybe_r_rep_to_go of
1115 Just r_rep -> r_rep /= PtrArg
1116 -- if it was, it would be impossible
1117 -- to create a valid return value
1118 -- placeholder on the stack
1119 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1122 --trace (showSDoc (ppr (a_reps, r_reps))) $
1123 if ok then maybe_r_rep_to_go else blargh
1125 -- Compile code which expects an unboxed Int on the top of stack,
1126 -- (call it i), and pushes the i'th closure in the supplied list
1127 -- as a consequence.
1128 implement_tagToId :: [Name] -> BcM BCInstrList
1129 implement_tagToId names
1130 = ASSERT( notNull names )
1131 do labels <- getLabelsBc (length names)
1132 label_fail <- getLabelBc
1133 label_exit <- getLabelBc
1134 let infos = zip4 labels (tail labels ++ [label_fail])
1136 steps = map (mkStep label_exit) infos
1137 return (concatOL steps
1139 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1141 mkStep l_exit (my_label, next_label, n, name_for_n)
1142 = toOL [LABEL my_label,
1143 TESTEQ_I n next_label,
1148 -- -----------------------------------------------------------------------------
1151 -- Push an atom onto the stack, returning suitable code & number of
1152 -- stack words used.
1154 -- The env p must map each variable to the highest- numbered stack
1155 -- slot for it. For example, if the stack has depth 4 and we
1156 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1157 -- the tag in stack[5], the stack will have depth 6, and p must map v
1158 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1159 -- depth 6 stack has valid words 0 .. 5.
1161 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1163 pushAtom d p (AnnApp f (_, AnnType _))
1164 = pushAtom d p (snd f)
1166 pushAtom d p (AnnNote note e)
1167 = pushAtom d p (snd e)
1169 pushAtom d p (AnnLam x e)
1171 = pushAtom d p (snd e)
1173 pushAtom d p (AnnVar v)
1175 | idCgRep v == VoidArg
1179 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1181 | Just primop <- isPrimOpId_maybe v
1182 = return (unitOL (PUSH_PRIMOP primop), 1)
1184 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1185 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1186 -- d - d_v the number of words between the TOS
1187 -- and the 1st slot of the object
1189 -- d - d_v - 1 the offset from the TOS of the 1st slot
1191 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1194 -- Having found the last slot, we proceed to copy the right number of
1195 -- slots on to the top of the stack.
1197 | otherwise -- v must be a global variable
1199 return (unitOL (PUSH_G (getName v)), sz)
1205 pushAtom d p (AnnLit lit)
1207 MachLabel fs _ -> code NonPtrArg
1208 MachWord w -> code NonPtrArg
1209 MachInt i -> code PtrArg
1210 MachFloat r -> code FloatArg
1211 MachDouble r -> code DoubleArg
1212 MachChar c -> code NonPtrArg
1213 MachStr s -> pushStr s
1216 = let size_host_words = cgRepSizeW rep
1217 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1221 = let getMallocvilleAddr
1223 FastString _ n _ fp _ ->
1224 -- we could grab the Ptr from the ForeignPtr,
1225 -- but then we have no way to control its lifetime.
1226 -- In reality it'll probably stay alive long enoungh
1227 -- by virtue of the global FastString table, but
1228 -- to be on the safe side we copy the string into
1229 -- a malloc'd area of memory.
1230 do ptr <- ioToBc (mallocBytes (n+1))
1233 withForeignPtr fp $ \p -> do
1234 memcpy ptr p (fromIntegral n)
1235 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1239 addr <- getMallocvilleAddr
1240 -- Get the addr on the stack, untaggedly
1241 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1243 pushAtom d p (AnnCast e _)
1244 = pushAtom d p (snd e)
1247 = pprPanic "ByteCodeGen.pushAtom"
1248 (pprCoreExpr (deAnnotate (undefined, other)))
1250 foreign import ccall unsafe "memcpy"
1251 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1254 -- -----------------------------------------------------------------------------
1255 -- Given a bunch of alts code and their discrs, do the donkey work
1256 -- of making a multiway branch using a switch tree.
1257 -- What a load of hassle!
1259 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1260 -- a hint; generates better code
1261 -- Nothing is always safe
1262 -> [(Discr, BCInstrList)]
1264 mkMultiBranch maybe_ncons raw_ways
1265 = let d_way = filter (isNoDiscr.fst) raw_ways
1267 (\w1 w2 -> leAlt (fst w1) (fst w2))
1268 (filter (not.isNoDiscr.fst) raw_ways)
1270 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1271 mkTree [] range_lo range_hi = return the_default
1273 mkTree [val] range_lo range_hi
1274 | range_lo `eqAlt` range_hi
1277 = do label_neq <- getLabelBc
1278 return (mkTestEQ (fst val) label_neq
1280 `appOL` unitOL (LABEL label_neq)
1281 `appOL` the_default))
1283 mkTree vals range_lo range_hi
1284 = let n = length vals `div` 2
1285 vals_lo = take n vals
1286 vals_hi = drop n vals
1287 v_mid = fst (head vals_hi)
1289 label_geq <- getLabelBc
1290 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1291 code_hi <- mkTree vals_hi v_mid range_hi
1292 return (mkTestLT v_mid label_geq
1294 `appOL` unitOL (LABEL label_geq)
1298 = case d_way of [] -> unitOL CASEFAIL
1301 -- None of these will be needed if there are no non-default alts
1302 (mkTestLT, mkTestEQ, init_lo, init_hi)
1304 = panic "mkMultiBranch: awesome foursome"
1306 = case fst (head notd_ways) of {
1307 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1308 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1311 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1312 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1315 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1316 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1319 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1320 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1322 DiscrP algMaxBound )
1325 (algMinBound, algMaxBound)
1326 = case maybe_ncons of
1327 Just n -> (0, n - 1)
1328 Nothing -> (minBound, maxBound)
1330 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1331 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1332 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1333 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1334 NoDiscr `eqAlt` NoDiscr = True
1337 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1338 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1339 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1340 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1341 NoDiscr `leAlt` NoDiscr = True
1344 isNoDiscr NoDiscr = True
1347 dec (DiscrI i) = DiscrI (i-1)
1348 dec (DiscrP i) = DiscrP (i-1)
1349 dec other = other -- not really right, but if you
1350 -- do cases on floating values, you'll get what you deserve
1352 -- same snotty comment applies to the following
1354 minD, maxD :: Double
1360 mkTree notd_ways init_lo init_hi
1363 -- -----------------------------------------------------------------------------
1364 -- Supporting junk for the compilation schemes
1366 -- Describes case alts
1374 instance Outputable Discr where
1375 ppr (DiscrI i) = int i
1376 ppr (DiscrF f) = text (show f)
1377 ppr (DiscrD d) = text (show d)
1378 ppr (DiscrP i) = int i
1379 ppr NoDiscr = text "DEF"
1382 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1383 lookupBCEnv_maybe = lookupFM
1385 idSizeW :: Id -> Int
1386 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1389 unboxedTupleException :: a
1390 unboxedTupleException
1393 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1394 " Possibly due to foreign import/export decls in source.\n"++
1395 " Workaround: use -fobject-code, or compile this module to .o separately."))
1398 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1401 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1402 -- The arguments are returned in *right-to-left* order
1403 splitApp (AnnApp (_,f) (_,a))
1404 | isTypeAtom a = splitApp f
1405 | otherwise = case splitApp f of
1406 (f', as) -> (f', a:as)
1407 splitApp (AnnNote n (_,e)) = splitApp e
1408 splitApp (AnnCast (_,e) _) = splitApp e
1409 splitApp e = (e, [])
1412 isTypeAtom :: AnnExpr' id ann -> Bool
1413 isTypeAtom (AnnType _) = True
1414 isTypeAtom _ = False
1416 isVoidArgAtom :: AnnExpr' id ann -> Bool
1417 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1418 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1419 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1420 isVoidArgAtom _ = False
1422 atomRep :: AnnExpr' Id ann -> CgRep
1423 atomRep (AnnVar v) = typeCgRep (idType v)
1424 atomRep (AnnLit l) = typeCgRep (literalType l)
1425 atomRep (AnnNote n b) = atomRep (snd b)
1426 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1427 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1428 atomRep (AnnCast b _) = atomRep (snd b)
1429 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1431 isPtrAtom :: AnnExpr' Id ann -> Bool
1432 isPtrAtom e = atomRep e == PtrArg
1434 -- Let szsw be the sizes in words of some items pushed onto the stack,
1435 -- which has initial depth d'. Return the values which the stack environment
1436 -- should map these items to.
1437 mkStackOffsets :: Int -> [Int] -> [Int]
1438 mkStackOffsets original_depth szsw
1439 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1441 -- -----------------------------------------------------------------------------
1442 -- The bytecode generator's monad
1444 type BcPtr = Either ItblPtr (Ptr ())
1448 uniqSupply :: UniqSupply, -- for generating fresh variable names
1449 nextlabel :: Int, -- for generating local labels
1450 malloced :: [BcPtr], -- thunks malloced for current BCO
1451 -- Should be free()d when it is GCd
1452 breakArray :: BreakArray -- array of breakpoint flags
1455 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1457 ioToBc :: IO a -> BcM a
1458 ioToBc io = BcM $ \st -> do
1462 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1463 runBc us modBreaks (BcM m)
1464 = m (BcM_State us 0 [] breakArray)
1466 breakArray = modBreaks_flags modBreaks
1468 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1469 thenBc (BcM expr) cont = BcM $ \st0 -> do
1470 (st1, q) <- expr st0
1475 thenBc_ :: BcM a -> BcM b -> BcM b
1476 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1477 (st1, q) <- expr st0
1478 (st2, r) <- cont st1
1481 returnBc :: a -> BcM a
1482 returnBc result = BcM $ \st -> (return (st, result))
1484 instance Monad BcM where
1489 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1491 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1493 recordMallocBc :: Ptr a -> BcM ()
1495 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1497 recordItblMallocBc :: ItblPtr -> BcM ()
1498 recordItblMallocBc a
1499 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1501 getLabelBc :: BcM Int
1503 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1505 getLabelsBc :: Int -> BcM [Int]
1507 = BcM $ \st -> let ctr = nextlabel st
1508 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1510 getBreakArray :: BcM BreakArray
1511 getBreakArray = BcM $ \st -> return (st, breakArray st)
1513 newUnique :: BcM Unique
1515 \st -> case splitUniqSupply (uniqSupply st) of
1516 (us1, us2) -> let newState = st { uniqSupply = us2 }
1517 in return (newState, uniqFromSupply us1)
1519 newId :: Type -> BcM Id
1522 return $ mkSysLocal FSLIT("ticked") uniq ty