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
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 | stack_overest >= iNTERP_STACK_CHECK_THRESH
186 = STKCHECK stack_overest : peep_d
188 = peep_d -- the supposedly common case
190 -- We assume that this sum doesn't wrap
191 stack_overest = sum (map bciStackUse peep_d)
193 -- Merge local pushes
194 peep_d = peep (fromOL instrs_ordlist)
196 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
197 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
198 peep (PUSH_L off1 : PUSH_L off2 : rest)
199 = PUSH_LL off1 (off2-1) : peep rest
205 argBits :: [CgRep] -> [Bool]
208 | isFollowableArg rep = False : argBits args
209 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
211 -- -----------------------------------------------------------------------------
214 -- Compile code for the right-hand side of a top-level binding
216 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
219 schemeTopBind (id, rhs)
220 | Just data_con <- isDataConWorkId_maybe id,
221 isNullaryRepDataCon data_con = do
222 -- Special case for the worker of a nullary data con.
223 -- It'll look like this: Nil = /\a -> Nil a
224 -- If we feed it into schemeR, we'll get
226 -- because mkConAppCode treats nullary constructor applications
227 -- by just re-using the single top-level definition. So
228 -- for the worker itself, we must allocate it directly.
229 -- ioToBc (putStrLn $ "top level BCO")
230 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
231 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
234 = schemeR [{- No free variables -}] (id, rhs)
237 -- -----------------------------------------------------------------------------
240 -- Compile code for a right-hand side, to give a BCO that,
241 -- when executed with the free variables and arguments on top of the stack,
242 -- will return with a pointer to the result on top of the stack, after
243 -- removing the free variables and arguments.
245 -- Park the resulting BCO in the monad. Also requires the
246 -- variable to which this value was bound, so as to give the
247 -- resulting BCO a name.
249 schemeR :: [Id] -- Free vars of the RHS, ordered as they
250 -- will appear in the thunk. Empty for
251 -- top-level things, which have no free vars.
252 -> (Id, AnnExpr Id VarSet)
253 -> BcM (ProtoBCO Name)
254 schemeR fvs (nm, rhs)
258 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
259 $$ pprCoreExpr (deAnnotate rhs)
265 = schemeR_wrk fvs nm rhs (collect [] rhs)
267 collect :: [Var] -> AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
268 collect xs (_, AnnNote note e) = collect xs e
269 collect xs (_, AnnCast e _) = collect xs e
270 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
271 collect xs (_, not_lambda) = (reverse xs, not_lambda)
273 schemeR_wrk :: [Id] -> Id -> AnnExpr Id VarSet -> ([Var], AnnExpr' Var VarSet) -> BcM (ProtoBCO Name)
274 schemeR_wrk fvs nm original_body (args, body)
276 all_args = reverse args ++ fvs
277 arity = length all_args
278 -- all_args are the args in reverse order. We're compiling a function
279 -- \fv1..fvn x1..xn -> e
280 -- i.e. the fvs come first
282 szsw_args = map idSizeW all_args
283 szw_args = sum szsw_args
284 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
286 -- make the arg bitmap
287 bits = argBits (reverse (map idCgRep all_args))
288 bitmap_size = length bits
289 bitmap = mkBitmap bits
291 body_code <- schemeER_wrk szw_args p_init body
293 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
294 arity bitmap_size bitmap False{-not alts-})
296 -- introduce break instructions for ticked expressions
297 schemeER_wrk :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
299 | Just (tickInfo, (_annot, newRhs)) <- isTickedExp' rhs = do
300 code <- schemeE d 0 p newRhs
302 let idOffSets = getVarOffSets d p tickInfo
303 let tickNumber = tickInfo_number tickInfo
304 let breakInfo = BreakInfo
305 { breakInfo_module = tickInfo_module tickInfo
306 , breakInfo_number = tickNumber
307 , breakInfo_vars = idOffSets
308 , breakInfo_resty = exprType (deAnnotate' newRhs)
310 let breakInstr = case arr of (BA arr#) -> BRK_FUN arr# tickNumber breakInfo
311 return $ breakInstr `consOL` code
312 | otherwise = schemeE d 0 p rhs
314 getVarOffSets :: Int -> BCEnv -> TickInfo -> [(Id, Int)]
315 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
317 getOffSet :: Int -> BCEnv -> Id -> Maybe (Id, Int)
319 = case lookupBCEnv_maybe env id of
321 Just offset -> Just (id, d - offset)
323 fvsToEnv :: BCEnv -> VarSet -> [Id]
324 -- Takes the free variables of a right-hand side, and
325 -- delivers an ordered list of the local variables that will
326 -- be captured in the thunk for the RHS
327 -- The BCEnv argument tells which variables are in the local
328 -- environment: these are the ones that should be captured
330 -- The code that constructs the thunk, and the code that executes
331 -- it, have to agree about this layout
332 fvsToEnv p fvs = [v | v <- varSetElems fvs,
333 isId v, -- Could be a type variable
336 -- -----------------------------------------------------------------------------
341 { tickInfo_number :: Int -- the (module) unique number of the tick
342 , tickInfo_module :: Module -- the origin of the ticked expression
343 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
346 instance Outputable TickInfo where
347 ppr info = text "TickInfo" <+>
348 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
349 ppr (tickInfo_locals info))
351 -- Compile code to apply the given expression to the remaining args
352 -- on the stack, returning a HNF.
353 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
355 -- Delegate tail-calls to schemeT.
356 schemeE d s p e@(AnnApp f a)
359 schemeE d s p e@(AnnVar v)
360 | not (isUnLiftedType v_type)
361 = -- Lifted-type thing; push it in the normal way
365 = do -- Returning an unlifted value.
366 -- Heave it on the stack, SLIDE, and RETURN.
367 (push, szw) <- pushAtom d p (AnnVar v)
368 return (push -- value onto stack
369 `appOL` mkSLIDE szw (d-s) -- clear to sequel
370 `snocOL` RETURN_UBX v_rep) -- go
373 v_rep = typeCgRep v_type
375 schemeE d s p (AnnLit literal)
376 = do (push, szw) <- pushAtom d p (AnnLit literal)
377 let l_rep = typeCgRep (literalType literal)
378 return (push -- value onto stack
379 `appOL` mkSLIDE szw (d-s) -- clear to sequel
380 `snocOL` RETURN_UBX l_rep) -- go
382 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
383 | (AnnVar v, args_r_to_l) <- splitApp rhs,
384 Just data_con <- isDataConWorkId_maybe v,
385 dataConRepArity data_con == length args_r_to_l
386 = do -- Special case for a non-recursive let whose RHS is a
387 -- saturatred constructor application.
388 -- Just allocate the constructor and carry on
389 alloc_code <- mkConAppCode d s p data_con args_r_to_l
390 body_code <- schemeE (d+1) s (addToFM p x d) body
391 return (alloc_code `appOL` body_code)
393 -- General case for let. Generates correct, if inefficient, code in
395 schemeE d s p (AnnLet binds (_,body))
396 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
397 AnnRec xs_n_rhss -> unzip xs_n_rhss
400 fvss = map (fvsToEnv p' . fst) rhss
402 -- Sizes of free vars
403 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
405 -- the arity of each rhs
406 arities = map (length . fst . collect []) rhss
408 -- This p', d' defn is safe because all the items being pushed
409 -- are ptrs, so all have size 1. d' and p' reflect the stack
410 -- after the closures have been allocated in the heap (but not
411 -- filled in), and pointers to them parked on the stack.
412 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
414 zipE = zipEqual "schemeE"
416 -- ToDo: don't build thunks for things with no free variables
417 build_thunk dd [] size bco off arity
418 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
420 mkap | arity == 0 = MKAP
422 build_thunk dd (fv:fvs) size bco off arity = do
423 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
424 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
425 return (push_code `appOL` more_push_code)
427 alloc_code = toOL (zipWith mkAlloc sizes arities)
428 where mkAlloc sz 0 = ALLOC_AP sz
429 mkAlloc sz arity = ALLOC_PAP arity sz
431 compile_bind d' fvs x rhs size arity off = do
432 bco <- schemeR fvs (x,rhs)
433 build_thunk d' fvs size bco off arity
436 [ compile_bind d' fvs x rhs size arity n
437 | (fvs, x, rhs, size, arity, n) <-
438 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
441 body_code <- schemeE d' s p' body
442 thunk_codes <- sequence compile_binds
443 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
445 -- introduce a let binding for a ticked case expression. This rule
446 -- *should* only fire when the expression was not already let-bound
447 -- (the code gen for let bindings should take care of that). Todo: we
448 -- call exprFreeVars on a deAnnotated expression, this may not be the
449 -- best way to calculate the free vars but it seemed like the least
450 -- intrusive thing to do
451 schemeE d s p exp@(AnnCase {})
452 | Just (tickInfo,rhs) <- isTickedExp' exp
453 = if isUnLiftedType ty
454 then schemeE d s p (snd rhs)
457 -- Todo: is emptyVarSet correct on the next line?
458 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
460 where exp' = deAnnotate' exp
461 fvs = exprFreeVars exp'
464 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
465 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
467 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
469 -- case .... of a { DEFAULT -> ... }
470 -- becuse the return convention for both are identical.
472 -- Note that it does not matter losing the void-rep thing from the
473 -- envt (it won't be bound now) because we never look such things up.
475 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
476 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
478 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
479 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
480 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
482 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
483 | isUnboxedTupleCon dc
484 -- Similarly, convert
485 -- case .... of x { (# a #) -> ... }
487 -- case .... of a { DEFAULT -> ... }
488 = --trace "automagic mashing of case alts (# a #)" $
489 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
491 schemeE d s p (AnnCase scrut bndr _ alts)
492 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
494 schemeE d s p (AnnNote note (_, body))
497 schemeE d s p (AnnCast (_, body) _)
501 = pprPanic "ByteCodeGen.schemeE: unhandled case"
502 (pprCoreExpr (deAnnotate' other))
508 A ticked expression looks like this:
510 case tick<n> var1 ... varN of DEFAULT -> e
512 (*) <n> is the number of the tick, which is unique within a module
513 (*) var1 ... varN are the local variables in scope at the tick site
515 If we find a ticked expression we return:
517 Just ((n, [var1 ... varN]), e)
519 otherwise we return Nothing.
521 The idea is that the "case tick<n> ..." is really just an annotation on
522 the code. When we find such a thing, we pull out the useful information,
523 and then compile the code as if it was just the expression "e".
527 isTickedExp :: AnnExpr Id a -> Maybe (TickInfo, AnnExpr Id a)
528 isTickedExp (annot, expr) = isTickedExp' expr
530 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
531 isTickedExp' (AnnCase scrut _bndr _type alts)
532 | Just tickInfo <- isTickedScrut scrut,
533 [(DEFAULT, _bndr, rhs)] <- alts
534 = Just (tickInfo, rhs)
536 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
539 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
540 = Just $ TickInfo { tickInfo_number = tickNumber
541 , tickInfo_module = modName
542 , tickInfo_locals = idsOfArgs args
544 | otherwise = Nothing
546 (f, args) = collectArgs $ deAnnotate expr
547 idsOfArgs :: [Expr Id] -> [Id]
548 idsOfArgs = catMaybes . map exprId
549 exprId :: Expr Id -> Maybe Id
550 exprId (Var id) = Just id
551 exprId other = Nothing
553 isTickedExp' other = Nothing
555 -- Compile code to do a tail call. Specifically, push the fn,
556 -- slide the on-stack app back down to the sequel depth,
557 -- and enter. Four cases:
560 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
561 -- The int will be on the stack. Generate a code sequence
562 -- to convert it to the relevant constructor, SLIDE and ENTER.
564 -- 1. The fn denotes a ccall. Defer to generateCCall.
566 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
567 -- it simply as b -- since the representations are identical
568 -- (the VoidArg takes up zero stack space). Also, spot
569 -- (# b #) and treat it as b.
571 -- 3. Application of a constructor, by defn saturated.
572 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
573 -- then the ptrs, and then do PACK and RETURN.
575 -- 4. Otherwise, it must be a function call. Push the args
576 -- right to left, SLIDE and ENTER.
578 schemeT :: Int -- Stack depth
579 -> Sequel -- Sequel depth
580 -> BCEnv -- stack env
581 -> AnnExpr' Id VarSet
586 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
587 -- = panic "schemeT ?!?!"
589 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
593 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
594 = do (push, arg_words) <- pushAtom d p arg
595 tagToId_sequence <- implement_tagToId constr_names
596 return (push `appOL` tagToId_sequence
597 `appOL` mkSLIDE 1 (d+arg_words-s)
601 | Just (CCall ccall_spec) <- isFCallId_maybe fn
602 = generateCCall d s p ccall_spec fn args_r_to_l
604 -- Case 2: Constructor application
605 | Just con <- maybe_saturated_dcon,
606 isUnboxedTupleCon con
607 = case args_r_to_l of
608 [arg1,arg2] | isVoidArgAtom arg1 ->
609 unboxedTupleReturn d s p arg2
610 [arg1,arg2] | isVoidArgAtom arg2 ->
611 unboxedTupleReturn d s p arg1
612 _other -> unboxedTupleException
614 -- Case 3: Ordinary data constructor
615 | Just con <- maybe_saturated_dcon
616 = do alloc_con <- mkConAppCode d s p con args_r_to_l
617 return (alloc_con `appOL`
618 mkSLIDE 1 (d - s) `snocOL`
621 -- Case 4: Tail call of function
623 = doTailCall d s p fn args_r_to_l
626 -- Detect and extract relevant info for the tagToEnum kludge.
627 maybe_is_tagToEnum_call
628 = let extract_constr_Names ty
629 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
631 = map (getName . dataConWorkId) (tyConDataCons tyc)
632 -- NOTE: use the worker name, not the source name of
633 -- the DataCon. See DataCon.lhs for details.
635 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
638 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
639 -> case isPrimOpId_maybe v of
640 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
644 -- Extract the args (R->L) and fn
645 -- The function will necessarily be a variable,
646 -- because we are compiling a tail call
647 (AnnVar fn, args_r_to_l) = splitApp app
649 -- Only consider this to be a constructor application iff it is
650 -- saturated. Otherwise, we'll call the constructor wrapper.
651 n_args = length args_r_to_l
653 = case isDataConWorkId_maybe fn of
654 Just con | dataConRepArity con == n_args -> Just con
657 -- -----------------------------------------------------------------------------
658 -- Generate code to build a constructor application,
659 -- leaving it on top of the stack
661 mkConAppCode :: Int -> Sequel -> BCEnv
662 -> DataCon -- The data constructor
663 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
666 mkConAppCode orig_d s p con [] -- Nullary constructor
667 = ASSERT( isNullaryRepDataCon con )
668 return (unitOL (PUSH_G (getName (dataConWorkId con))))
669 -- Instead of doing a PACK, which would allocate a fresh
670 -- copy of this constructor, use the single shared version.
672 mkConAppCode orig_d s p con args_r_to_l
673 = ASSERT( dataConRepArity con == length args_r_to_l )
674 do_pushery orig_d (non_ptr_args ++ ptr_args)
676 -- The args are already in reverse order, which is the way PACK
677 -- expects them to be. We must push the non-ptrs after the ptrs.
678 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
680 do_pushery d (arg:args)
681 = do (push, arg_words) <- pushAtom d p arg
682 more_push_code <- do_pushery (d+arg_words) args
683 return (push `appOL` more_push_code)
685 = return (unitOL (PACK con n_arg_words))
687 n_arg_words = d - orig_d
690 -- -----------------------------------------------------------------------------
691 -- Returning an unboxed tuple with one non-void component (the only
692 -- case we can handle).
694 -- Remember, we don't want to *evaluate* the component that is being
695 -- returned, even if it is a pointed type. We always just return.
698 :: Int -> Sequel -> BCEnv
699 -> AnnExpr' Id VarSet -> BcM BCInstrList
700 unboxedTupleReturn d s p arg = do
701 (push, sz) <- pushAtom d p arg
703 mkSLIDE sz (d-s) `snocOL`
704 RETURN_UBX (atomRep arg))
706 -- -----------------------------------------------------------------------------
707 -- Generate code for a tail-call
710 :: Int -> Sequel -> BCEnv
711 -> Id -> [AnnExpr' Id VarSet]
713 doTailCall init_d s p fn args
714 = do_pushes init_d args (map atomRep args)
716 do_pushes d [] reps = do
717 ASSERT( null reps ) return ()
718 (push_fn, sz) <- pushAtom d p (AnnVar fn)
719 ASSERT( sz == 1 ) return ()
720 return (push_fn `appOL` (
721 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
723 do_pushes d args reps = do
724 let (push_apply, n, rest_of_reps) = findPushSeq reps
725 (these_args, rest_of_args) = splitAt n args
726 (next_d, push_code) <- push_seq d these_args
727 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
728 -- ^^^ for the PUSH_APPLY_ instruction
729 return (push_code `appOL` (push_apply `consOL` instrs))
731 push_seq d [] = return (d, nilOL)
732 push_seq d (arg:args) = do
733 (push_code, sz) <- pushAtom d p arg
734 (final_d, more_push_code) <- push_seq (d+sz) args
735 return (final_d, push_code `appOL` more_push_code)
737 -- v. similar to CgStackery.findMatch, ToDo: merge
738 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
739 = (PUSH_APPLY_PPPPPP, 6, rest)
740 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
741 = (PUSH_APPLY_PPPPP, 5, rest)
742 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
743 = (PUSH_APPLY_PPPP, 4, rest)
744 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
745 = (PUSH_APPLY_PPP, 3, rest)
746 findPushSeq (PtrArg: PtrArg: rest)
747 = (PUSH_APPLY_PP, 2, rest)
748 findPushSeq (PtrArg: rest)
749 = (PUSH_APPLY_P, 1, rest)
750 findPushSeq (VoidArg: rest)
751 = (PUSH_APPLY_V, 1, rest)
752 findPushSeq (NonPtrArg: rest)
753 = (PUSH_APPLY_N, 1, rest)
754 findPushSeq (FloatArg: rest)
755 = (PUSH_APPLY_F, 1, rest)
756 findPushSeq (DoubleArg: rest)
757 = (PUSH_APPLY_D, 1, rest)
758 findPushSeq (LongArg: rest)
759 = (PUSH_APPLY_L, 1, rest)
761 = panic "ByteCodeGen.findPushSeq"
763 -- -----------------------------------------------------------------------------
766 doCase :: Int -> Sequel -> BCEnv
767 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
768 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
770 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
772 -- Top of stack is the return itbl, as usual.
773 -- underneath it is the pointer to the alt_code BCO.
774 -- When an alt is entered, it assumes the returned value is
775 -- on top of the itbl.
778 -- An unlifted value gets an extra info table pushed on top
779 -- when it is returned.
780 unlifted_itbl_sizeW | isAlgCase = 0
783 -- depth of stack after the return value has been pushed
784 d_bndr = d + ret_frame_sizeW + idSizeW bndr
786 -- depth of stack after the extra info table for an unboxed return
787 -- has been pushed, if any. This is the stack depth at the
789 d_alts = d_bndr + unlifted_itbl_sizeW
791 -- Env in which to compile the alts, not including
792 -- any vars bound by the alts themselves
793 p_alts = addToFM p bndr (d_bndr - 1)
795 bndr_ty = idType bndr
796 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
798 -- given an alt, return a discr and code for it.
799 codeAlt alt@(DEFAULT, _, (_,rhs))
800 = do rhs_code <- schemeE d_alts s p_alts rhs
801 return (NoDiscr, rhs_code)
803 codeAlt alt@(discr, bndrs, (_,rhs))
804 -- primitive or nullary constructor alt: no need to UNPACK
805 | null real_bndrs = do
806 rhs_code <- schemeE d_alts s p_alts rhs
807 return (my_discr alt, rhs_code)
808 -- algebraic alt with some binders
809 | ASSERT(isAlgCase) otherwise =
811 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
812 ptr_sizes = map idSizeW ptrs
813 nptrs_sizes = map idSizeW nptrs
814 bind_sizes = ptr_sizes ++ nptrs_sizes
815 size = sum ptr_sizes + sum nptrs_sizes
816 -- the UNPACK instruction unpacks in reverse order...
817 p' = addListToFM p_alts
818 (zip (reverse (ptrs ++ nptrs))
819 (mkStackOffsets d_alts (reverse bind_sizes)))
821 rhs_code <- schemeE (d_alts+size) s p' rhs
822 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
824 real_bndrs = filter (not.isTyVar) bndrs
826 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
827 my_discr (DataAlt dc, binds, rhs)
828 | isUnboxedTupleCon dc
829 = unboxedTupleException
831 = DiscrP (dataConTag dc - fIRST_TAG)
832 my_discr (LitAlt l, binds, rhs)
833 = case l of MachInt i -> DiscrI (fromInteger i)
834 MachFloat r -> DiscrF (fromRational r)
835 MachDouble r -> DiscrD (fromRational r)
836 MachChar i -> DiscrI (ord i)
837 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
840 | not isAlgCase = Nothing
842 = case [dc | (DataAlt dc, _, _) <- alts] of
844 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
846 -- the bitmap is relative to stack depth d, i.e. before the
847 -- BCO, info table and return value are pushed on.
848 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
849 -- except that here we build the bitmap from the known bindings of
850 -- things that are pointers, whereas in CgBindery the code builds the
851 -- bitmap from the free slots and unboxed bindings.
854 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
855 -- The bitmap must cover the portion of the stack up to the sequel only.
856 -- Previously we were building a bitmap for the whole depth (d), but we
857 -- really want a bitmap up to depth (d-s). This affects compilation of
858 -- case-of-case expressions, which is the only time we can be compiling a
859 -- case expression with s /= 0.
861 bitmap = intsToReverseBitmap bitmap_size{-size-}
862 (sortLe (<=) (filter (< bitmap_size) rel_slots))
865 rel_slots = concat (map spread binds)
867 | isFollowableArg (idCgRep id) = [ rel_offset ]
869 where rel_offset = d - offset - 1
872 alt_stuff <- mapM codeAlt alts
873 alt_final <- mkMultiBranch maybe_ncons alt_stuff
876 alt_bco_name = getName bndr
877 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
878 0{-no arity-} bitmap_size bitmap True{-is alts-}
880 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
881 -- "\n bitmap = " ++ show bitmap) $ do
882 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
883 alt_bco' <- emitBc alt_bco
885 | isAlgCase = PUSH_ALTS alt_bco'
886 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
887 return (push_alts `consOL` scrut_code)
890 -- -----------------------------------------------------------------------------
891 -- Deal with a CCall.
893 -- Taggedly push the args onto the stack R->L,
894 -- deferencing ForeignObj#s and adjusting addrs to point to
895 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
896 -- (machine) code for the ccall, and create bytecodes to call that and
897 -- then return in the right way.
899 generateCCall :: Int -> Sequel -- stack and sequel depths
901 -> CCallSpec -- where to call
902 -> Id -- of target, for type info
903 -> [AnnExpr' Id VarSet] -- args (atoms)
906 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
909 addr_sizeW = cgRepSizeW NonPtrArg
911 -- Get the args on the stack, with tags and suitably
912 -- dereferenced for the CCall. For each arg, return the
913 -- depth to the first word of the bits for that arg, and the
914 -- CgRep of what was actually pushed.
916 pargs d [] = return []
918 = let arg_ty = repType (exprType (deAnnotate' a))
920 in case splitTyConApp_maybe arg_ty of
921 -- Don't push the FO; instead push the Addr# it
924 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
925 -> do rest <- pargs (d + addr_sizeW) az
926 code <- parg_ArrayishRep arrPtrsHdrSize d p a
927 return ((code,NonPtrArg):rest)
929 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
930 -> do rest <- pargs (d + addr_sizeW) az
931 code <- parg_ArrayishRep arrWordsHdrSize d p a
932 return ((code,NonPtrArg):rest)
934 -- Default case: push taggedly, but otherwise intact.
936 -> do (code_a, sz_a) <- pushAtom d p a
937 rest <- pargs (d+sz_a) az
938 return ((code_a, atomRep a) : rest)
940 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
941 -- the stack but then advance it over the headers, so as to
942 -- point to the payload.
943 parg_ArrayishRep hdrSize d p a
944 = do (push_fo, _) <- pushAtom d p a
945 -- The ptr points at the header. Advance it over the
946 -- header and then pretend this is an Addr#.
947 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
950 code_n_reps <- pargs d0 args_r_to_l
952 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
954 push_args = concatOL pushs_arg
955 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
957 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
958 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
960 = reverse (tail a_reps_pushed_r_to_l)
962 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
963 -- push_args is the code to do that.
964 -- d_after_args is the stack depth once the args are on.
966 -- Get the result rep.
967 (returns_void, r_rep)
968 = case maybe_getCCallReturnRep (idType fn) of
969 Nothing -> (True, VoidArg)
970 Just rr -> (False, rr)
972 Because the Haskell stack grows down, the a_reps refer to
973 lowest to highest addresses in that order. The args for the call
974 are on the stack. Now push an unboxed Addr# indicating
975 the C function to call. Then push a dummy placeholder for the
976 result. Finally, emit a CCALL insn with an offset pointing to the
977 Addr# just pushed, and a literal field holding the mallocville
978 address of the piece of marshalling code we generate.
979 So, just prior to the CCALL insn, the stack looks like this
980 (growing down, as usual):
985 Addr# address_of_C_fn
986 <placeholder-for-result#> (must be an unboxed type)
988 The interpreter then calls the marshall code mentioned
989 in the CCALL insn, passing it (& <placeholder-for-result#>),
990 that is, the addr of the topmost word in the stack.
991 When this returns, the placeholder will have been
992 filled in. The placeholder is slid down to the sequel
993 depth, and we RETURN.
995 This arrangement makes it simple to do f-i-dynamic since the Addr#
996 value is the first arg anyway.
998 The marshalling code is generated specifically for this
999 call site, and so knows exactly the (Haskell) stack
1000 offsets of the args, fn address and placeholder. It
1001 copies the args to the C stack, calls the stacked addr,
1002 and parks the result back in the placeholder. The interpreter
1003 calls it as a normal C call, assuming it has a signature
1004 void marshall_code ( StgWord* ptr_to_top_of_stack )
1006 -- resolve static address
1010 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
1012 -> do res <- ioToBc (lookupStaticPtr target)
1015 (is_static, static_target_addr) <- get_target_info
1018 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1019 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1020 | is_static = a_reps_pushed_RAW
1021 | otherwise = if null a_reps_pushed_RAW
1022 then panic "ByteCodeGen.generateCCall: dyn with no args"
1023 else tail a_reps_pushed_RAW
1026 (push_Addr, d_after_Addr)
1028 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1029 d_after_args + addr_sizeW)
1030 | otherwise -- is already on the stack
1031 = (nilOL, d_after_args)
1033 -- Push the return placeholder. For a call returning nothing,
1034 -- this is a VoidArg (tag).
1035 r_sizeW = cgRepSizeW r_rep
1036 d_after_r = d_after_Addr + r_sizeW
1037 r_lit = mkDummyLiteral r_rep
1038 push_r = (if returns_void
1040 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1042 -- generate the marshalling code we're going to call
1045 arg1_offW = r_sizeW + addr_sizeW
1046 args_offW = map (arg1_offW +)
1047 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
1049 addr_of_marshaller <- ioToBc (mkMarshalCode cconv
1050 (r_offW, r_rep) addr_offW
1051 (zip args_offW a_reps))
1052 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1054 -- Offset of the next stack frame down the stack. The CCALL
1055 -- instruction needs to describe the chunk of stack containing
1056 -- the ccall args to the GC, so it needs to know how large it
1057 -- is. See comment in Interpreter.c with the CCALL instruction.
1058 stk_offset = d_after_r - s
1061 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1063 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1064 `snocOL` RETURN_UBX r_rep
1066 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1069 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1073 -- Make a dummy literal, to be used as a placeholder for FFI return
1074 -- values on the stack.
1075 mkDummyLiteral :: CgRep -> Literal
1078 NonPtrArg -> MachWord 0
1079 DoubleArg -> MachDouble 0
1080 FloatArg -> MachFloat 0
1081 LongArg -> MachWord64 0
1082 _ -> moan64 "mkDummyLiteral" (ppr pr)
1086 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1087 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1090 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1092 -- Alternatively, for call-targets returning nothing, convert
1094 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1095 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1099 maybe_getCCallReturnRep :: Type -> Maybe CgRep
1100 maybe_getCCallReturnRep fn_ty
1101 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1103 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1105 = case splitTyConApp_maybe (repType r_ty) of
1106 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
1108 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
1109 || r_reps == [VoidArg] )
1110 && isUnboxedTupleTyCon r_tycon
1111 && case maybe_r_rep_to_go of
1113 Just r_rep -> r_rep /= PtrArg
1114 -- if it was, it would be impossible
1115 -- to create a valid return value
1116 -- placeholder on the stack
1117 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1120 --trace (showSDoc (ppr (a_reps, r_reps))) $
1121 if ok then maybe_r_rep_to_go else blargh
1123 -- Compile code which expects an unboxed Int on the top of stack,
1124 -- (call it i), and pushes the i'th closure in the supplied list
1125 -- as a consequence.
1126 implement_tagToId :: [Name] -> BcM BCInstrList
1127 implement_tagToId names
1128 = ASSERT( notNull names )
1129 do labels <- getLabelsBc (length names)
1130 label_fail <- getLabelBc
1131 label_exit <- getLabelBc
1132 let infos = zip4 labels (tail labels ++ [label_fail])
1134 steps = map (mkStep label_exit) infos
1135 return (concatOL steps
1137 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1139 mkStep l_exit (my_label, next_label, n, name_for_n)
1140 = toOL [LABEL my_label,
1141 TESTEQ_I n next_label,
1146 -- -----------------------------------------------------------------------------
1149 -- Push an atom onto the stack, returning suitable code & number of
1150 -- stack words used.
1152 -- The env p must map each variable to the highest- numbered stack
1153 -- slot for it. For example, if the stack has depth 4 and we
1154 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1155 -- the tag in stack[5], the stack will have depth 6, and p must map v
1156 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1157 -- depth 6 stack has valid words 0 .. 5.
1159 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1161 pushAtom d p (AnnApp f (_, AnnType _))
1162 = pushAtom d p (snd f)
1164 pushAtom d p (AnnNote note e)
1165 = pushAtom d p (snd e)
1167 pushAtom d p (AnnLam x e)
1169 = pushAtom d p (snd e)
1171 pushAtom d p (AnnVar v)
1173 | idCgRep v == VoidArg
1177 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1179 | Just primop <- isPrimOpId_maybe v
1180 = return (unitOL (PUSH_PRIMOP primop), 1)
1182 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1183 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1184 -- d - d_v the number of words between the TOS
1185 -- and the 1st slot of the object
1187 -- d - d_v - 1 the offset from the TOS of the 1st slot
1189 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1192 -- Having found the last slot, we proceed to copy the right number of
1193 -- slots on to the top of the stack.
1195 | otherwise -- v must be a global variable
1197 return (unitOL (PUSH_G (getName v)), sz)
1203 pushAtom d p (AnnLit lit)
1205 MachLabel fs _ -> code NonPtrArg
1206 MachWord w -> code NonPtrArg
1207 MachInt i -> code PtrArg
1208 MachFloat r -> code FloatArg
1209 MachDouble r -> code DoubleArg
1210 MachChar c -> code NonPtrArg
1211 MachStr s -> pushStr s
1214 = let size_host_words = cgRepSizeW rep
1215 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1219 = let getMallocvilleAddr
1221 FastString _ n _ fp _ ->
1222 -- we could grab the Ptr from the ForeignPtr,
1223 -- but then we have no way to control its lifetime.
1224 -- In reality it'll probably stay alive long enoungh
1225 -- by virtue of the global FastString table, but
1226 -- to be on the safe side we copy the string into
1227 -- a malloc'd area of memory.
1228 do ptr <- ioToBc (mallocBytes (n+1))
1231 withForeignPtr fp $ \p -> do
1232 memcpy ptr p (fromIntegral n)
1233 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1237 addr <- getMallocvilleAddr
1238 -- Get the addr on the stack, untaggedly
1239 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1241 pushAtom d p (AnnCast e _)
1242 = pushAtom d p (snd e)
1245 = pprPanic "ByteCodeGen.pushAtom"
1246 (pprCoreExpr (deAnnotate (undefined, other)))
1248 foreign import ccall unsafe "memcpy"
1249 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1252 -- -----------------------------------------------------------------------------
1253 -- Given a bunch of alts code and their discrs, do the donkey work
1254 -- of making a multiway branch using a switch tree.
1255 -- What a load of hassle!
1257 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1258 -- a hint; generates better code
1259 -- Nothing is always safe
1260 -> [(Discr, BCInstrList)]
1262 mkMultiBranch maybe_ncons raw_ways
1263 = let d_way = filter (isNoDiscr.fst) raw_ways
1265 (\w1 w2 -> leAlt (fst w1) (fst w2))
1266 (filter (not.isNoDiscr.fst) raw_ways)
1268 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1269 mkTree [] range_lo range_hi = return the_default
1271 mkTree [val] range_lo range_hi
1272 | range_lo `eqAlt` range_hi
1275 = do label_neq <- getLabelBc
1276 return (mkTestEQ (fst val) label_neq
1278 `appOL` unitOL (LABEL label_neq)
1279 `appOL` the_default))
1281 mkTree vals range_lo range_hi
1282 = let n = length vals `div` 2
1283 vals_lo = take n vals
1284 vals_hi = drop n vals
1285 v_mid = fst (head vals_hi)
1287 label_geq <- getLabelBc
1288 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1289 code_hi <- mkTree vals_hi v_mid range_hi
1290 return (mkTestLT v_mid label_geq
1292 `appOL` unitOL (LABEL label_geq)
1296 = case d_way of [] -> unitOL CASEFAIL
1299 -- None of these will be needed if there are no non-default alts
1300 (mkTestLT, mkTestEQ, init_lo, init_hi)
1302 = panic "mkMultiBranch: awesome foursome"
1304 = case fst (head notd_ways) of {
1305 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1306 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1309 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1310 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1313 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1314 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1317 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1318 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1320 DiscrP algMaxBound )
1323 (algMinBound, algMaxBound)
1324 = case maybe_ncons of
1325 Just n -> (0, n - 1)
1326 Nothing -> (minBound, maxBound)
1328 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1329 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1330 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1331 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1332 NoDiscr `eqAlt` NoDiscr = True
1335 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1336 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1337 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1338 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1339 NoDiscr `leAlt` NoDiscr = True
1342 isNoDiscr NoDiscr = True
1345 dec (DiscrI i) = DiscrI (i-1)
1346 dec (DiscrP i) = DiscrP (i-1)
1347 dec other = other -- not really right, but if you
1348 -- do cases on floating values, you'll get what you deserve
1350 -- same snotty comment applies to the following
1352 minD, maxD :: Double
1358 mkTree notd_ways init_lo init_hi
1361 -- -----------------------------------------------------------------------------
1362 -- Supporting junk for the compilation schemes
1364 -- Describes case alts
1372 instance Outputable Discr where
1373 ppr (DiscrI i) = int i
1374 ppr (DiscrF f) = text (show f)
1375 ppr (DiscrD d) = text (show d)
1376 ppr (DiscrP i) = int i
1377 ppr NoDiscr = text "DEF"
1380 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1381 lookupBCEnv_maybe = lookupFM
1383 idSizeW :: Id -> Int
1384 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1387 unboxedTupleException :: a
1388 unboxedTupleException
1391 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1392 " Possibly due to foreign import/export decls in source.\n"++
1393 " Workaround: use -fobject-code, or compile this module to .o separately."))
1396 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1399 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1400 -- The arguments are returned in *right-to-left* order
1401 splitApp (AnnApp (_,f) (_,a))
1402 | isTypeAtom a = splitApp f
1403 | otherwise = case splitApp f of
1404 (f', as) -> (f', a:as)
1405 splitApp (AnnNote n (_,e)) = splitApp e
1406 splitApp (AnnCast (_,e) _) = splitApp e
1407 splitApp e = (e, [])
1410 isTypeAtom :: AnnExpr' id ann -> Bool
1411 isTypeAtom (AnnType _) = True
1412 isTypeAtom _ = False
1414 isVoidArgAtom :: AnnExpr' id ann -> Bool
1415 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1416 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1417 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1418 isVoidArgAtom _ = False
1420 atomRep :: AnnExpr' Id ann -> CgRep
1421 atomRep (AnnVar v) = typeCgRep (idType v)
1422 atomRep (AnnLit l) = typeCgRep (literalType l)
1423 atomRep (AnnNote n b) = atomRep (snd b)
1424 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1425 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1426 atomRep (AnnCast b _) = atomRep (snd b)
1427 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1429 isPtrAtom :: AnnExpr' Id ann -> Bool
1430 isPtrAtom e = atomRep e == PtrArg
1432 -- Let szsw be the sizes in words of some items pushed onto the stack,
1433 -- which has initial depth d'. Return the values which the stack environment
1434 -- should map these items to.
1435 mkStackOffsets :: Int -> [Int] -> [Int]
1436 mkStackOffsets original_depth szsw
1437 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1439 -- -----------------------------------------------------------------------------
1440 -- The bytecode generator's monad
1442 type BcPtr = Either ItblPtr (Ptr ())
1446 uniqSupply :: UniqSupply, -- for generating fresh variable names
1447 nextlabel :: Int, -- for generating local labels
1448 malloced :: [BcPtr], -- thunks malloced for current BCO
1449 -- Should be free()d when it is GCd
1450 breakArray :: BreakArray -- array of breakpoint flags
1453 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1455 ioToBc :: IO a -> BcM a
1456 ioToBc io = BcM $ \st -> do
1460 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1461 runBc us modBreaks (BcM m)
1462 = m (BcM_State us 0 [] breakArray)
1464 breakArray = modBreaks_flags modBreaks
1466 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1467 thenBc (BcM expr) cont = BcM $ \st0 -> do
1468 (st1, q) <- expr st0
1473 thenBc_ :: BcM a -> BcM b -> BcM b
1474 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1475 (st1, q) <- expr st0
1476 (st2, r) <- cont st1
1479 returnBc :: a -> BcM a
1480 returnBc result = BcM $ \st -> (return (st, result))
1482 instance Monad BcM where
1487 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1489 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1491 recordMallocBc :: Ptr a -> BcM ()
1493 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1495 recordItblMallocBc :: ItblPtr -> BcM ()
1496 recordItblMallocBc a
1497 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1499 getLabelBc :: BcM Int
1501 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1503 getLabelsBc :: Int -> BcM [Int]
1505 = BcM $ \st -> let ctr = nextlabel st
1506 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1508 getBreakArray :: BcM BreakArray
1509 getBreakArray = BcM $ \st -> return (st, breakArray st)
1511 newUnique :: BcM Unique
1513 \st -> case splitUniqSupply (uniqSupply st) of
1514 (us1, us2) -> let newState = st { uniqSupply = us2 }
1515 in return (newState, uniqFromSupply us1)
1517 newId :: Type -> BcM Id
1520 return $ mkSysLocal FSLIT("ticked") uniq ty