2 % (c) The University of Glasgow 2002-2006
5 ByteCodeGen: Generate bytecode from Core
8 module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where
10 #include "HsVersions.h"
53 import GHC.Exts ( Int(..), ByteArray# )
55 import Control.Monad ( when )
64 -- -----------------------------------------------------------------------------
65 -- Generating byte code for a complete module
67 byteCodeGen :: DynFlags
71 -> IO CompiledByteCode
72 byteCodeGen dflags binds tycs modBreaks
73 = do showPass dflags "ByteCodeGen"
75 let flatBinds = [ (bndr, freeVars rhs)
76 | (bndr, rhs) <- flattenBinds binds]
78 us <- mkSplitUniqSupply 'y'
79 (BcM_State _us _final_ctr mallocd _, proto_bcos)
80 <- runBc us modBreaks (mapM schemeTopBind flatBinds)
82 when (notNull mallocd)
83 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
85 dumpIfSet_dyn dflags Opt_D_dump_BCOs
86 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
88 assembleBCOs proto_bcos tycs
90 -- -----------------------------------------------------------------------------
91 -- Generating byte code for an expression
93 -- Returns: (the root BCO for this expression,
94 -- a list of auxilary BCOs resulting from compiling closures)
95 coreExprToBCOs :: DynFlags
98 coreExprToBCOs dflags expr
99 = do showPass dflags "ByteCodeGen"
101 -- create a totally bogus name for the top-level BCO; this
102 -- should be harmless, since it's never used for anything
103 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) (fsLit "ExprTopLevel")
104 invented_id = Id.mkLocalId invented_name (panic "invented_id's type")
106 -- the uniques are needed to generate fresh variables when we introduce new
107 -- let bindings for ticked expressions
108 us <- mkSplitUniqSupply 'y'
109 (BcM_State _us _final_ctr mallocd _ , proto_bco)
110 <- runBc us emptyModBreaks (schemeTopBind (invented_id, freeVars expr))
112 when (notNull mallocd)
113 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
115 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
117 assembleBCO proto_bco
120 -- -----------------------------------------------------------------------------
121 -- Compilation schema for the bytecode generator
123 type BCInstrList = OrdList BCInstr
125 type Sequel = Int -- back off to this depth before ENTER
127 -- Maps Ids to the offset from the stack _base_ so we don't have
128 -- to mess with it after each push/pop.
129 type BCEnv = FiniteMap Id Int -- To find vars on the stack
132 ppBCEnv :: BCEnv -> SDoc
135 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
138 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
139 cmp_snd x y = compare (snd x) (snd y)
142 -- Create a BCO and do a spot of peephole optimisation on the insns
147 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
151 -> Bool -- True <=> is a return point, rather than a function
154 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap is_ret mallocd_blocks
157 protoBCOInstrs = maybe_with_stack_check,
158 protoBCOBitmap = bitmap,
159 protoBCOBitmapSize = bitmap_size,
160 protoBCOArity = arity,
161 protoBCOExpr = origin,
162 protoBCOPtrs = mallocd_blocks
165 -- Overestimate the stack usage (in words) of this BCO,
166 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
167 -- stack check. (The interpreter always does a stack check
168 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
169 -- BCO anyway, so we only need to add an explicit one in the
170 -- (hopefully rare) cases when the (overestimated) stack use
171 -- exceeds iNTERP_STACK_CHECK_THRESH.
172 maybe_with_stack_check
173 | is_ret && stack_usage < aP_STACK_SPLIM = peep_d
174 -- don't do stack checks at return points,
175 -- everything is aggregated up to the top BCO
176 -- (which must be a function).
177 -- That is, unless the stack usage is >= AP_STACK_SPLIM,
179 | stack_usage >= iNTERP_STACK_CHECK_THRESH
180 = STKCHECK stack_usage : peep_d
182 = peep_d -- the supposedly common case
184 -- We assume that this sum doesn't wrap
185 stack_usage = sum (map bciStackUse peep_d)
187 -- Merge local pushes
188 peep_d = peep (fromOL instrs_ordlist)
190 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
191 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
192 peep (PUSH_L off1 : PUSH_L off2 : rest)
193 = PUSH_LL off1 (off2-1) : peep rest
199 argBits :: [CgRep] -> [Bool]
202 | isFollowableArg rep = False : argBits args
203 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
205 -- -----------------------------------------------------------------------------
208 -- Compile code for the right-hand side of a top-level binding
210 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
213 schemeTopBind (id, rhs)
214 | Just data_con <- isDataConWorkId_maybe id,
215 isNullaryRepDataCon data_con = do
216 -- Special case for the worker of a nullary data con.
217 -- It'll look like this: Nil = /\a -> Nil a
218 -- If we feed it into schemeR, we'll get
220 -- because mkConAppCode treats nullary constructor applications
221 -- by just re-using the single top-level definition. So
222 -- for the worker itself, we must allocate it directly.
223 -- ioToBc (putStrLn $ "top level BCO")
224 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
225 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
228 = schemeR [{- No free variables -}] (id, rhs)
231 -- -----------------------------------------------------------------------------
234 -- Compile code for a right-hand side, to give a BCO that,
235 -- when executed with the free variables and arguments on top of the stack,
236 -- will return with a pointer to the result on top of the stack, after
237 -- removing the free variables and arguments.
239 -- Park the resulting BCO in the monad. Also requires the
240 -- variable to which this value was bound, so as to give the
241 -- resulting BCO a name.
243 schemeR :: [Id] -- Free vars of the RHS, ordered as they
244 -- will appear in the thunk. Empty for
245 -- top-level things, which have no free vars.
246 -> (Id, AnnExpr Id VarSet)
247 -> BcM (ProtoBCO Name)
248 schemeR fvs (nm, rhs)
252 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
253 $$ pprCoreExpr (deAnnotate rhs)
259 = schemeR_wrk fvs nm rhs (collect rhs)
261 collect :: AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
262 collect (_, e) = go [] e
264 go xs e | Just e' <- bcView e = go xs e'
265 go xs (AnnLam x (_,e)) = go (x:xs) e
266 go xs not_lambda = (reverse xs, not_lambda)
268 schemeR_wrk :: [Id] -> Id -> AnnExpr Id VarSet -> ([Var], AnnExpr' Var VarSet) -> BcM (ProtoBCO Name)
269 schemeR_wrk fvs nm original_body (args, body)
271 all_args = reverse args ++ fvs
272 arity = length all_args
273 -- all_args are the args in reverse order. We're compiling a function
274 -- \fv1..fvn x1..xn -> e
275 -- i.e. the fvs come first
277 szsw_args = map idSizeW all_args
278 szw_args = sum szsw_args
279 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
281 -- make the arg bitmap
282 bits = argBits (reverse (map idCgRep all_args))
283 bitmap_size = length bits
284 bitmap = mkBitmap bits
286 body_code <- schemeER_wrk szw_args p_init body
288 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
289 arity bitmap_size bitmap False{-not alts-})
291 -- introduce break instructions for ticked expressions
292 schemeER_wrk :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
294 | Just (tickInfo, (_annot, newRhs)) <- isTickedExp' rhs = do
295 code <- schemeE d 0 p newRhs
297 let idOffSets = getVarOffSets d p tickInfo
298 let tickNumber = tickInfo_number tickInfo
299 let breakInfo = BreakInfo
300 { breakInfo_module = tickInfo_module tickInfo
301 , breakInfo_number = tickNumber
302 , breakInfo_vars = idOffSets
303 , breakInfo_resty = exprType (deAnnotate' newRhs)
305 let breakInstr = case arr of (BA arr#) -> BRK_FUN arr# tickNumber breakInfo
306 return $ breakInstr `consOL` code
307 | otherwise = schemeE d 0 p rhs
309 getVarOffSets :: Int -> BCEnv -> TickInfo -> [(Id, Int)]
310 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
312 getOffSet :: Int -> BCEnv -> Id -> Maybe (Id, Int)
314 = case lookupBCEnv_maybe env id of
316 Just offset -> Just (id, d - offset)
318 fvsToEnv :: BCEnv -> VarSet -> [Id]
319 -- Takes the free variables of a right-hand side, and
320 -- delivers an ordered list of the local variables that will
321 -- be captured in the thunk for the RHS
322 -- The BCEnv argument tells which variables are in the local
323 -- environment: these are the ones that should be captured
325 -- The code that constructs the thunk, and the code that executes
326 -- it, have to agree about this layout
327 fvsToEnv p fvs = [v | v <- varSetElems fvs,
328 isId v, -- Could be a type variable
331 -- -----------------------------------------------------------------------------
336 { tickInfo_number :: Int -- the (module) unique number of the tick
337 , tickInfo_module :: Module -- the origin of the ticked expression
338 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
341 instance Outputable TickInfo where
342 ppr info = text "TickInfo" <+>
343 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
344 ppr (tickInfo_locals info))
346 -- Compile code to apply the given expression to the remaining args
347 -- on the stack, returning a HNF.
348 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
351 | Just e' <- bcView e
354 -- Delegate tail-calls to schemeT.
355 schemeE d s p e@(AnnApp _ _)
358 schemeE d s p e@(AnnVar v)
359 | not (isUnLiftedType v_type)
360 = -- Lifted-type thing; push it in the normal way
364 = do -- Returning an unlifted value.
365 -- Heave it on the stack, SLIDE, and RETURN.
366 (push, szw) <- pushAtom d p (AnnVar v)
367 return (push -- value onto stack
368 `appOL` mkSLIDE szw (d-s) -- clear to sequel
369 `snocOL` RETURN_UBX v_rep) -- go
372 v_rep = typeCgRep v_type
374 schemeE d s p (AnnLit literal)
375 = do (push, szw) <- pushAtom d p (AnnLit literal)
376 let l_rep = typeCgRep (literalType literal)
377 return (push -- value onto stack
378 `appOL` mkSLIDE szw (d-s) -- clear to sequel
379 `snocOL` RETURN_UBX l_rep) -- go
381 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
382 | (AnnVar v, args_r_to_l) <- splitApp rhs,
383 Just data_con <- isDataConWorkId_maybe v,
384 dataConRepArity data_con == length args_r_to_l
385 = do -- Special case for a non-recursive let whose RHS is a
386 -- saturatred constructor application.
387 -- Just allocate the constructor and carry on
388 alloc_code <- mkConAppCode d s p data_con args_r_to_l
389 body_code <- schemeE (d+1) s (addToFM p x d) body
390 return (alloc_code `appOL` body_code)
392 -- General case for let. Generates correct, if inefficient, code in
394 schemeE d s p (AnnLet binds (_,body))
395 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
396 AnnRec xs_n_rhss -> unzip xs_n_rhss
399 fvss = map (fvsToEnv p' . fst) rhss
401 -- Sizes of free vars
402 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
404 -- the arity of each rhs
405 arities = map (length . fst . collect) rhss
407 -- This p', d' defn is safe because all the items being pushed
408 -- are ptrs, so all have size 1. d' and p' reflect the stack
409 -- after the closures have been allocated in the heap (but not
410 -- filled in), and pointers to them parked on the stack.
411 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
413 zipE = zipEqual "schemeE"
415 -- ToDo: don't build thunks for things with no free variables
416 build_thunk _ [] size bco off arity
417 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
419 mkap | arity == 0 = MKAP
421 build_thunk dd (fv:fvs) size bco off arity = do
422 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
423 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
424 return (push_code `appOL` more_push_code)
426 alloc_code = toOL (zipWith mkAlloc sizes arities)
428 | is_tick = ALLOC_AP_NOUPD sz
429 | otherwise = ALLOC_AP sz
430 mkAlloc sz arity = ALLOC_PAP arity sz
432 is_tick = case binds of
433 AnnNonRec id _ -> occNameFS (getOccName id) == tickFS
436 compile_bind d' fvs x rhs size arity off = do
437 bco <- schemeR fvs (x,rhs)
438 build_thunk d' fvs size bco off arity
441 [ compile_bind d' fvs x rhs size arity n
442 | (fvs, x, rhs, size, arity, n) <-
443 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
446 body_code <- schemeE d' s p' body
447 thunk_codes <- sequence compile_binds
448 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
450 -- introduce a let binding for a ticked case expression. This rule
451 -- *should* only fire when the expression was not already let-bound
452 -- (the code gen for let bindings should take care of that). Todo: we
453 -- call exprFreeVars on a deAnnotated expression, this may not be the
454 -- best way to calculate the free vars but it seemed like the least
455 -- intrusive thing to do
456 schemeE d s p exp@(AnnCase {})
457 | Just (_tickInfo, rhs) <- isTickedExp' exp
458 = if isUnLiftedType ty
459 then schemeE d s p (snd rhs)
462 -- Todo: is emptyVarSet correct on the next line?
463 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
465 where exp' = deAnnotate' exp
466 fvs = exprFreeVars exp'
469 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1, bind2], rhs)])
470 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
472 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
474 -- case .... of a { DEFAULT -> ... }
475 -- becuse the return convention for both are identical.
477 -- Note that it does not matter losing the void-rep thing from the
478 -- envt (it won't be bound now) because we never look such things up.
480 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
481 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
483 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
484 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
485 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
487 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1], rhs)])
488 | isUnboxedTupleCon dc
489 -- Similarly, convert
490 -- case .... of x { (# a #) -> ... }
492 -- case .... of a { DEFAULT -> ... }
493 = --trace "automagic mashing of case alts (# a #)" $
494 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
496 schemeE d s p (AnnCase scrut bndr _ alts)
497 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
500 = pprPanic "ByteCodeGen.schemeE: unhandled case"
501 (pprCoreExpr (deAnnotate' expr))
507 A ticked expression looks like this:
509 case tick<n> var1 ... varN of DEFAULT -> e
511 (*) <n> is the number of the tick, which is unique within a module
512 (*) var1 ... varN are the local variables in scope at the tick site
514 If we find a ticked expression we return:
516 Just ((n, [var1 ... varN]), e)
518 otherwise we return Nothing.
520 The idea is that the "case tick<n> ..." is really just an annotation on
521 the code. When we find such a thing, we pull out the useful information,
522 and then compile the code as if it was just the expression "e".
526 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
527 isTickedExp' (AnnCase scrut _bndr _type alts)
528 | Just tickInfo <- isTickedScrut scrut,
529 [(DEFAULT, _bndr, rhs)] <- alts
530 = Just (tickInfo, rhs)
532 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
535 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
536 = Just $ TickInfo { tickInfo_number = tickNumber
537 , tickInfo_module = modName
538 , tickInfo_locals = idsOfArgs args
540 | otherwise = Nothing
542 (f, args) = collectArgs $ deAnnotate expr
543 idsOfArgs :: [Expr Id] -> [Id]
544 idsOfArgs = catMaybes . map exprId
545 exprId :: Expr Id -> Maybe Id
546 exprId (Var id) = Just id
549 isTickedExp' _ = Nothing
551 -- Compile code to do a tail call. Specifically, push the fn,
552 -- slide the on-stack app back down to the sequel depth,
553 -- and enter. Four cases:
556 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
557 -- The int will be on the stack. Generate a code sequence
558 -- to convert it to the relevant constructor, SLIDE and ENTER.
560 -- 1. The fn denotes a ccall. Defer to generateCCall.
562 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
563 -- it simply as b -- since the representations are identical
564 -- (the VoidArg takes up zero stack space). Also, spot
565 -- (# b #) and treat it as b.
567 -- 3. Application of a constructor, by defn saturated.
568 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
569 -- then the ptrs, and then do PACK and RETURN.
571 -- 4. Otherwise, it must be a function call. Push the args
572 -- right to left, SLIDE and ENTER.
574 schemeT :: Int -- Stack depth
575 -> Sequel -- Sequel depth
576 -> BCEnv -- stack env
577 -> AnnExpr' Id VarSet
582 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
583 -- = panic "schemeT ?!?!"
585 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
589 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
590 = do (push, arg_words) <- pushAtom d p arg
591 tagToId_sequence <- implement_tagToId constr_names
592 return (push `appOL` tagToId_sequence
593 `appOL` mkSLIDE 1 (d+arg_words-s)
597 | Just (CCall ccall_spec) <- isFCallId_maybe fn
598 = generateCCall d s p ccall_spec fn args_r_to_l
600 -- Case 2: Constructor application
601 | Just con <- maybe_saturated_dcon,
602 isUnboxedTupleCon con
603 = case args_r_to_l of
604 [arg1,arg2] | isVoidArgAtom arg1 ->
605 unboxedTupleReturn d s p arg2
606 [arg1,arg2] | isVoidArgAtom arg2 ->
607 unboxedTupleReturn d s p arg1
608 _other -> unboxedTupleException
610 -- Case 3: Ordinary data constructor
611 | Just con <- maybe_saturated_dcon
612 = do alloc_con <- mkConAppCode d s p con args_r_to_l
613 return (alloc_con `appOL`
614 mkSLIDE 1 (d - s) `snocOL`
617 -- Case 4: Tail call of function
619 = doTailCall d s p fn args_r_to_l
622 -- Detect and extract relevant info for the tagToEnum kludge.
623 maybe_is_tagToEnum_call
624 = let extract_constr_Names ty
625 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
627 = map (getName . dataConWorkId) (tyConDataCons tyc)
628 -- NOTE: use the worker name, not the source name of
629 -- the DataCon. See DataCon.lhs for details.
631 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
634 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
635 -> case isPrimOpId_maybe v of
636 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
640 -- Extract the args (R->L) and fn
641 -- The function will necessarily be a variable,
642 -- because we are compiling a tail call
643 (AnnVar fn, args_r_to_l) = splitApp app
645 -- Only consider this to be a constructor application iff it is
646 -- saturated. Otherwise, we'll call the constructor wrapper.
647 n_args = length args_r_to_l
649 = case isDataConWorkId_maybe fn of
650 Just con | dataConRepArity con == n_args -> Just con
653 -- -----------------------------------------------------------------------------
654 -- Generate code to build a constructor application,
655 -- leaving it on top of the stack
657 mkConAppCode :: Int -> Sequel -> BCEnv
658 -> DataCon -- The data constructor
659 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
662 mkConAppCode _ _ _ con [] -- Nullary constructor
663 = ASSERT( isNullaryRepDataCon con )
664 return (unitOL (PUSH_G (getName (dataConWorkId con))))
665 -- Instead of doing a PACK, which would allocate a fresh
666 -- copy of this constructor, use the single shared version.
668 mkConAppCode orig_d _ p con args_r_to_l
669 = ASSERT( dataConRepArity con == length args_r_to_l )
670 do_pushery orig_d (non_ptr_args ++ ptr_args)
672 -- The args are already in reverse order, which is the way PACK
673 -- expects them to be. We must push the non-ptrs after the ptrs.
674 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
676 do_pushery d (arg:args)
677 = do (push, arg_words) <- pushAtom d p arg
678 more_push_code <- do_pushery (d+arg_words) args
679 return (push `appOL` more_push_code)
681 = return (unitOL (PACK con n_arg_words))
683 n_arg_words = d - orig_d
686 -- -----------------------------------------------------------------------------
687 -- Returning an unboxed tuple with one non-void component (the only
688 -- case we can handle).
690 -- Remember, we don't want to *evaluate* the component that is being
691 -- returned, even if it is a pointed type. We always just return.
694 :: Int -> Sequel -> BCEnv
695 -> AnnExpr' Id VarSet -> BcM BCInstrList
696 unboxedTupleReturn d s p arg = do
697 (push, sz) <- pushAtom d p arg
699 mkSLIDE sz (d-s) `snocOL`
700 RETURN_UBX (atomRep arg))
702 -- -----------------------------------------------------------------------------
703 -- Generate code for a tail-call
706 :: Int -> Sequel -> BCEnv
707 -> Id -> [AnnExpr' Id VarSet]
709 doTailCall init_d s p fn args
710 = do_pushes init_d args (map atomRep args)
712 do_pushes d [] reps = do
713 ASSERT( null reps ) return ()
714 (push_fn, sz) <- pushAtom d p (AnnVar fn)
715 ASSERT( sz == 1 ) return ()
716 return (push_fn `appOL` (
717 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
719 do_pushes d args reps = do
720 let (push_apply, n, rest_of_reps) = findPushSeq reps
721 (these_args, rest_of_args) = splitAt n args
722 (next_d, push_code) <- push_seq d these_args
723 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
724 -- ^^^ for the PUSH_APPLY_ instruction
725 return (push_code `appOL` (push_apply `consOL` instrs))
727 push_seq d [] = return (d, nilOL)
728 push_seq d (arg:args) = do
729 (push_code, sz) <- pushAtom d p arg
730 (final_d, more_push_code) <- push_seq (d+sz) args
731 return (final_d, push_code `appOL` more_push_code)
733 -- v. similar to CgStackery.findMatch, ToDo: merge
734 findPushSeq :: [CgRep] -> (BCInstr, Int, [CgRep])
735 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
736 = (PUSH_APPLY_PPPPPP, 6, rest)
737 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
738 = (PUSH_APPLY_PPPPP, 5, rest)
739 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
740 = (PUSH_APPLY_PPPP, 4, rest)
741 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
742 = (PUSH_APPLY_PPP, 3, rest)
743 findPushSeq (PtrArg: PtrArg: rest)
744 = (PUSH_APPLY_PP, 2, rest)
745 findPushSeq (PtrArg: rest)
746 = (PUSH_APPLY_P, 1, rest)
747 findPushSeq (VoidArg: rest)
748 = (PUSH_APPLY_V, 1, rest)
749 findPushSeq (NonPtrArg: rest)
750 = (PUSH_APPLY_N, 1, rest)
751 findPushSeq (FloatArg: rest)
752 = (PUSH_APPLY_F, 1, rest)
753 findPushSeq (DoubleArg: rest)
754 = (PUSH_APPLY_D, 1, rest)
755 findPushSeq (LongArg: rest)
756 = (PUSH_APPLY_L, 1, rest)
758 = panic "ByteCodeGen.findPushSeq"
760 -- -----------------------------------------------------------------------------
763 doCase :: Int -> Sequel -> BCEnv
764 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
765 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
767 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
769 -- Top of stack is the return itbl, as usual.
770 -- underneath it is the pointer to the alt_code BCO.
771 -- When an alt is entered, it assumes the returned value is
772 -- on top of the itbl.
775 -- An unlifted value gets an extra info table pushed on top
776 -- when it is returned.
777 unlifted_itbl_sizeW | isAlgCase = 0
780 -- depth of stack after the return value has been pushed
781 d_bndr = d + ret_frame_sizeW + idSizeW bndr
783 -- depth of stack after the extra info table for an unboxed return
784 -- has been pushed, if any. This is the stack depth at the
786 d_alts = d_bndr + unlifted_itbl_sizeW
788 -- Env in which to compile the alts, not including
789 -- any vars bound by the alts themselves
790 p_alts = addToFM p bndr (d_bndr - 1)
792 bndr_ty = idType bndr
793 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
795 -- given an alt, return a discr and code for it.
796 codeAlt (DEFAULT, _, (_,rhs))
797 = do rhs_code <- schemeE d_alts s p_alts rhs
798 return (NoDiscr, rhs_code)
800 codeAlt alt@(_, bndrs, (_,rhs))
801 -- primitive or nullary constructor alt: no need to UNPACK
802 | null real_bndrs = do
803 rhs_code <- schemeE d_alts s p_alts rhs
804 return (my_discr alt, rhs_code)
805 -- algebraic alt with some binders
808 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
809 ptr_sizes = map idSizeW ptrs
810 nptrs_sizes = map idSizeW nptrs
811 bind_sizes = ptr_sizes ++ nptrs_sizes
812 size = sum ptr_sizes + sum nptrs_sizes
813 -- the UNPACK instruction unpacks in reverse order...
814 p' = addListToFM p_alts
815 (zip (reverse (ptrs ++ nptrs))
816 (mkStackOffsets d_alts (reverse bind_sizes)))
819 rhs_code <- schemeE (d_alts+size) s p' rhs
820 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
822 real_bndrs = filter (not.isTyVar) bndrs
824 my_discr (DEFAULT, _, _) = NoDiscr {-shouldn't really happen-}
825 my_discr (DataAlt dc, _, _)
826 | isUnboxedTupleCon dc
827 = unboxedTupleException
829 = DiscrP (dataConTag dc - fIRST_TAG)
830 my_discr (LitAlt l, _, _)
831 = case l of MachInt i -> DiscrI (fromInteger i)
832 MachFloat r -> DiscrF (fromRational r)
833 MachDouble r -> DiscrD (fromRational r)
834 MachChar i -> DiscrI (ord i)
835 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
838 | not isAlgCase = Nothing
840 = case [dc | (DataAlt dc, _, _) <- alts] of
842 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
844 -- the bitmap is relative to stack depth d, i.e. before the
845 -- BCO, info table and return value are pushed on.
846 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
847 -- except that here we build the bitmap from the known bindings of
848 -- things that are pointers, whereas in CgBindery the code builds the
849 -- bitmap from the free slots and unboxed bindings.
852 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
853 -- The bitmap must cover the portion of the stack up to the sequel only.
854 -- Previously we were building a bitmap for the whole depth (d), but we
855 -- really want a bitmap up to depth (d-s). This affects compilation of
856 -- case-of-case expressions, which is the only time we can be compiling a
857 -- case expression with s /= 0.
859 bitmap = intsToReverseBitmap bitmap_size{-size-}
860 (sortLe (<=) (filter (< bitmap_size) rel_slots))
863 rel_slots = concat (map spread binds)
865 | isFollowableArg (idCgRep id) = [ rel_offset ]
867 where rel_offset = d - offset - 1
870 alt_stuff <- mapM codeAlt alts
871 alt_final <- mkMultiBranch maybe_ncons alt_stuff
874 alt_bco_name = getName bndr
875 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
876 0{-no arity-} bitmap_size bitmap True{-is alts-}
878 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
879 -- "\n bitmap = " ++ show bitmap) $ do
880 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
881 alt_bco' <- emitBc alt_bco
883 | isAlgCase = PUSH_ALTS alt_bco'
884 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
885 return (push_alts `consOL` scrut_code)
888 -- -----------------------------------------------------------------------------
889 -- Deal with a CCall.
891 -- Taggedly push the args onto the stack R->L,
892 -- deferencing ForeignObj#s and adjusting addrs to point to
893 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
894 -- (machine) code for the ccall, and create bytecodes to call that and
895 -- then return in the right way.
897 generateCCall :: Int -> Sequel -- stack and sequel depths
899 -> CCallSpec -- where to call
900 -> Id -- of target, for type info
901 -> [AnnExpr' Id VarSet] -- args (atoms)
904 generateCCall d0 s p (CCallSpec target cconv _) fn args_r_to_l
907 addr_sizeW = cgRepSizeW NonPtrArg
909 -- Get the args on the stack, with tags and suitably
910 -- dereferenced for the CCall. For each arg, return the
911 -- depth to the first word of the bits for that arg, and the
912 -- CgRep of what was actually pushed.
914 pargs _ [] = return []
916 = let arg_ty = repType (exprType (deAnnotate' a))
918 in case splitTyConApp_maybe arg_ty of
919 -- Don't push the FO; instead push the Addr# it
922 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
923 -> do rest <- pargs (d + addr_sizeW) az
924 code <- parg_ArrayishRep arrPtrsHdrSize d p a
925 return ((code,AddrRep):rest)
927 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
928 -> do rest <- pargs (d + addr_sizeW) az
929 code <- parg_ArrayishRep arrWordsHdrSize d p a
930 return ((code,AddrRep):rest)
932 -- Default case: push taggedly, but otherwise intact.
934 -> do (code_a, sz_a) <- pushAtom d p a
935 rest <- pargs (d+sz_a) az
936 return ((code_a, atomPrimRep a) : rest)
938 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
939 -- the stack but then advance it over the headers, so as to
940 -- point to the payload.
941 parg_ArrayishRep hdrSize d p a
942 = do (push_fo, _) <- pushAtom d p a
943 -- The ptr points at the header. Advance it over the
944 -- header and then pretend this is an Addr#.
945 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
948 code_n_reps <- pargs d0 args_r_to_l
950 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
951 a_reps_sizeW = sum (map primRepSizeW a_reps_pushed_r_to_l)
953 push_args = concatOL pushs_arg
954 d_after_args = d0 + a_reps_sizeW
956 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
957 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
959 = reverse (tail a_reps_pushed_r_to_l)
961 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
962 -- push_args is the code to do that.
963 -- d_after_args is the stack depth once the args are on.
965 -- Get the result rep.
966 (returns_void, r_rep)
967 = case maybe_getCCallReturnRep (idType fn) of
968 Nothing -> (True, VoidRep)
969 Just rr -> (False, rr)
971 Because the Haskell stack grows down, the a_reps refer to
972 lowest to highest addresses in that order. The args for the call
973 are on the stack. Now push an unboxed Addr# indicating
974 the C function to call. Then push a dummy placeholder for the
975 result. Finally, emit a CCALL insn with an offset pointing to the
976 Addr# just pushed, and a literal field holding the mallocville
977 address of the piece of marshalling code we generate.
978 So, just prior to the CCALL insn, the stack looks like this
979 (growing down, as usual):
984 Addr# address_of_C_fn
985 <placeholder-for-result#> (must be an unboxed type)
987 The interpreter then calls the marshall code mentioned
988 in the CCALL insn, passing it (& <placeholder-for-result#>),
989 that is, the addr of the topmost word in the stack.
990 When this returns, the placeholder will have been
991 filled in. The placeholder is slid down to the sequel
992 depth, and we RETURN.
994 This arrangement makes it simple to do f-i-dynamic since the Addr#
995 value is the first arg anyway.
997 The marshalling code is generated specifically for this
998 call site, and so knows exactly the (Haskell) stack
999 offsets of the args, fn address and placeholder. It
1000 copies the args to the C stack, calls the stacked addr,
1001 and parks the result back in the placeholder. The interpreter
1002 calls it as a normal C call, assuming it has a signature
1003 void marshall_code ( StgWord* ptr_to_top_of_stack )
1005 -- resolve static address
1009 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
1011 -> do res <- ioToBc (lookupStaticPtr stdcall_adj_target)
1015 #ifdef mingw32_TARGET_OS
1016 | StdCallConv <- cconv
1017 = let size = a_reps_sizeW * wORD_SIZE in
1018 mkFastString (unpackFS target ++ '@':show size)
1024 (is_static, static_target_addr) <- get_target_info
1027 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1028 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1029 | is_static = a_reps_pushed_RAW
1030 | otherwise = if null a_reps_pushed_RAW
1031 then panic "ByteCodeGen.generateCCall: dyn with no args"
1032 else tail a_reps_pushed_RAW
1035 (push_Addr, d_after_Addr)
1037 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1038 d_after_args + addr_sizeW)
1039 | otherwise -- is already on the stack
1040 = (nilOL, d_after_args)
1042 -- Push the return placeholder. For a call returning nothing,
1043 -- this is a VoidArg (tag).
1044 r_sizeW = primRepSizeW r_rep
1045 d_after_r = d_after_Addr + r_sizeW
1046 r_lit = mkDummyLiteral r_rep
1047 push_r = (if returns_void
1049 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1051 -- generate the marshalling code we're going to call
1053 -- Offset of the next stack frame down the stack. The CCALL
1054 -- instruction needs to describe the chunk of stack containing
1055 -- the ccall args to the GC, so it needs to know how large it
1056 -- is. See comment in Interpreter.c with the CCALL instruction.
1057 stk_offset = d_after_r - s
1060 -- the only difference in libffi mode is that we prepare a cif
1061 -- describing the call type by calling libffi, and we attach the
1062 -- address of this to the CCALL instruction.
1063 token <- ioToBc $ prepForeignCall cconv a_reps r_rep
1064 let addr_of_marshaller = castPtrToFunPtr token
1066 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1069 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1071 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1072 `snocOL` RETURN_UBX (primRepToCgRep r_rep)
1074 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1077 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1080 -- Make a dummy literal, to be used as a placeholder for FFI return
1081 -- values on the stack.
1082 mkDummyLiteral :: PrimRep -> Literal
1086 WordRep -> MachWord 0
1087 AddrRep -> MachNullAddr
1088 DoubleRep -> MachDouble 0
1089 FloatRep -> MachFloat 0
1090 Int64Rep -> MachInt64 0
1091 Word64Rep -> MachWord64 0
1092 _ -> panic "mkDummyLiteral"
1096 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1097 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1100 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1102 -- Alternatively, for call-targets returning nothing, convert
1104 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1105 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1109 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
1110 maybe_getCCallReturnRep fn_ty
1111 = let (_a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1113 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1115 = case splitTyConApp_maybe (repType r_ty) of
1116 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
1118 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
1119 || r_reps == [VoidRep] )
1120 && isUnboxedTupleTyCon r_tycon
1121 && case maybe_r_rep_to_go of
1123 Just r_rep -> r_rep /= PtrRep
1124 -- if it was, it would be impossible
1125 -- to create a valid return value
1126 -- placeholder on the stack
1127 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1130 --trace (showSDoc (ppr (a_reps, r_reps))) $
1131 if ok then maybe_r_rep_to_go else blargh
1133 -- Compile code which expects an unboxed Int on the top of stack,
1134 -- (call it i), and pushes the i'th closure in the supplied list
1135 -- as a consequence.
1136 implement_tagToId :: [Name] -> BcM BCInstrList
1137 implement_tagToId names
1138 = ASSERT( notNull names )
1139 do labels <- getLabelsBc (length names)
1140 label_fail <- getLabelBc
1141 label_exit <- getLabelBc
1142 let infos = zip4 labels (tail labels ++ [label_fail])
1144 steps = map (mkStep label_exit) infos
1145 return (concatOL steps
1147 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1149 mkStep l_exit (my_label, next_label, n, name_for_n)
1150 = toOL [LABEL my_label,
1151 TESTEQ_I n next_label,
1156 -- -----------------------------------------------------------------------------
1159 -- Push an atom onto the stack, returning suitable code & number of
1160 -- stack words used.
1162 -- The env p must map each variable to the highest- numbered stack
1163 -- slot for it. For example, if the stack has depth 4 and we
1164 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1165 -- the tag in stack[5], the stack will have depth 6, and p must map v
1166 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1167 -- depth 6 stack has valid words 0 .. 5.
1169 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1172 | Just e' <- bcView e
1175 pushAtom d p (AnnVar v)
1176 | idCgRep v == VoidArg
1180 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1182 | Just primop <- isPrimOpId_maybe v
1183 = return (unitOL (PUSH_PRIMOP primop), 1)
1185 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1186 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1187 -- d - d_v the number of words between the TOS
1188 -- and the 1st slot of the object
1190 -- d - d_v - 1 the offset from the TOS of the 1st slot
1192 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1195 -- Having found the last slot, we proceed to copy the right number of
1196 -- slots on to the top of the stack.
1198 | otherwise -- v must be a global variable
1200 return (unitOL (PUSH_G (getName v)), sz)
1206 pushAtom _ _ (AnnLit lit)
1208 MachLabel _ _ -> code NonPtrArg
1209 MachWord _ -> code NonPtrArg
1210 MachInt _ -> code PtrArg
1211 MachFloat _ -> code FloatArg
1212 MachDouble _ -> code DoubleArg
1213 MachChar _ -> code NonPtrArg
1214 MachNullAddr -> code NonPtrArg
1215 MachStr s -> pushStr s
1216 l -> pprPanic "pushAtom" (ppr l)
1219 = let size_host_words = cgRepSizeW rep
1220 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1224 = let getMallocvilleAddr
1226 FastString _ n _ fp _ ->
1227 -- we could grab the Ptr from the ForeignPtr,
1228 -- but then we have no way to control its lifetime.
1229 -- In reality it'll probably stay alive long enoungh
1230 -- by virtue of the global FastString table, but
1231 -- to be on the safe side we copy the string into
1232 -- a malloc'd area of memory.
1233 do ptr <- ioToBc (mallocBytes (n+1))
1236 withForeignPtr fp $ \p -> do
1237 memcpy ptr p (fromIntegral n)
1238 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1242 addr <- getMallocvilleAddr
1243 -- Get the addr on the stack, untaggedly
1244 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1246 pushAtom d p (AnnCast e _)
1247 = pushAtom d p (snd e)
1250 = pprPanic "ByteCodeGen.pushAtom"
1251 (pprCoreExpr (deAnnotate (undefined, expr)))
1253 foreign import ccall unsafe "memcpy"
1254 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1257 -- -----------------------------------------------------------------------------
1258 -- Given a bunch of alts code and their discrs, do the donkey work
1259 -- of making a multiway branch using a switch tree.
1260 -- What a load of hassle!
1262 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1263 -- a hint; generates better code
1264 -- Nothing is always safe
1265 -> [(Discr, BCInstrList)]
1267 mkMultiBranch maybe_ncons raw_ways
1268 = let d_way = filter (isNoDiscr.fst) raw_ways
1270 (\w1 w2 -> leAlt (fst w1) (fst w2))
1271 (filter (not.isNoDiscr.fst) raw_ways)
1273 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1274 mkTree [] _range_lo _range_hi = return the_default
1276 mkTree [val] range_lo range_hi
1277 | range_lo `eqAlt` range_hi
1280 = do label_neq <- getLabelBc
1281 return (mkTestEQ (fst val) label_neq
1283 `appOL` unitOL (LABEL label_neq)
1284 `appOL` the_default))
1286 mkTree vals range_lo range_hi
1287 = let n = length vals `div` 2
1288 vals_lo = take n vals
1289 vals_hi = drop n vals
1290 v_mid = fst (head vals_hi)
1292 label_geq <- getLabelBc
1293 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1294 code_hi <- mkTree vals_hi v_mid range_hi
1295 return (mkTestLT v_mid label_geq
1297 `appOL` unitOL (LABEL label_geq)
1301 = case d_way of [] -> unitOL CASEFAIL
1303 _ -> panic "mkMultiBranch/the_default"
1305 -- None of these will be needed if there are no non-default alts
1306 (mkTestLT, mkTestEQ, init_lo, init_hi)
1308 = panic "mkMultiBranch: awesome foursome"
1310 = case fst (head notd_ways) of {
1311 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1312 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1315 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1316 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1319 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1320 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1323 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1324 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1326 DiscrP algMaxBound );
1327 NoDiscr -> panic "mkMultiBranch NoDiscr"
1330 (algMinBound, algMaxBound)
1331 = case maybe_ncons of
1332 Just n -> (0, n - 1)
1333 Nothing -> (minBound, maxBound)
1335 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1336 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1337 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1338 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1339 NoDiscr `eqAlt` NoDiscr = True
1342 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1343 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1344 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1345 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1346 NoDiscr `leAlt` NoDiscr = True
1349 isNoDiscr NoDiscr = True
1352 dec (DiscrI i) = DiscrI (i-1)
1353 dec (DiscrP i) = DiscrP (i-1)
1354 dec other = other -- not really right, but if you
1355 -- do cases on floating values, you'll get what you deserve
1357 -- same snotty comment applies to the following
1359 minD, maxD :: Double
1365 mkTree notd_ways init_lo init_hi
1368 -- -----------------------------------------------------------------------------
1369 -- Supporting junk for the compilation schemes
1371 -- Describes case alts
1379 instance Outputable Discr where
1380 ppr (DiscrI i) = int i
1381 ppr (DiscrF f) = text (show f)
1382 ppr (DiscrD d) = text (show d)
1383 ppr (DiscrP i) = int i
1384 ppr NoDiscr = text "DEF"
1387 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1388 lookupBCEnv_maybe = lookupFM
1390 idSizeW :: Id -> Int
1391 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1394 unboxedTupleException :: a
1395 unboxedTupleException
1398 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1399 " Possibly due to foreign import/export decls in source.\n"++
1400 " Workaround: use -fobject-code, or compile this module to .o separately."))
1403 mkSLIDE :: Int -> Int -> OrdList BCInstr
1404 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1406 splitApp :: AnnExpr' Var ann -> (AnnExpr' Var ann, [AnnExpr' Var ann])
1407 -- The arguments are returned in *right-to-left* order
1408 splitApp e | Just e' <- bcView e = splitApp e'
1409 splitApp (AnnApp (_,f) (_,a)) = case splitApp f of
1410 (f', as) -> (f', a:as)
1411 splitApp e = (e, [])
1414 bcView :: AnnExpr' Var ann -> Maybe (AnnExpr' Var ann)
1415 -- The "bytecode view" of a term discards
1416 -- a) type abstractions
1417 -- b) type applications
1420 -- Type lambdas *can* occur in random expressions,
1421 -- whereas value lambdas cannot; that is why they are nuked here
1422 bcView (AnnNote _ (_,e)) = Just e
1423 bcView (AnnCast (_,e) _) = Just e
1424 bcView (AnnLam v (_,e)) | isTyVar v = Just e
1425 bcView (AnnApp (_,e) (_, AnnType _)) = Just e
1428 isVoidArgAtom :: AnnExpr' Var ann -> Bool
1429 isVoidArgAtom e | Just e' <- bcView e = isVoidArgAtom e'
1430 isVoidArgAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1431 isVoidArgAtom _ = False
1433 atomPrimRep :: AnnExpr' Id ann -> PrimRep
1434 atomPrimRep e | Just e' <- bcView e = atomPrimRep e'
1435 atomPrimRep (AnnVar v) = typePrimRep (idType v)
1436 atomPrimRep (AnnLit l) = typePrimRep (literalType l)
1437 atomPrimRep other = pprPanic "atomPrimRep" (ppr (deAnnotate (undefined,other)))
1439 atomRep :: AnnExpr' Id ann -> CgRep
1440 atomRep e = primRepToCgRep (atomPrimRep e)
1442 isPtrAtom :: AnnExpr' Id ann -> Bool
1443 isPtrAtom e = atomRep e == PtrArg
1445 -- Let szsw be the sizes in words of some items pushed onto the stack,
1446 -- which has initial depth d'. Return the values which the stack environment
1447 -- should map these items to.
1448 mkStackOffsets :: Int -> [Int] -> [Int]
1449 mkStackOffsets original_depth szsw
1450 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1452 -- -----------------------------------------------------------------------------
1453 -- The bytecode generator's monad
1455 type BcPtr = Either ItblPtr (Ptr ())
1459 uniqSupply :: UniqSupply, -- for generating fresh variable names
1460 nextlabel :: Int, -- for generating local labels
1461 malloced :: [BcPtr], -- thunks malloced for current BCO
1462 -- Should be free()d when it is GCd
1463 breakArray :: BreakArray -- array of breakpoint flags
1466 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1468 ioToBc :: IO a -> BcM a
1469 ioToBc io = BcM $ \st -> do
1473 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1474 runBc us modBreaks (BcM m)
1475 = m (BcM_State us 0 [] breakArray)
1477 breakArray = modBreaks_flags modBreaks
1479 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1480 thenBc (BcM expr) cont = BcM $ \st0 -> do
1481 (st1, q) <- expr st0
1486 thenBc_ :: BcM a -> BcM b -> BcM b
1487 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1488 (st1, _) <- expr st0
1489 (st2, r) <- cont st1
1492 returnBc :: a -> BcM a
1493 returnBc result = BcM $ \st -> (return (st, result))
1495 instance Monad BcM where
1500 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1502 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1504 recordMallocBc :: Ptr a -> BcM ()
1506 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1508 recordItblMallocBc :: ItblPtr -> BcM ()
1509 recordItblMallocBc a
1510 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1512 getLabelBc :: BcM Int
1514 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1516 getLabelsBc :: Int -> BcM [Int]
1518 = BcM $ \st -> let ctr = nextlabel st
1519 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1521 getBreakArray :: BcM BreakArray
1522 getBreakArray = BcM $ \st -> return (st, breakArray st)
1524 newUnique :: BcM Unique
1526 \st -> case splitUniqSupply (uniqSupply st) of
1527 (us1, us2) -> let newState = st { uniqSupply = us2 }
1528 in return (newState, uniqFromSupply us1)
1530 newId :: Type -> BcM Id
1533 return $ mkSysLocal tickFS uniq ty
1535 tickFS :: FastString
1536 tickFS = fsLit "ticked"