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"
52 import Control.Exception ( throwDyn )
54 import GHC.Exts ( Int(..), ByteArray# )
56 import Control.Monad ( when )
65 -- -----------------------------------------------------------------------------
66 -- Generating byte code for a complete module
68 byteCodeGen :: DynFlags
72 -> IO CompiledByteCode
73 byteCodeGen dflags binds tycs modBreaks
74 = do showPass dflags "ByteCodeGen"
76 let flatBinds = [ (bndr, freeVars rhs)
77 | (bndr, rhs) <- flattenBinds binds]
79 us <- mkSplitUniqSupply 'y'
80 (BcM_State _us _final_ctr mallocd _, proto_bcos)
81 <- runBc us modBreaks (mapM schemeTopBind flatBinds)
83 when (notNull mallocd)
84 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
86 dumpIfSet_dyn dflags Opt_D_dump_BCOs
87 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
89 assembleBCOs proto_bcos tycs
91 -- -----------------------------------------------------------------------------
92 -- Generating byte code for an expression
94 -- Returns: (the root BCO for this expression,
95 -- a list of auxilary BCOs resulting from compiling closures)
96 coreExprToBCOs :: DynFlags
99 coreExprToBCOs dflags expr
100 = do showPass dflags "ByteCodeGen"
102 -- create a totally bogus name for the top-level BCO; this
103 -- should be harmless, since it's never used for anything
104 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) (fsLit "ExprTopLevel")
105 invented_id = Id.mkLocalId invented_name (panic "invented_id's type")
107 -- the uniques are needed to generate fresh variables when we introduce new
108 -- let bindings for ticked expressions
109 us <- mkSplitUniqSupply 'y'
110 (BcM_State _us _final_ctr mallocd _ , proto_bco)
111 <- runBc us emptyModBreaks (schemeTopBind (invented_id, freeVars expr))
113 when (notNull mallocd)
114 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
116 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
118 assembleBCO proto_bco
121 -- -----------------------------------------------------------------------------
122 -- Compilation schema for the bytecode generator
124 type BCInstrList = OrdList BCInstr
126 type Sequel = Int -- back off to this depth before ENTER
128 -- Maps Ids to the offset from the stack _base_ so we don't have
129 -- to mess with it after each push/pop.
130 type BCEnv = FiniteMap Id Int -- To find vars on the stack
133 ppBCEnv :: BCEnv -> SDoc
136 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
139 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
140 cmp_snd x y = compare (snd x) (snd y)
143 -- Create a BCO and do a spot of peephole optimisation on the insns
148 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
152 -> Bool -- True <=> is a return point, rather than a function
155 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap is_ret mallocd_blocks
158 protoBCOInstrs = maybe_with_stack_check,
159 protoBCOBitmap = bitmap,
160 protoBCOBitmapSize = bitmap_size,
161 protoBCOArity = arity,
162 protoBCOExpr = origin,
163 protoBCOPtrs = mallocd_blocks
166 -- Overestimate the stack usage (in words) of this BCO,
167 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
168 -- stack check. (The interpreter always does a stack check
169 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
170 -- BCO anyway, so we only need to add an explicit on in the
171 -- (hopefully rare) cases when the (overestimated) stack use
172 -- exceeds iNTERP_STACK_CHECK_THRESH.
173 maybe_with_stack_check
174 | is_ret && stack_usage < aP_STACK_SPLIM = peep_d
175 -- don't do stack checks at return points,
176 -- everything is aggregated up to the top BCO
177 -- (which must be a function).
178 -- That is, unless the stack usage is >= AP_STACK_SPLIM,
180 | stack_usage >= iNTERP_STACK_CHECK_THRESH
181 = STKCHECK stack_usage : peep_d
183 = peep_d -- the supposedly common case
185 -- We assume that this sum doesn't wrap
186 stack_usage = sum (map bciStackUse peep_d)
188 -- Merge local pushes
189 peep_d = peep (fromOL instrs_ordlist)
191 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
192 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
193 peep (PUSH_L off1 : PUSH_L off2 : rest)
194 = PUSH_LL off1 (off2-1) : peep rest
200 argBits :: [CgRep] -> [Bool]
203 | isFollowableArg rep = False : argBits args
204 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
206 -- -----------------------------------------------------------------------------
209 -- Compile code for the right-hand side of a top-level binding
211 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
214 schemeTopBind (id, rhs)
215 | Just data_con <- isDataConWorkId_maybe id,
216 isNullaryRepDataCon data_con = do
217 -- Special case for the worker of a nullary data con.
218 -- It'll look like this: Nil = /\a -> Nil a
219 -- If we feed it into schemeR, we'll get
221 -- because mkConAppCode treats nullary constructor applications
222 -- by just re-using the single top-level definition. So
223 -- for the worker itself, we must allocate it directly.
224 -- ioToBc (putStrLn $ "top level BCO")
225 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
226 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
229 = schemeR [{- No free variables -}] (id, rhs)
232 -- -----------------------------------------------------------------------------
235 -- Compile code for a right-hand side, to give a BCO that,
236 -- when executed with the free variables and arguments on top of the stack,
237 -- will return with a pointer to the result on top of the stack, after
238 -- removing the free variables and arguments.
240 -- Park the resulting BCO in the monad. Also requires the
241 -- variable to which this value was bound, so as to give the
242 -- resulting BCO a name.
244 schemeR :: [Id] -- Free vars of the RHS, ordered as they
245 -- will appear in the thunk. Empty for
246 -- top-level things, which have no free vars.
247 -> (Id, AnnExpr Id VarSet)
248 -> BcM (ProtoBCO Name)
249 schemeR fvs (nm, rhs)
253 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
254 $$ pprCoreExpr (deAnnotate rhs)
260 = schemeR_wrk fvs nm rhs (collect [] rhs)
262 collect :: [Var] -> AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
263 collect xs (_, AnnNote _ e) = collect xs e
264 collect xs (_, AnnCast e _) = collect xs e
265 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
266 collect 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
350 -- Delegate tail-calls to schemeT.
351 schemeE d s p e@(AnnApp _ _)
354 schemeE d s p e@(AnnVar v)
355 | not (isUnLiftedType v_type)
356 = -- Lifted-type thing; push it in the normal way
360 = do -- Returning an unlifted value.
361 -- Heave it on the stack, SLIDE, and RETURN.
362 (push, szw) <- pushAtom d p (AnnVar v)
363 return (push -- value onto stack
364 `appOL` mkSLIDE szw (d-s) -- clear to sequel
365 `snocOL` RETURN_UBX v_rep) -- go
368 v_rep = typeCgRep v_type
370 schemeE d s p (AnnLit literal)
371 = do (push, szw) <- pushAtom d p (AnnLit literal)
372 let l_rep = typeCgRep (literalType literal)
373 return (push -- value onto stack
374 `appOL` mkSLIDE szw (d-s) -- clear to sequel
375 `snocOL` RETURN_UBX l_rep) -- go
377 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
378 | (AnnVar v, args_r_to_l) <- splitApp rhs,
379 Just data_con <- isDataConWorkId_maybe v,
380 dataConRepArity data_con == length args_r_to_l
381 = do -- Special case for a non-recursive let whose RHS is a
382 -- saturatred constructor application.
383 -- Just allocate the constructor and carry on
384 alloc_code <- mkConAppCode d s p data_con args_r_to_l
385 body_code <- schemeE (d+1) s (addToFM p x d) body
386 return (alloc_code `appOL` body_code)
388 -- General case for let. Generates correct, if inefficient, code in
390 schemeE d s p (AnnLet binds (_,body))
391 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
392 AnnRec xs_n_rhss -> unzip xs_n_rhss
395 fvss = map (fvsToEnv p' . fst) rhss
397 -- Sizes of free vars
398 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
400 -- the arity of each rhs
401 arities = map (length . fst . collect []) rhss
403 -- This p', d' defn is safe because all the items being pushed
404 -- are ptrs, so all have size 1. d' and p' reflect the stack
405 -- after the closures have been allocated in the heap (but not
406 -- filled in), and pointers to them parked on the stack.
407 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
409 zipE = zipEqual "schemeE"
411 -- ToDo: don't build thunks for things with no free variables
412 build_thunk _ [] size bco off arity
413 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
415 mkap | arity == 0 = MKAP
417 build_thunk dd (fv:fvs) size bco off arity = do
418 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
419 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
420 return (push_code `appOL` more_push_code)
422 alloc_code = toOL (zipWith mkAlloc sizes arities)
424 | is_tick = ALLOC_AP_NOUPD sz
425 | otherwise = ALLOC_AP sz
426 mkAlloc sz arity = ALLOC_PAP arity sz
428 is_tick = case binds of
429 AnnNonRec id _ -> occNameFS (getOccName id) == tickFS
432 compile_bind d' fvs x rhs size arity off = do
433 bco <- schemeR fvs (x,rhs)
434 build_thunk d' fvs size bco off arity
437 [ compile_bind d' fvs x rhs size arity n
438 | (fvs, x, rhs, size, arity, n) <-
439 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
442 body_code <- schemeE d' s p' body
443 thunk_codes <- sequence compile_binds
444 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
446 -- introduce a let binding for a ticked case expression. This rule
447 -- *should* only fire when the expression was not already let-bound
448 -- (the code gen for let bindings should take care of that). Todo: we
449 -- call exprFreeVars on a deAnnotated expression, this may not be the
450 -- best way to calculate the free vars but it seemed like the least
451 -- intrusive thing to do
452 schemeE d s p exp@(AnnCase {})
453 | Just (_tickInfo, rhs) <- isTickedExp' exp
454 = if isUnLiftedType ty
455 then schemeE d s p (snd rhs)
458 -- Todo: is emptyVarSet correct on the next line?
459 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
461 where exp' = deAnnotate' exp
462 fvs = exprFreeVars exp'
465 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1, bind2], rhs)])
466 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
468 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
470 -- case .... of a { DEFAULT -> ... }
471 -- becuse the return convention for both are identical.
473 -- Note that it does not matter losing the void-rep thing from the
474 -- envt (it won't be bound now) because we never look such things up.
476 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
477 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
479 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
480 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
481 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
483 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1], rhs)])
484 | isUnboxedTupleCon dc
485 -- Similarly, convert
486 -- case .... of x { (# a #) -> ... }
488 -- case .... of a { DEFAULT -> ... }
489 = --trace "automagic mashing of case alts (# a #)" $
490 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
492 schemeE d s p (AnnCase scrut bndr _ alts)
493 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
495 schemeE d s p (AnnNote _ (_, body))
498 schemeE d s p (AnnCast (_, body) _)
502 = pprPanic "ByteCodeGen.schemeE: unhandled case"
503 (pprCoreExpr (deAnnotate' expr))
509 A ticked expression looks like this:
511 case tick<n> var1 ... varN of DEFAULT -> e
513 (*) <n> is the number of the tick, which is unique within a module
514 (*) var1 ... varN are the local variables in scope at the tick site
516 If we find a ticked expression we return:
518 Just ((n, [var1 ... varN]), e)
520 otherwise we return Nothing.
522 The idea is that the "case tick<n> ..." is really just an annotation on
523 the code. When we find such a thing, we pull out the useful information,
524 and then compile the code as if it was just the expression "e".
528 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
529 isTickedExp' (AnnCase scrut _bndr _type alts)
530 | Just tickInfo <- isTickedScrut scrut,
531 [(DEFAULT, _bndr, rhs)] <- alts
532 = Just (tickInfo, rhs)
534 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
537 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
538 = Just $ TickInfo { tickInfo_number = tickNumber
539 , tickInfo_module = modName
540 , tickInfo_locals = idsOfArgs args
542 | otherwise = Nothing
544 (f, args) = collectArgs $ deAnnotate expr
545 idsOfArgs :: [Expr Id] -> [Id]
546 idsOfArgs = catMaybes . map exprId
547 exprId :: Expr Id -> Maybe Id
548 exprId (Var id) = Just id
551 isTickedExp' _ = Nothing
553 -- Compile code to do a tail call. Specifically, push the fn,
554 -- slide the on-stack app back down to the sequel depth,
555 -- and enter. Four cases:
558 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
559 -- The int will be on the stack. Generate a code sequence
560 -- to convert it to the relevant constructor, SLIDE and ENTER.
562 -- 1. The fn denotes a ccall. Defer to generateCCall.
564 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
565 -- it simply as b -- since the representations are identical
566 -- (the VoidArg takes up zero stack space). Also, spot
567 -- (# b #) and treat it as b.
569 -- 3. Application of a constructor, by defn saturated.
570 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
571 -- then the ptrs, and then do PACK and RETURN.
573 -- 4. Otherwise, it must be a function call. Push the args
574 -- right to left, SLIDE and ENTER.
576 schemeT :: Int -- Stack depth
577 -> Sequel -- Sequel depth
578 -> BCEnv -- stack env
579 -> AnnExpr' Id VarSet
584 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
585 -- = panic "schemeT ?!?!"
587 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
591 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
592 = do (push, arg_words) <- pushAtom d p arg
593 tagToId_sequence <- implement_tagToId constr_names
594 return (push `appOL` tagToId_sequence
595 `appOL` mkSLIDE 1 (d+arg_words-s)
599 | Just (CCall ccall_spec) <- isFCallId_maybe fn
600 = generateCCall d s p ccall_spec fn args_r_to_l
602 -- Case 2: Constructor application
603 | Just con <- maybe_saturated_dcon,
604 isUnboxedTupleCon con
605 = case args_r_to_l of
606 [arg1,arg2] | isVoidArgAtom arg1 ->
607 unboxedTupleReturn d s p arg2
608 [arg1,arg2] | isVoidArgAtom arg2 ->
609 unboxedTupleReturn d s p arg1
610 _other -> unboxedTupleException
612 -- Case 3: Ordinary data constructor
613 | Just con <- maybe_saturated_dcon
614 = do alloc_con <- mkConAppCode d s p con args_r_to_l
615 return (alloc_con `appOL`
616 mkSLIDE 1 (d - s) `snocOL`
619 -- Case 4: Tail call of function
621 = doTailCall d s p fn args_r_to_l
624 -- Detect and extract relevant info for the tagToEnum kludge.
625 maybe_is_tagToEnum_call
626 = let extract_constr_Names ty
627 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
629 = map (getName . dataConWorkId) (tyConDataCons tyc)
630 -- NOTE: use the worker name, not the source name of
631 -- the DataCon. See DataCon.lhs for details.
633 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
636 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
637 -> case isPrimOpId_maybe v of
638 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
642 -- Extract the args (R->L) and fn
643 -- The function will necessarily be a variable,
644 -- because we are compiling a tail call
645 (AnnVar fn, args_r_to_l) = splitApp app
647 -- Only consider this to be a constructor application iff it is
648 -- saturated. Otherwise, we'll call the constructor wrapper.
649 n_args = length args_r_to_l
651 = case isDataConWorkId_maybe fn of
652 Just con | dataConRepArity con == n_args -> Just con
655 -- -----------------------------------------------------------------------------
656 -- Generate code to build a constructor application,
657 -- leaving it on top of the stack
659 mkConAppCode :: Int -> Sequel -> BCEnv
660 -> DataCon -- The data constructor
661 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
664 mkConAppCode _ _ _ con [] -- Nullary constructor
665 = ASSERT( isNullaryRepDataCon con )
666 return (unitOL (PUSH_G (getName (dataConWorkId con))))
667 -- Instead of doing a PACK, which would allocate a fresh
668 -- copy of this constructor, use the single shared version.
670 mkConAppCode orig_d _ p con args_r_to_l
671 = ASSERT( dataConRepArity con == length args_r_to_l )
672 do_pushery orig_d (non_ptr_args ++ ptr_args)
674 -- The args are already in reverse order, which is the way PACK
675 -- expects them to be. We must push the non-ptrs after the ptrs.
676 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
678 do_pushery d (arg:args)
679 = do (push, arg_words) <- pushAtom d p arg
680 more_push_code <- do_pushery (d+arg_words) args
681 return (push `appOL` more_push_code)
683 = return (unitOL (PACK con n_arg_words))
685 n_arg_words = d - orig_d
688 -- -----------------------------------------------------------------------------
689 -- Returning an unboxed tuple with one non-void component (the only
690 -- case we can handle).
692 -- Remember, we don't want to *evaluate* the component that is being
693 -- returned, even if it is a pointed type. We always just return.
696 :: Int -> Sequel -> BCEnv
697 -> AnnExpr' Id VarSet -> BcM BCInstrList
698 unboxedTupleReturn d s p arg = do
699 (push, sz) <- pushAtom d p arg
701 mkSLIDE sz (d-s) `snocOL`
702 RETURN_UBX (atomRep arg))
704 -- -----------------------------------------------------------------------------
705 -- Generate code for a tail-call
708 :: Int -> Sequel -> BCEnv
709 -> Id -> [AnnExpr' Id VarSet]
711 doTailCall init_d s p fn args
712 = do_pushes init_d args (map atomRep args)
714 do_pushes d [] reps = do
715 ASSERT( null reps ) return ()
716 (push_fn, sz) <- pushAtom d p (AnnVar fn)
717 ASSERT( sz == 1 ) return ()
718 return (push_fn `appOL` (
719 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
721 do_pushes d args reps = do
722 let (push_apply, n, rest_of_reps) = findPushSeq reps
723 (these_args, rest_of_args) = splitAt n args
724 (next_d, push_code) <- push_seq d these_args
725 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
726 -- ^^^ for the PUSH_APPLY_ instruction
727 return (push_code `appOL` (push_apply `consOL` instrs))
729 push_seq d [] = return (d, nilOL)
730 push_seq d (arg:args) = do
731 (push_code, sz) <- pushAtom d p arg
732 (final_d, more_push_code) <- push_seq (d+sz) args
733 return (final_d, push_code `appOL` more_push_code)
735 -- v. similar to CgStackery.findMatch, ToDo: merge
736 findPushSeq :: [CgRep] -> (BCInstr, Int, [CgRep])
737 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
738 = (PUSH_APPLY_PPPPPP, 6, rest)
739 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
740 = (PUSH_APPLY_PPPPP, 5, rest)
741 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
742 = (PUSH_APPLY_PPPP, 4, rest)
743 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
744 = (PUSH_APPLY_PPP, 3, rest)
745 findPushSeq (PtrArg: PtrArg: rest)
746 = (PUSH_APPLY_PP, 2, rest)
747 findPushSeq (PtrArg: rest)
748 = (PUSH_APPLY_P, 1, rest)
749 findPushSeq (VoidArg: rest)
750 = (PUSH_APPLY_V, 1, rest)
751 findPushSeq (NonPtrArg: rest)
752 = (PUSH_APPLY_N, 1, rest)
753 findPushSeq (FloatArg: rest)
754 = (PUSH_APPLY_F, 1, rest)
755 findPushSeq (DoubleArg: rest)
756 = (PUSH_APPLY_D, 1, rest)
757 findPushSeq (LongArg: rest)
758 = (PUSH_APPLY_L, 1, rest)
760 = panic "ByteCodeGen.findPushSeq"
762 -- -----------------------------------------------------------------------------
765 doCase :: Int -> Sequel -> BCEnv
766 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
767 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
769 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
771 -- Top of stack is the return itbl, as usual.
772 -- underneath it is the pointer to the alt_code BCO.
773 -- When an alt is entered, it assumes the returned value is
774 -- on top of the itbl.
777 -- An unlifted value gets an extra info table pushed on top
778 -- when it is returned.
779 unlifted_itbl_sizeW | isAlgCase = 0
782 -- depth of stack after the return value has been pushed
783 d_bndr = d + ret_frame_sizeW + idSizeW bndr
785 -- depth of stack after the extra info table for an unboxed return
786 -- has been pushed, if any. This is the stack depth at the
788 d_alts = d_bndr + unlifted_itbl_sizeW
790 -- Env in which to compile the alts, not including
791 -- any vars bound by the alts themselves
792 p_alts = addToFM p bndr (d_bndr - 1)
794 bndr_ty = idType bndr
795 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
797 -- given an alt, return a discr and code for it.
798 codeAlt (DEFAULT, _, (_,rhs))
799 = do rhs_code <- schemeE d_alts s p_alts rhs
800 return (NoDiscr, rhs_code)
802 codeAlt alt@(_, bndrs, (_,rhs))
803 -- primitive or nullary constructor alt: no need to UNPACK
804 | null real_bndrs = do
805 rhs_code <- schemeE d_alts s p_alts rhs
806 return (my_discr alt, rhs_code)
807 -- algebraic alt with some binders
810 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
811 ptr_sizes = map idSizeW ptrs
812 nptrs_sizes = map idSizeW nptrs
813 bind_sizes = ptr_sizes ++ nptrs_sizes
814 size = sum ptr_sizes + sum nptrs_sizes
815 -- the UNPACK instruction unpacks in reverse order...
816 p' = addListToFM p_alts
817 (zip (reverse (ptrs ++ nptrs))
818 (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, _, _) = NoDiscr {-shouldn't really happen-}
827 my_discr (DataAlt dc, _, _)
828 | isUnboxedTupleCon dc
829 = unboxedTupleException
831 = DiscrP (dataConTag dc - fIRST_TAG)
832 my_discr (LitAlt l, _, _)
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 (CCallSpec target cconv _) 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 _ [] = 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,AddrRep):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,AddrRep):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, atomPrimRep 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 primRepSizeW a_reps_pushed_r_to_l)
957 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
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, VoidRep)
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 = primRepSizeW 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
1044 -- Offset of the next stack frame down the stack. The CCALL
1045 -- instruction needs to describe the chunk of stack containing
1046 -- the ccall args to the GC, so it needs to know how large it
1047 -- is. See comment in Interpreter.c with the CCALL instruction.
1048 stk_offset = d_after_r - s
1051 -- the only difference in libffi mode is that we prepare a cif
1052 -- describing the call type by calling libffi, and we attach the
1053 -- address of this to the CCALL instruction.
1054 token <- ioToBc $ prepForeignCall cconv a_reps r_rep
1055 let addr_of_marshaller = castPtrToFunPtr token
1057 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1060 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1062 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1063 `snocOL` RETURN_UBX (primRepToCgRep r_rep)
1065 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1068 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1071 -- Make a dummy literal, to be used as a placeholder for FFI return
1072 -- values on the stack.
1073 mkDummyLiteral :: PrimRep -> Literal
1077 WordRep -> MachWord 0
1078 AddrRep -> MachNullAddr
1079 DoubleRep -> MachDouble 0
1080 FloatRep -> MachFloat 0
1081 Int64Rep -> MachInt64 0
1082 Word64Rep -> MachWord64 0
1083 _ -> panic "mkDummyLiteral"
1087 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1088 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1091 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1093 -- Alternatively, for call-targets returning nothing, convert
1095 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1096 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1100 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
1101 maybe_getCCallReturnRep fn_ty
1102 = let (_a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1104 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1106 = case splitTyConApp_maybe (repType r_ty) of
1107 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
1109 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
1110 || r_reps == [VoidRep] )
1111 && isUnboxedTupleTyCon r_tycon
1112 && case maybe_r_rep_to_go of
1114 Just r_rep -> r_rep /= PtrRep
1115 -- if it was, it would be impossible
1116 -- to create a valid return value
1117 -- placeholder on the stack
1118 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1121 --trace (showSDoc (ppr (a_reps, r_reps))) $
1122 if ok then maybe_r_rep_to_go else blargh
1124 -- Compile code which expects an unboxed Int on the top of stack,
1125 -- (call it i), and pushes the i'th closure in the supplied list
1126 -- as a consequence.
1127 implement_tagToId :: [Name] -> BcM BCInstrList
1128 implement_tagToId names
1129 = ASSERT( notNull names )
1130 do labels <- getLabelsBc (length names)
1131 label_fail <- getLabelBc
1132 label_exit <- getLabelBc
1133 let infos = zip4 labels (tail labels ++ [label_fail])
1135 steps = map (mkStep label_exit) infos
1136 return (concatOL steps
1138 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1140 mkStep l_exit (my_label, next_label, n, name_for_n)
1141 = toOL [LABEL my_label,
1142 TESTEQ_I n next_label,
1147 -- -----------------------------------------------------------------------------
1150 -- Push an atom onto the stack, returning suitable code & number of
1151 -- stack words used.
1153 -- The env p must map each variable to the highest- numbered stack
1154 -- slot for it. For example, if the stack has depth 4 and we
1155 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1156 -- the tag in stack[5], the stack will have depth 6, and p must map v
1157 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1158 -- depth 6 stack has valid words 0 .. 5.
1160 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1162 pushAtom d p (AnnApp f (_, AnnType _))
1163 = pushAtom d p (snd f)
1165 pushAtom d p (AnnNote _ e)
1166 = pushAtom d p (snd e)
1168 pushAtom d p (AnnLam x e)
1170 = pushAtom d p (snd e)
1172 pushAtom d p (AnnVar v)
1174 | idCgRep v == VoidArg
1178 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1180 | Just primop <- isPrimOpId_maybe v
1181 = return (unitOL (PUSH_PRIMOP primop), 1)
1183 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1184 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1185 -- d - d_v the number of words between the TOS
1186 -- and the 1st slot of the object
1188 -- d - d_v - 1 the offset from the TOS of the 1st slot
1190 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1193 -- Having found the last slot, we proceed to copy the right number of
1194 -- slots on to the top of the stack.
1196 | otherwise -- v must be a global variable
1198 return (unitOL (PUSH_G (getName v)), sz)
1204 pushAtom _ _ (AnnLit lit)
1206 MachLabel _ _ -> code NonPtrArg
1207 MachWord _ -> code NonPtrArg
1208 MachInt _ -> code PtrArg
1209 MachFloat _ -> code FloatArg
1210 MachDouble _ -> code DoubleArg
1211 MachChar _ -> code NonPtrArg
1212 MachStr s -> pushStr s
1213 l -> pprPanic "pushAtom" (ppr l)
1216 = let size_host_words = cgRepSizeW rep
1217 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1221 = let getMallocvilleAddr
1223 FastString _ n _ fp _ ->
1224 -- we could grab the Ptr from the ForeignPtr,
1225 -- but then we have no way to control its lifetime.
1226 -- In reality it'll probably stay alive long enoungh
1227 -- by virtue of the global FastString table, but
1228 -- to be on the safe side we copy the string into
1229 -- a malloc'd area of memory.
1230 do ptr <- ioToBc (mallocBytes (n+1))
1233 withForeignPtr fp $ \p -> do
1234 memcpy ptr p (fromIntegral n)
1235 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1239 addr <- getMallocvilleAddr
1240 -- Get the addr on the stack, untaggedly
1241 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1243 pushAtom d p (AnnCast e _)
1244 = pushAtom d p (snd e)
1247 = pprPanic "ByteCodeGen.pushAtom"
1248 (pprCoreExpr (deAnnotate (undefined, expr)))
1250 foreign import ccall unsafe "memcpy"
1251 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1254 -- -----------------------------------------------------------------------------
1255 -- Given a bunch of alts code and their discrs, do the donkey work
1256 -- of making a multiway branch using a switch tree.
1257 -- What a load of hassle!
1259 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1260 -- a hint; generates better code
1261 -- Nothing is always safe
1262 -> [(Discr, BCInstrList)]
1264 mkMultiBranch maybe_ncons raw_ways
1265 = let d_way = filter (isNoDiscr.fst) raw_ways
1267 (\w1 w2 -> leAlt (fst w1) (fst w2))
1268 (filter (not.isNoDiscr.fst) raw_ways)
1270 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1271 mkTree [] _range_lo _range_hi = return the_default
1273 mkTree [val] range_lo range_hi
1274 | range_lo `eqAlt` range_hi
1277 = do label_neq <- getLabelBc
1278 return (mkTestEQ (fst val) label_neq
1280 `appOL` unitOL (LABEL label_neq)
1281 `appOL` the_default))
1283 mkTree vals range_lo range_hi
1284 = let n = length vals `div` 2
1285 vals_lo = take n vals
1286 vals_hi = drop n vals
1287 v_mid = fst (head vals_hi)
1289 label_geq <- getLabelBc
1290 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1291 code_hi <- mkTree vals_hi v_mid range_hi
1292 return (mkTestLT v_mid label_geq
1294 `appOL` unitOL (LABEL label_geq)
1298 = case d_way of [] -> unitOL CASEFAIL
1300 _ -> panic "mkMultiBranch/the_default"
1302 -- None of these will be needed if there are no non-default alts
1303 (mkTestLT, mkTestEQ, init_lo, init_hi)
1305 = panic "mkMultiBranch: awesome foursome"
1307 = case fst (head notd_ways) of {
1308 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1309 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1312 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1313 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1316 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1317 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1320 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1321 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1323 DiscrP algMaxBound );
1324 NoDiscr -> panic "mkMultiBranch NoDiscr"
1327 (algMinBound, algMaxBound)
1328 = case maybe_ncons of
1329 Just n -> (0, n - 1)
1330 Nothing -> (minBound, maxBound)
1332 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1333 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1334 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1335 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1336 NoDiscr `eqAlt` NoDiscr = True
1339 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1340 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1341 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1342 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1343 NoDiscr `leAlt` NoDiscr = True
1346 isNoDiscr NoDiscr = True
1349 dec (DiscrI i) = DiscrI (i-1)
1350 dec (DiscrP i) = DiscrP (i-1)
1351 dec other = other -- not really right, but if you
1352 -- do cases on floating values, you'll get what you deserve
1354 -- same snotty comment applies to the following
1356 minD, maxD :: Double
1362 mkTree notd_ways init_lo init_hi
1365 -- -----------------------------------------------------------------------------
1366 -- Supporting junk for the compilation schemes
1368 -- Describes case alts
1376 instance Outputable Discr where
1377 ppr (DiscrI i) = int i
1378 ppr (DiscrF f) = text (show f)
1379 ppr (DiscrD d) = text (show d)
1380 ppr (DiscrP i) = int i
1381 ppr NoDiscr = text "DEF"
1384 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1385 lookupBCEnv_maybe = lookupFM
1387 idSizeW :: Id -> Int
1388 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1391 unboxedTupleException :: a
1392 unboxedTupleException
1395 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1396 " Possibly due to foreign import/export decls in source.\n"++
1397 " Workaround: use -fobject-code, or compile this module to .o separately."))
1400 mkSLIDE :: Int -> Int -> OrdList BCInstr
1401 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1403 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1404 -- The arguments are returned in *right-to-left* order
1405 splitApp (AnnApp (_,f) (_,a))
1406 | isTypeAtom a = splitApp f
1407 | otherwise = case splitApp f of
1408 (f', as) -> (f', a:as)
1409 splitApp (AnnNote _ (_,e)) = splitApp e
1410 splitApp (AnnCast (_,e) _) = splitApp e
1411 splitApp e = (e, [])
1414 isTypeAtom :: AnnExpr' id ann -> Bool
1415 isTypeAtom (AnnType _) = True
1416 isTypeAtom _ = False
1418 isVoidArgAtom :: AnnExpr' id ann -> Bool
1419 isVoidArgAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1420 isVoidArgAtom (AnnNote _ (_,e)) = isVoidArgAtom e
1421 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1422 isVoidArgAtom _ = False
1424 atomPrimRep :: AnnExpr' Id ann -> PrimRep
1425 atomPrimRep (AnnVar v) = typePrimRep (idType v)
1426 atomPrimRep (AnnLit l) = typePrimRep (literalType l)
1427 atomPrimRep (AnnNote _ b) = atomPrimRep (snd b)
1428 atomPrimRep (AnnApp f (_, AnnType _)) = atomPrimRep (snd f)
1429 atomPrimRep (AnnLam x e) | isTyVar x = atomPrimRep (snd e)
1430 atomPrimRep (AnnCast b _) = atomPrimRep (snd b)
1431 atomPrimRep other = pprPanic "atomPrimRep" (ppr (deAnnotate (undefined,other)))
1433 atomRep :: AnnExpr' Id ann -> CgRep
1434 atomRep e = primRepToCgRep (atomPrimRep e)
1436 isPtrAtom :: AnnExpr' Id ann -> Bool
1437 isPtrAtom e = atomRep e == PtrArg
1439 -- Let szsw be the sizes in words of some items pushed onto the stack,
1440 -- which has initial depth d'. Return the values which the stack environment
1441 -- should map these items to.
1442 mkStackOffsets :: Int -> [Int] -> [Int]
1443 mkStackOffsets original_depth szsw
1444 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1446 -- -----------------------------------------------------------------------------
1447 -- The bytecode generator's monad
1449 type BcPtr = Either ItblPtr (Ptr ())
1453 uniqSupply :: UniqSupply, -- for generating fresh variable names
1454 nextlabel :: Int, -- for generating local labels
1455 malloced :: [BcPtr], -- thunks malloced for current BCO
1456 -- Should be free()d when it is GCd
1457 breakArray :: BreakArray -- array of breakpoint flags
1460 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1462 ioToBc :: IO a -> BcM a
1463 ioToBc io = BcM $ \st -> do
1467 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1468 runBc us modBreaks (BcM m)
1469 = m (BcM_State us 0 [] breakArray)
1471 breakArray = modBreaks_flags modBreaks
1473 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1474 thenBc (BcM expr) cont = BcM $ \st0 -> do
1475 (st1, q) <- expr st0
1480 thenBc_ :: BcM a -> BcM b -> BcM b
1481 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1482 (st1, _) <- expr st0
1483 (st2, r) <- cont st1
1486 returnBc :: a -> BcM a
1487 returnBc result = BcM $ \st -> (return (st, result))
1489 instance Monad BcM where
1494 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1496 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1498 recordMallocBc :: Ptr a -> BcM ()
1500 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1502 recordItblMallocBc :: ItblPtr -> BcM ()
1503 recordItblMallocBc a
1504 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1506 getLabelBc :: BcM Int
1508 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1510 getLabelsBc :: Int -> BcM [Int]
1512 = BcM $ \st -> let ctr = nextlabel st
1513 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1515 getBreakArray :: BcM BreakArray
1516 getBreakArray = BcM $ \st -> return (st, breakArray st)
1518 newUnique :: BcM Unique
1520 \st -> case splitUniqSupply (uniqSupply st) of
1521 (us1, us2) -> let newState = st { uniqSupply = us2 }
1522 in return (newState, uniqFromSupply us1)
1524 newId :: Type -> BcM Id
1527 return $ mkSysLocal tickFS uniq ty
1529 tickFS :: FastString
1530 tickFS = fsLit "ticked"