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"
61 import qualified Data.Map as Map
62 import qualified FiniteMap as Map
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 dflags 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 dflags proto_bco
120 -- -----------------------------------------------------------------------------
121 -- Compilation schema for the bytecode generator
123 type BCInstrList = OrdList BCInstr
125 type Sequel = Word16 -- 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 = Map Id Word16 -- To find vars on the stack
132 ppBCEnv :: BCEnv -> SDoc
135 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (Map.toList 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 < fromIntegral 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 >= fromIntegral 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.isTyCoVar).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 (fromIntegral . idSizeW) all_args
278 szw_args = sum szsw_args
279 p_init = Map.fromList (zip all_args (mkStackOffsets 0 szsw_args))
281 -- make the arg bitmap
282 bits = argBits (reverse (map idCgRep all_args))
283 bitmap_size = genericLength 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 :: Word16 -> 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
307 BRK_FUN arr# (fromIntegral tickNumber) breakInfo
308 return $ breakInstr `consOL` code
309 | otherwise = schemeE d 0 p rhs
311 getVarOffSets :: Word16 -> BCEnv -> TickInfo -> [(Id, Word16)]
312 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
314 getOffSet :: Word16 -> BCEnv -> Id -> Maybe (Id, Word16)
316 = case lookupBCEnv_maybe id env of
318 Just offset -> Just (id, d - offset)
320 fvsToEnv :: BCEnv -> VarSet -> [Id]
321 -- Takes the free variables of a right-hand side, and
322 -- delivers an ordered list of the local variables that will
323 -- be captured in the thunk for the RHS
324 -- The BCEnv argument tells which variables are in the local
325 -- environment: these are the ones that should be captured
327 -- The code that constructs the thunk, and the code that executes
328 -- it, have to agree about this layout
329 fvsToEnv p fvs = [v | v <- varSetElems fvs,
330 isId v, -- Could be a type variable
333 -- -----------------------------------------------------------------------------
338 { tickInfo_number :: Int -- the (module) unique number of the tick
339 , tickInfo_module :: Module -- the origin of the ticked expression
340 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
343 instance Outputable TickInfo where
344 ppr info = text "TickInfo" <+>
345 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
346 ppr (tickInfo_locals info))
348 -- Compile code to apply the given expression to the remaining args
349 -- on the stack, returning a HNF.
350 schemeE :: Word16 -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
353 | Just e' <- bcView e
356 -- Delegate tail-calls to schemeT.
357 schemeE d s p e@(AnnApp _ _)
360 schemeE d s p e@(AnnVar v)
361 | not (isUnLiftedType v_type)
362 = -- Lifted-type thing; push it in the normal way
366 = do -- Returning an unlifted value.
367 -- Heave it on the stack, SLIDE, and RETURN.
368 (push, szw) <- pushAtom d p (AnnVar v)
369 return (push -- value onto stack
370 `appOL` mkSLIDE szw (d-s) -- clear to sequel
371 `snocOL` RETURN_UBX v_rep) -- go
374 v_rep = typeCgRep v_type
376 schemeE d s p (AnnLit literal)
377 = do (push, szw) <- pushAtom d p (AnnLit literal)
378 let l_rep = typeCgRep (literalType literal)
379 return (push -- value onto stack
380 `appOL` mkSLIDE szw (d-s) -- clear to sequel
381 `snocOL` RETURN_UBX l_rep) -- go
383 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
384 | (AnnVar v, args_r_to_l) <- splitApp rhs,
385 Just data_con <- isDataConWorkId_maybe v,
386 dataConRepArity data_con == length args_r_to_l
387 = do -- Special case for a non-recursive let whose RHS is a
388 -- saturatred constructor application.
389 -- Just allocate the constructor and carry on
390 alloc_code <- mkConAppCode d s p data_con args_r_to_l
391 body_code <- schemeE (d+1) s (Map.insert x d p) body
392 return (alloc_code `appOL` body_code)
394 -- General case for let. Generates correct, if inefficient, code in
396 schemeE d s p (AnnLet binds (_,body))
397 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
398 AnnRec xs_n_rhss -> unzip xs_n_rhss
399 n_binds = genericLength xs
401 fvss = map (fvsToEnv p' . fst) rhss
403 -- Sizes of free vars
404 sizes = map (\rhs_fvs -> sum (map (fromIntegral . idSizeW) rhs_fvs)) fvss
406 -- the arity of each rhs
407 arities = map (genericLength . fst . collect) rhss
409 -- This p', d' defn is safe because all the items being pushed
410 -- are ptrs, so all have size 1. d' and p' reflect the stack
411 -- after the closures have been allocated in the heap (but not
412 -- filled in), and pointers to them parked on the stack.
413 p' = Map.insertList (zipE xs (mkStackOffsets d (genericReplicate n_binds 1))) p
415 zipE = zipEqual "schemeE"
417 -- ToDo: don't build thunks for things with no free variables
418 build_thunk _ [] size bco off arity
419 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
421 mkap | arity == 0 = MKAP
423 build_thunk dd (fv:fvs) size bco off arity = do
424 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
425 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
426 return (push_code `appOL` more_push_code)
428 alloc_code = toOL (zipWith mkAlloc sizes arities)
430 | is_tick = ALLOC_AP_NOUPD sz
431 | otherwise = ALLOC_AP sz
432 mkAlloc sz arity = ALLOC_PAP arity sz
434 is_tick = case binds of
435 AnnNonRec id _ -> occNameFS (getOccName id) == tickFS
438 compile_bind d' fvs x rhs size arity off = do
439 bco <- schemeR fvs (x,rhs)
440 build_thunk d' fvs size bco off arity
443 [ compile_bind d' fvs x rhs size arity n
444 | (fvs, x, rhs, size, arity, n) <-
445 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
448 body_code <- schemeE d' s p' body
449 thunk_codes <- sequence compile_binds
450 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
452 -- introduce a let binding for a ticked case expression. This rule
453 -- *should* only fire when the expression was not already let-bound
454 -- (the code gen for let bindings should take care of that). Todo: we
455 -- call exprFreeVars on a deAnnotated expression, this may not be the
456 -- best way to calculate the free vars but it seemed like the least
457 -- intrusive thing to do
458 schemeE d s p exp@(AnnCase {})
459 | Just (_tickInfo, _rhs) <- isTickedExp' exp
460 = if isUnLiftedType ty
462 -- If the result type is unlifted, then we must generate
463 -- let f = \s . case tick# of _ -> e
465 -- When we stop at the breakpoint, _result will have an unlifted
466 -- type and hence won't be bound in the environment, but the
467 -- breakpoint will otherwise work fine.
468 id <- newId (mkFunTy realWorldStatePrimTy ty)
469 st <- newId realWorldStatePrimTy
470 let letExp = AnnLet (AnnNonRec id (fvs, AnnLam st (emptyVarSet, exp)))
471 (emptyVarSet, (AnnApp (emptyVarSet, AnnVar id)
472 (emptyVarSet, AnnVar realWorldPrimId)))
476 -- Todo: is emptyVarSet correct on the next line?
477 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
479 where exp' = deAnnotate' exp
480 fvs = exprFreeVars exp'
483 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1, bind2], rhs)])
484 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
486 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
488 -- case .... of a { DEFAULT -> ... }
489 -- becuse the return convention for both are identical.
491 -- Note that it does not matter losing the void-rep thing from the
492 -- envt (it won't be bound now) because we never look such things up.
494 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
495 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
497 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
498 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
499 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
501 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1], rhs)])
502 | isUnboxedTupleCon dc
503 -- Similarly, convert
504 -- case .... of x { (# a #) -> ... }
506 -- case .... of a { DEFAULT -> ... }
507 = --trace "automagic mashing of case alts (# a #)" $
508 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
510 schemeE d s p (AnnCase scrut bndr _ alts)
511 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
514 = pprPanic "ByteCodeGen.schemeE: unhandled case"
515 (pprCoreExpr (deAnnotate' expr))
521 A ticked expression looks like this:
523 case tick<n> var1 ... varN of DEFAULT -> e
525 (*) <n> is the number of the tick, which is unique within a module
526 (*) var1 ... varN are the local variables in scope at the tick site
528 If we find a ticked expression we return:
530 Just ((n, [var1 ... varN]), e)
532 otherwise we return Nothing.
534 The idea is that the "case tick<n> ..." is really just an annotation on
535 the code. When we find such a thing, we pull out the useful information,
536 and then compile the code as if it was just the expression "e".
540 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
541 isTickedExp' (AnnCase scrut _bndr _type alts)
542 | Just tickInfo <- isTickedScrut scrut,
543 [(DEFAULT, _bndr, rhs)] <- alts
544 = Just (tickInfo, rhs)
546 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
549 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
550 = Just $ TickInfo { tickInfo_number = tickNumber
551 , tickInfo_module = modName
552 , tickInfo_locals = idsOfArgs args
554 | otherwise = Nothing
556 (f, args) = collectArgs $ deAnnotate expr
557 idsOfArgs :: [Expr Id] -> [Id]
558 idsOfArgs = catMaybes . map exprId
559 exprId :: Expr Id -> Maybe Id
560 exprId (Var id) = Just id
563 isTickedExp' _ = Nothing
565 -- Compile code to do a tail call. Specifically, push the fn,
566 -- slide the on-stack app back down to the sequel depth,
567 -- and enter. Four cases:
570 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
571 -- The int will be on the stack. Generate a code sequence
572 -- to convert it to the relevant constructor, SLIDE and ENTER.
574 -- 1. The fn denotes a ccall. Defer to generateCCall.
576 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
577 -- it simply as b -- since the representations are identical
578 -- (the VoidArg takes up zero stack space). Also, spot
579 -- (# b #) and treat it as b.
581 -- 3. Application of a constructor, by defn saturated.
582 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
583 -- then the ptrs, and then do PACK and RETURN.
585 -- 4. Otherwise, it must be a function call. Push the args
586 -- right to left, SLIDE and ENTER.
588 schemeT :: Word16 -- Stack depth
589 -> Sequel -- Sequel depth
590 -> BCEnv -- stack env
591 -> AnnExpr' Id VarSet
596 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
597 -- = panic "schemeT ?!?!"
599 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
603 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
604 = do (push, arg_words) <- pushAtom d p arg
605 tagToId_sequence <- implement_tagToId constr_names
606 return (push `appOL` tagToId_sequence
607 `appOL` mkSLIDE 1 (d+arg_words-s)
611 | Just (CCall ccall_spec) <- isFCallId_maybe fn
612 = generateCCall d s p ccall_spec fn args_r_to_l
614 -- Case 2: Constructor application
615 | Just con <- maybe_saturated_dcon,
616 isUnboxedTupleCon con
617 = case args_r_to_l of
618 [arg1,arg2] | isVoidArgAtom arg1 ->
619 unboxedTupleReturn d s p arg2
620 [arg1,arg2] | isVoidArgAtom arg2 ->
621 unboxedTupleReturn d s p arg1
622 _other -> unboxedTupleException
624 -- Case 3: Ordinary data constructor
625 | Just con <- maybe_saturated_dcon
626 = do alloc_con <- mkConAppCode d s p con args_r_to_l
627 return (alloc_con `appOL`
628 mkSLIDE 1 (d - s) `snocOL`
631 -- Case 4: Tail call of function
633 = doTailCall d s p fn args_r_to_l
636 -- Detect and extract relevant info for the tagToEnum kludge.
637 maybe_is_tagToEnum_call
638 = let extract_constr_Names ty
639 | Just (tyc, _) <- splitTyConApp_maybe (repType ty),
641 = map (getName . dataConWorkId) (tyConDataCons tyc)
642 -- NOTE: use the worker name, not the source name of
643 -- the DataCon. See DataCon.lhs for details.
645 = pprPanic "maybe_is_tagToEnum_call.extract_constr_Ids" (ppr ty)
648 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
649 -> case isPrimOpId_maybe v of
650 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
654 -- Extract the args (R->L) and fn
655 -- The function will necessarily be a variable,
656 -- because we are compiling a tail call
657 (AnnVar fn, args_r_to_l) = splitApp app
659 -- Only consider this to be a constructor application iff it is
660 -- saturated. Otherwise, we'll call the constructor wrapper.
661 n_args = length args_r_to_l
663 = case isDataConWorkId_maybe fn of
664 Just con | dataConRepArity con == n_args -> Just con
667 -- -----------------------------------------------------------------------------
668 -- Generate code to build a constructor application,
669 -- leaving it on top of the stack
671 mkConAppCode :: Word16 -> Sequel -> BCEnv
672 -> DataCon -- The data constructor
673 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
676 mkConAppCode _ _ _ con [] -- Nullary constructor
677 = ASSERT( isNullaryRepDataCon con )
678 return (unitOL (PUSH_G (getName (dataConWorkId con))))
679 -- Instead of doing a PACK, which would allocate a fresh
680 -- copy of this constructor, use the single shared version.
682 mkConAppCode orig_d _ p con args_r_to_l
683 = ASSERT( dataConRepArity con == length args_r_to_l )
684 do_pushery orig_d (non_ptr_args ++ ptr_args)
686 -- The args are already in reverse order, which is the way PACK
687 -- expects them to be. We must push the non-ptrs after the ptrs.
688 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
690 do_pushery d (arg:args)
691 = do (push, arg_words) <- pushAtom d p arg
692 more_push_code <- do_pushery (d+arg_words) args
693 return (push `appOL` more_push_code)
695 = return (unitOL (PACK con n_arg_words))
697 n_arg_words = d - orig_d
700 -- -----------------------------------------------------------------------------
701 -- Returning an unboxed tuple with one non-void component (the only
702 -- case we can handle).
704 -- Remember, we don't want to *evaluate* the component that is being
705 -- returned, even if it is a pointed type. We always just return.
708 :: Word16 -> Sequel -> BCEnv
709 -> AnnExpr' Id VarSet -> BcM BCInstrList
710 unboxedTupleReturn d s p arg = do
711 (push, sz) <- pushAtom d p arg
713 mkSLIDE sz (d-s) `snocOL`
714 RETURN_UBX (atomRep arg))
716 -- -----------------------------------------------------------------------------
717 -- Generate code for a tail-call
720 :: Word16 -> Sequel -> BCEnv
721 -> Id -> [AnnExpr' Id VarSet]
723 doTailCall init_d s p fn args
724 = do_pushes init_d args (map atomRep args)
726 do_pushes d [] reps = do
727 ASSERT( null reps ) return ()
728 (push_fn, sz) <- pushAtom d p (AnnVar fn)
729 ASSERT( sz == 1 ) return ()
730 return (push_fn `appOL` (
731 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
733 do_pushes d args reps = do
734 let (push_apply, n, rest_of_reps) = findPushSeq reps
735 (these_args, rest_of_args) = splitAt n args
736 (next_d, push_code) <- push_seq d these_args
737 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
738 -- ^^^ for the PUSH_APPLY_ instruction
739 return (push_code `appOL` (push_apply `consOL` instrs))
741 push_seq d [] = return (d, nilOL)
742 push_seq d (arg:args) = do
743 (push_code, sz) <- pushAtom d p arg
744 (final_d, more_push_code) <- push_seq (d+sz) args
745 return (final_d, push_code `appOL` more_push_code)
747 -- v. similar to CgStackery.findMatch, ToDo: merge
748 findPushSeq :: [CgRep] -> (BCInstr, Int, [CgRep])
749 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
750 = (PUSH_APPLY_PPPPPP, 6, rest)
751 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
752 = (PUSH_APPLY_PPPPP, 5, rest)
753 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
754 = (PUSH_APPLY_PPPP, 4, rest)
755 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
756 = (PUSH_APPLY_PPP, 3, rest)
757 findPushSeq (PtrArg: PtrArg: rest)
758 = (PUSH_APPLY_PP, 2, rest)
759 findPushSeq (PtrArg: rest)
760 = (PUSH_APPLY_P, 1, rest)
761 findPushSeq (VoidArg: rest)
762 = (PUSH_APPLY_V, 1, rest)
763 findPushSeq (NonPtrArg: rest)
764 = (PUSH_APPLY_N, 1, rest)
765 findPushSeq (FloatArg: rest)
766 = (PUSH_APPLY_F, 1, rest)
767 findPushSeq (DoubleArg: rest)
768 = (PUSH_APPLY_D, 1, rest)
769 findPushSeq (LongArg: rest)
770 = (PUSH_APPLY_L, 1, rest)
772 = panic "ByteCodeGen.findPushSeq"
774 -- -----------------------------------------------------------------------------
777 doCase :: Word16 -> Sequel -> BCEnv
778 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
779 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
781 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
783 -- Top of stack is the return itbl, as usual.
784 -- underneath it is the pointer to the alt_code BCO.
785 -- When an alt is entered, it assumes the returned value is
786 -- on top of the itbl.
789 -- An unlifted value gets an extra info table pushed on top
790 -- when it is returned.
791 unlifted_itbl_sizeW | isAlgCase = 0
794 -- depth of stack after the return value has been pushed
795 d_bndr = d + ret_frame_sizeW + fromIntegral (idSizeW bndr)
797 -- depth of stack after the extra info table for an unboxed return
798 -- has been pushed, if any. This is the stack depth at the
800 d_alts = d_bndr + unlifted_itbl_sizeW
802 -- Env in which to compile the alts, not including
803 -- any vars bound by the alts themselves
804 p_alts = Map.insert bndr (d_bndr - 1) p
806 bndr_ty = idType bndr
807 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
809 -- given an alt, return a discr and code for it.
810 codeAlt (DEFAULT, _, (_,rhs))
811 = do rhs_code <- schemeE d_alts s p_alts rhs
812 return (NoDiscr, rhs_code)
814 codeAlt alt@(_, bndrs, (_,rhs))
815 -- primitive or nullary constructor alt: no need to UNPACK
816 | null real_bndrs = do
817 rhs_code <- schemeE d_alts s p_alts rhs
818 return (my_discr alt, rhs_code)
819 -- algebraic alt with some binders
822 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
823 ptr_sizes = map (fromIntegral . idSizeW) ptrs
824 nptrs_sizes = map (fromIntegral . idSizeW) nptrs
825 bind_sizes = ptr_sizes ++ nptrs_sizes
826 size = sum ptr_sizes + sum nptrs_sizes
827 -- the UNPACK instruction unpacks in reverse order...
829 (zip (reverse (ptrs ++ nptrs))
830 (mkStackOffsets d_alts (reverse bind_sizes)))
834 rhs_code <- schemeE (d_alts+size) s p' rhs
835 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
837 real_bndrs = filter (not.isTyCoVar) bndrs
839 my_discr (DEFAULT, _, _) = NoDiscr {-shouldn't really happen-}
840 my_discr (DataAlt dc, _, _)
841 | isUnboxedTupleCon dc
842 = unboxedTupleException
844 = DiscrP (fromIntegral (dataConTag dc - fIRST_TAG))
845 my_discr (LitAlt l, _, _)
846 = case l of MachInt i -> DiscrI (fromInteger i)
847 MachWord w -> DiscrW (fromInteger w)
848 MachFloat r -> DiscrF (fromRational r)
849 MachDouble r -> DiscrD (fromRational r)
850 MachChar i -> DiscrI (ord i)
851 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
854 | not isAlgCase = Nothing
856 = case [dc | (DataAlt dc, _, _) <- alts] of
858 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
860 -- the bitmap is relative to stack depth d, i.e. before the
861 -- BCO, info table and return value are pushed on.
862 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
863 -- except that here we build the bitmap from the known bindings of
864 -- things that are pointers, whereas in CgBindery the code builds the
865 -- bitmap from the free slots and unboxed bindings.
868 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
869 -- The bitmap must cover the portion of the stack up to the sequel only.
870 -- Previously we were building a bitmap for the whole depth (d), but we
871 -- really want a bitmap up to depth (d-s). This affects compilation of
872 -- case-of-case expressions, which is the only time we can be compiling a
873 -- case expression with s /= 0.
876 bitmap_size' = fromIntegral bitmap_size
877 bitmap = intsToReverseBitmap bitmap_size'{-size-}
878 (sortLe (<=) (filter (< bitmap_size') rel_slots))
881 rel_slots = map fromIntegral $ concat (map spread binds)
883 | isFollowableArg (idCgRep id) = [ rel_offset ]
885 where rel_offset = d - offset - 1
888 alt_stuff <- mapM codeAlt alts
889 alt_final <- mkMultiBranch maybe_ncons alt_stuff
892 alt_bco_name = getName bndr
893 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
894 0{-no arity-} bitmap_size bitmap True{-is alts-}
896 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
897 -- "\n bitmap = " ++ show bitmap) $ do
898 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
899 alt_bco' <- emitBc alt_bco
901 | isAlgCase = PUSH_ALTS alt_bco'
902 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
903 return (push_alts `consOL` scrut_code)
906 -- -----------------------------------------------------------------------------
907 -- Deal with a CCall.
909 -- Taggedly push the args onto the stack R->L,
910 -- deferencing ForeignObj#s and adjusting addrs to point to
911 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
912 -- (machine) code for the ccall, and create bytecodes to call that and
913 -- then return in the right way.
915 generateCCall :: Word16 -> Sequel -- stack and sequel depths
917 -> CCallSpec -- where to call
918 -> Id -- of target, for type info
919 -> [AnnExpr' Id VarSet] -- args (atoms)
922 generateCCall d0 s p (CCallSpec target cconv safety) fn args_r_to_l
926 addr_sizeW = fromIntegral (cgRepSizeW NonPtrArg)
928 -- Get the args on the stack, with tags and suitably
929 -- dereferenced for the CCall. For each arg, return the
930 -- depth to the first word of the bits for that arg, and the
931 -- CgRep of what was actually pushed.
933 pargs _ [] = return []
935 = let arg_ty = repType (exprType (deAnnotate' a))
937 in case splitTyConApp_maybe arg_ty of
938 -- Don't push the FO; instead push the Addr# it
941 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
942 -> do rest <- pargs (d + addr_sizeW) az
943 code <- parg_ArrayishRep (fromIntegral arrPtrsHdrSize) d p a
944 return ((code,AddrRep):rest)
946 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
947 -> do rest <- pargs (d + addr_sizeW) az
948 code <- parg_ArrayishRep (fromIntegral arrWordsHdrSize) d p a
949 return ((code,AddrRep):rest)
951 -- Default case: push taggedly, but otherwise intact.
953 -> do (code_a, sz_a) <- pushAtom d p a
954 rest <- pargs (d+sz_a) az
955 return ((code_a, atomPrimRep a) : rest)
957 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
958 -- the stack but then advance it over the headers, so as to
959 -- point to the payload.
960 parg_ArrayishRep :: Word16 -> Word16 -> BCEnv -> AnnExpr' Id VarSet
962 parg_ArrayishRep hdrSize d p a
963 = do (push_fo, _) <- pushAtom d p a
964 -- The ptr points at the header. Advance it over the
965 -- header and then pretend this is an Addr#.
966 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
969 code_n_reps <- pargs d0 args_r_to_l
971 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
972 a_reps_sizeW = fromIntegral (sum (map primRepSizeW a_reps_pushed_r_to_l))
974 push_args = concatOL pushs_arg
975 d_after_args = d0 + a_reps_sizeW
977 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
978 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
980 = reverse (tail a_reps_pushed_r_to_l)
982 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
983 -- push_args is the code to do that.
984 -- d_after_args is the stack depth once the args are on.
986 -- Get the result rep.
987 (returns_void, r_rep)
988 = case maybe_getCCallReturnRep (idType fn) of
989 Nothing -> (True, VoidRep)
990 Just rr -> (False, rr)
992 Because the Haskell stack grows down, the a_reps refer to
993 lowest to highest addresses in that order. The args for the call
994 are on the stack. Now push an unboxed Addr# indicating
995 the C function to call. Then push a dummy placeholder for the
996 result. Finally, emit a CCALL insn with an offset pointing to the
997 Addr# just pushed, and a literal field holding the mallocville
998 address of the piece of marshalling code we generate.
999 So, just prior to the CCALL insn, the stack looks like this
1000 (growing down, as usual):
1005 Addr# address_of_C_fn
1006 <placeholder-for-result#> (must be an unboxed type)
1008 The interpreter then calls the marshall code mentioned
1009 in the CCALL insn, passing it (& <placeholder-for-result#>),
1010 that is, the addr of the topmost word in the stack.
1011 When this returns, the placeholder will have been
1012 filled in. The placeholder is slid down to the sequel
1013 depth, and we RETURN.
1015 This arrangement makes it simple to do f-i-dynamic since the Addr#
1016 value is the first arg anyway.
1018 The marshalling code is generated specifically for this
1019 call site, and so knows exactly the (Haskell) stack
1020 offsets of the args, fn address and placeholder. It
1021 copies the args to the C stack, calls the stacked addr,
1022 and parks the result back in the placeholder. The interpreter
1023 calls it as a normal C call, assuming it has a signature
1024 void marshall_code ( StgWord* ptr_to_top_of_stack )
1026 -- resolve static address
1030 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
1032 StaticTarget target _
1033 -> do res <- ioToBc (lookupStaticPtr stdcall_adj_target)
1037 #ifdef mingw32_TARGET_OS
1038 | StdCallConv <- cconv
1039 = let size = fromIntegral a_reps_sizeW * wORD_SIZE in
1040 mkFastString (unpackFS target ++ '@':show size)
1046 (is_static, static_target_addr) <- get_target_info
1049 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1050 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1051 | is_static = a_reps_pushed_RAW
1052 | otherwise = if null a_reps_pushed_RAW
1053 then panic "ByteCodeGen.generateCCall: dyn with no args"
1054 else tail a_reps_pushed_RAW
1057 (push_Addr, d_after_Addr)
1059 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1060 d_after_args + addr_sizeW)
1061 | otherwise -- is already on the stack
1062 = (nilOL, d_after_args)
1064 -- Push the return placeholder. For a call returning nothing,
1065 -- this is a VoidArg (tag).
1066 r_sizeW = fromIntegral (primRepSizeW r_rep)
1067 d_after_r = d_after_Addr + r_sizeW
1068 r_lit = mkDummyLiteral r_rep
1069 push_r = (if returns_void
1071 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1073 -- generate the marshalling code we're going to call
1075 -- Offset of the next stack frame down the stack. The CCALL
1076 -- instruction needs to describe the chunk of stack containing
1077 -- the ccall args to the GC, so it needs to know how large it
1078 -- is. See comment in Interpreter.c with the CCALL instruction.
1079 stk_offset = d_after_r - s
1082 -- the only difference in libffi mode is that we prepare a cif
1083 -- describing the call type by calling libffi, and we attach the
1084 -- address of this to the CCALL instruction.
1085 token <- ioToBc $ prepForeignCall cconv a_reps r_rep
1086 let addr_of_marshaller = castPtrToFunPtr token
1088 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1091 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller)
1092 (fromIntegral (fromEnum (playInterruptible safety))))
1094 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1095 `snocOL` RETURN_UBX (primRepToCgRep r_rep)
1097 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1100 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1103 -- Make a dummy literal, to be used as a placeholder for FFI return
1104 -- values on the stack.
1105 mkDummyLiteral :: PrimRep -> Literal
1109 WordRep -> MachWord 0
1110 AddrRep -> MachNullAddr
1111 DoubleRep -> MachDouble 0
1112 FloatRep -> MachFloat 0
1113 Int64Rep -> MachInt64 0
1114 Word64Rep -> MachWord64 0
1115 _ -> panic "mkDummyLiteral"
1119 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1120 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1123 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1125 -- Alternatively, for call-targets returning nothing, convert
1127 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1128 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1132 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
1133 maybe_getCCallReturnRep fn_ty
1134 = let (_a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1136 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1138 = case splitTyConApp_maybe (repType r_ty) of
1139 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
1141 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
1142 || r_reps == [VoidRep] )
1143 && isUnboxedTupleTyCon r_tycon
1144 && case maybe_r_rep_to_go of
1146 Just r_rep -> r_rep /= PtrRep
1147 -- if it was, it would be impossible
1148 -- to create a valid return value
1149 -- placeholder on the stack
1151 blargh :: a -- Used at more than one type
1152 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1155 --trace (showSDoc (ppr (a_reps, r_reps))) $
1156 if ok then maybe_r_rep_to_go else blargh
1158 -- Compile code which expects an unboxed Int on the top of stack,
1159 -- (call it i), and pushes the i'th closure in the supplied list
1160 -- as a consequence.
1161 implement_tagToId :: [Name] -> BcM BCInstrList
1162 implement_tagToId names
1163 = ASSERT( notNull names )
1164 do labels <- getLabelsBc (genericLength names)
1165 label_fail <- getLabelBc
1166 label_exit <- getLabelBc
1167 let infos = zip4 labels (tail labels ++ [label_fail])
1169 steps = map (mkStep label_exit) infos
1170 return (concatOL steps
1172 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1174 mkStep l_exit (my_label, next_label, n, name_for_n)
1175 = toOL [LABEL my_label,
1176 TESTEQ_I n next_label,
1181 -- -----------------------------------------------------------------------------
1184 -- Push an atom onto the stack, returning suitable code & number of
1185 -- stack words used.
1187 -- The env p must map each variable to the highest- numbered stack
1188 -- slot for it. For example, if the stack has depth 4 and we
1189 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1190 -- the tag in stack[5], the stack will have depth 6, and p must map v
1191 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1192 -- depth 6 stack has valid words 0 .. 5.
1194 pushAtom :: Word16 -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Word16)
1197 | Just e' <- bcView e
1200 pushAtom d p (AnnVar v)
1201 | idCgRep v == VoidArg
1205 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1207 | Just primop <- isPrimOpId_maybe v
1208 = return (unitOL (PUSH_PRIMOP primop), 1)
1210 | Just d_v <- lookupBCEnv_maybe v p -- v is a local variable
1211 = let l = d - d_v + sz - 2
1212 in return (toOL (genericReplicate sz (PUSH_L l)), sz)
1213 -- d - d_v the number of words between the TOS
1214 -- and the 1st slot of the object
1216 -- d - d_v - 1 the offset from the TOS of the 1st slot
1218 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1221 -- Having found the last slot, we proceed to copy the right number of
1222 -- slots on to the top of the stack.
1224 | otherwise -- v must be a global variable
1226 return (unitOL (PUSH_G (getName v)), sz)
1230 sz = fromIntegral (idSizeW v)
1233 pushAtom _ _ (AnnLit lit)
1235 MachLabel _ _ _ -> code NonPtrArg
1236 MachWord _ -> code NonPtrArg
1237 MachInt _ -> code PtrArg
1238 MachFloat _ -> code FloatArg
1239 MachDouble _ -> code DoubleArg
1240 MachChar _ -> code NonPtrArg
1241 MachNullAddr -> code NonPtrArg
1242 MachStr s -> pushStr s
1243 l -> pprPanic "pushAtom" (ppr l)
1246 = let size_host_words = fromIntegral (cgRepSizeW rep)
1247 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1251 = let getMallocvilleAddr
1253 FastString _ n _ fp _ ->
1254 -- we could grab the Ptr from the ForeignPtr,
1255 -- but then we have no way to control its lifetime.
1256 -- In reality it'll probably stay alive long enoungh
1257 -- by virtue of the global FastString table, but
1258 -- to be on the safe side we copy the string into
1259 -- a malloc'd area of memory.
1260 do ptr <- ioToBc (mallocBytes (n+1))
1263 withForeignPtr fp $ \p -> do
1264 memcpy ptr p (fromIntegral n)
1265 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1269 addr <- getMallocvilleAddr
1270 -- Get the addr on the stack, untaggedly
1271 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1273 pushAtom d p (AnnCast e _)
1274 = pushAtom d p (snd e)
1277 = pprPanic "ByteCodeGen.pushAtom"
1278 (pprCoreExpr (deAnnotate (undefined, expr)))
1280 foreign import ccall unsafe "memcpy"
1281 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1284 -- -----------------------------------------------------------------------------
1285 -- Given a bunch of alts code and their discrs, do the donkey work
1286 -- of making a multiway branch using a switch tree.
1287 -- What a load of hassle!
1289 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1290 -- a hint; generates better code
1291 -- Nothing is always safe
1292 -> [(Discr, BCInstrList)]
1294 mkMultiBranch maybe_ncons raw_ways
1295 = let d_way = filter (isNoDiscr.fst) raw_ways
1297 (\w1 w2 -> leAlt (fst w1) (fst w2))
1298 (filter (not.isNoDiscr.fst) raw_ways)
1300 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1301 mkTree [] _range_lo _range_hi = return the_default
1303 mkTree [val] range_lo range_hi
1304 | range_lo `eqAlt` range_hi
1307 = do label_neq <- getLabelBc
1308 return (testEQ (fst val) label_neq
1310 `appOL` unitOL (LABEL label_neq)
1311 `appOL` the_default))
1313 mkTree vals range_lo range_hi
1314 = let n = length vals `div` 2
1315 vals_lo = take n vals
1316 vals_hi = drop n vals
1317 v_mid = fst (head vals_hi)
1319 label_geq <- getLabelBc
1320 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1321 code_hi <- mkTree vals_hi v_mid range_hi
1322 return (testLT v_mid label_geq
1324 `appOL` unitOL (LABEL label_geq)
1328 = case d_way of [] -> unitOL CASEFAIL
1330 _ -> panic "mkMultiBranch/the_default"
1332 testLT (DiscrI i) fail_label = TESTLT_I i fail_label
1333 testLT (DiscrW i) fail_label = TESTLT_W i fail_label
1334 testLT (DiscrF i) fail_label = TESTLT_F i fail_label
1335 testLT (DiscrD i) fail_label = TESTLT_D i fail_label
1336 testLT (DiscrP i) fail_label = TESTLT_P i fail_label
1337 testLT NoDiscr _ = panic "mkMultiBranch NoDiscr"
1339 testEQ (DiscrI i) fail_label = TESTEQ_I i fail_label
1340 testEQ (DiscrW i) fail_label = TESTEQ_W i fail_label
1341 testEQ (DiscrF i) fail_label = TESTEQ_F i fail_label
1342 testEQ (DiscrD i) fail_label = TESTEQ_D i fail_label
1343 testEQ (DiscrP i) fail_label = TESTEQ_P i fail_label
1344 testEQ NoDiscr _ = panic "mkMultiBranch NoDiscr"
1346 -- None of these will be needed if there are no non-default alts
1349 = panic "mkMultiBranch: awesome foursome"
1351 = case fst (head notd_ways) of
1352 DiscrI _ -> ( DiscrI minBound, DiscrI maxBound )
1353 DiscrW _ -> ( DiscrW minBound, DiscrW maxBound )
1354 DiscrF _ -> ( DiscrF minF, DiscrF maxF )
1355 DiscrD _ -> ( DiscrD minD, DiscrD maxD )
1356 DiscrP _ -> ( DiscrP algMinBound, DiscrP algMaxBound )
1357 NoDiscr -> panic "mkMultiBranch NoDiscr"
1359 (algMinBound, algMaxBound)
1360 = case maybe_ncons of
1361 -- XXX What happens when n == 0?
1362 Just n -> (0, fromIntegral n - 1)
1363 Nothing -> (minBound, maxBound)
1365 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1366 (DiscrW w1) `eqAlt` (DiscrW w2) = w1 == w2
1367 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1368 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1369 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1370 NoDiscr `eqAlt` NoDiscr = True
1373 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1374 (DiscrW w1) `leAlt` (DiscrW w2) = w1 <= w2
1375 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1376 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1377 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1378 NoDiscr `leAlt` NoDiscr = True
1381 isNoDiscr NoDiscr = True
1384 dec (DiscrI i) = DiscrI (i-1)
1385 dec (DiscrW w) = DiscrW (w-1)
1386 dec (DiscrP i) = DiscrP (i-1)
1387 dec other = other -- not really right, but if you
1388 -- do cases on floating values, you'll get what you deserve
1390 -- same snotty comment applies to the following
1392 minD, maxD :: Double
1398 mkTree notd_ways init_lo init_hi
1401 -- -----------------------------------------------------------------------------
1402 -- Supporting junk for the compilation schemes
1404 -- Describes case alts
1413 instance Outputable Discr where
1414 ppr (DiscrI i) = int i
1415 ppr (DiscrW w) = text (show w)
1416 ppr (DiscrF f) = text (show f)
1417 ppr (DiscrD d) = text (show d)
1418 ppr (DiscrP i) = ppr i
1419 ppr NoDiscr = text "DEF"
1422 lookupBCEnv_maybe :: Id -> BCEnv -> Maybe Word16
1423 lookupBCEnv_maybe = Map.lookup
1425 idSizeW :: Id -> Int
1426 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1429 unboxedTupleException :: a
1430 unboxedTupleException
1433 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1434 " Possibly due to foreign import/export decls in source.\n"++
1435 " Workaround: use -fobject-code, or compile this module to .o separately."))
1438 mkSLIDE :: Word16 -> Word16 -> OrdList BCInstr
1439 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1441 splitApp :: AnnExpr' Var ann -> (AnnExpr' Var ann, [AnnExpr' Var ann])
1442 -- The arguments are returned in *right-to-left* order
1443 splitApp e | Just e' <- bcView e = splitApp e'
1444 splitApp (AnnApp (_,f) (_,a)) = case splitApp f of
1445 (f', as) -> (f', a:as)
1446 splitApp e = (e, [])
1449 bcView :: AnnExpr' Var ann -> Maybe (AnnExpr' Var ann)
1450 -- The "bytecode view" of a term discards
1451 -- a) type abstractions
1452 -- b) type applications
1455 -- Type lambdas *can* occur in random expressions,
1456 -- whereas value lambdas cannot; that is why they are nuked here
1457 bcView (AnnNote _ (_,e)) = Just e
1458 bcView (AnnCast (_,e) _) = Just e
1459 bcView (AnnLam v (_,e)) | isTyCoVar v = Just e
1460 bcView (AnnApp (_,e) (_, AnnType _)) = Just e
1463 isVoidArgAtom :: AnnExpr' Var ann -> Bool
1464 isVoidArgAtom e | Just e' <- bcView e = isVoidArgAtom e'
1465 isVoidArgAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1466 isVoidArgAtom _ = False
1468 atomPrimRep :: AnnExpr' Id ann -> PrimRep
1469 atomPrimRep e | Just e' <- bcView e = atomPrimRep e'
1470 atomPrimRep (AnnVar v) = typePrimRep (idType v)
1471 atomPrimRep (AnnLit l) = typePrimRep (literalType l)
1472 atomPrimRep other = pprPanic "atomPrimRep" (ppr (deAnnotate (undefined,other)))
1474 atomRep :: AnnExpr' Id ann -> CgRep
1475 atomRep e = primRepToCgRep (atomPrimRep e)
1477 isPtrAtom :: AnnExpr' Id ann -> Bool
1478 isPtrAtom e = atomRep e == PtrArg
1480 -- Let szsw be the sizes in words of some items pushed onto the stack,
1481 -- which has initial depth d'. Return the values which the stack environment
1482 -- should map these items to.
1483 mkStackOffsets :: Word16 -> [Word16] -> [Word16]
1484 mkStackOffsets original_depth szsw
1485 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1487 -- -----------------------------------------------------------------------------
1488 -- The bytecode generator's monad
1490 type BcPtr = Either ItblPtr (Ptr ())
1494 uniqSupply :: UniqSupply, -- for generating fresh variable names
1495 nextlabel :: Word16, -- for generating local labels
1496 malloced :: [BcPtr], -- thunks malloced for current BCO
1497 -- Should be free()d when it is GCd
1498 breakArray :: BreakArray -- array of breakpoint flags
1501 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1503 ioToBc :: IO a -> BcM a
1504 ioToBc io = BcM $ \st -> do
1508 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1509 runBc us modBreaks (BcM m)
1510 = m (BcM_State us 0 [] breakArray)
1512 breakArray = modBreaks_flags modBreaks
1514 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1515 thenBc (BcM expr) cont = BcM $ \st0 -> do
1516 (st1, q) <- expr st0
1521 thenBc_ :: BcM a -> BcM b -> BcM b
1522 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1523 (st1, _) <- expr st0
1524 (st2, r) <- cont st1
1527 returnBc :: a -> BcM a
1528 returnBc result = BcM $ \st -> (return (st, result))
1530 instance Monad BcM where
1535 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1537 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1539 recordMallocBc :: Ptr a -> BcM ()
1541 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1543 recordItblMallocBc :: ItblPtr -> BcM ()
1544 recordItblMallocBc a
1545 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1547 getLabelBc :: BcM Word16
1549 = BcM $ \st -> do let nl = nextlabel st
1550 when (nl == maxBound) $
1551 panic "getLabelBc: Ran out of labels"
1552 return (st{nextlabel = nl + 1}, nl)
1554 getLabelsBc :: Word16 -> BcM [Word16]
1556 = BcM $ \st -> let ctr = nextlabel st
1557 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1559 getBreakArray :: BcM BreakArray
1560 getBreakArray = BcM $ \st -> return (st, breakArray st)
1562 newUnique :: BcM Unique
1564 \st -> case takeUniqFromSupply (uniqSupply st) of
1565 (uniq, us) -> let newState = st { uniqSupply = us }
1566 in return (newState, uniq)
1568 newId :: Type -> BcM Id
1571 return $ mkSysLocal tickFS uniq ty
1573 tickFS :: FastString
1574 tickFS = fsLit "ticked"