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"
54 -- import GHC.Exts ( Int(..) )
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 = Word16 -- 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 Word16 -- 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 one 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 < fromIntegral 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 >= fromIntegral 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 :: AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
263 collect (_, e) = go [] e
265 go xs e | Just e' <- bcView e = go xs e'
266 go xs (AnnLam x (_,e)) = go (x:xs) e
267 go xs not_lambda = (reverse xs, not_lambda)
269 schemeR_wrk :: [Id] -> Id -> AnnExpr Id VarSet -> ([Var], AnnExpr' Var VarSet) -> BcM (ProtoBCO Name)
270 schemeR_wrk fvs nm original_body (args, body)
272 all_args = reverse args ++ fvs
273 arity = length all_args
274 -- all_args are the args in reverse order. We're compiling a function
275 -- \fv1..fvn x1..xn -> e
276 -- i.e. the fvs come first
278 szsw_args = map (fromIntegral . idSizeW) all_args
279 szw_args = sum szsw_args
280 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
282 -- make the arg bitmap
283 bits = argBits (reverse (map idCgRep all_args))
284 bitmap_size = genericLength bits
285 bitmap = mkBitmap bits
287 body_code <- schemeER_wrk szw_args p_init body
289 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
290 arity bitmap_size bitmap False{-not alts-})
292 -- introduce break instructions for ticked expressions
293 schemeER_wrk :: Word16 -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
295 | Just (tickInfo, (_annot, newRhs)) <- isTickedExp' rhs = do
296 code <- schemeE d 0 p newRhs
298 let idOffSets = getVarOffSets (fromIntegral d) p tickInfo
299 let tickNumber = tickInfo_number tickInfo
300 let breakInfo = BreakInfo
301 { breakInfo_module = tickInfo_module tickInfo
302 , breakInfo_number = tickNumber
303 , breakInfo_vars = idOffSets
304 , breakInfo_resty = exprType (deAnnotate' newRhs)
306 let breakInstr = case arr of
308 BRK_FUN arr# (fromIntegral tickNumber) breakInfo
309 return $ breakInstr `consOL` code
310 | otherwise = schemeE d 0 p rhs
312 getVarOffSets :: Word16 -> BCEnv -> TickInfo -> [(Id, Word16)]
313 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
315 getOffSet :: Word16 -> BCEnv -> Id -> Maybe (Id, Word16)
317 = case lookupBCEnv_maybe env id of
319 Just offset -> Just (id, d - offset)
321 fvsToEnv :: BCEnv -> VarSet -> [Id]
322 -- Takes the free variables of a right-hand side, and
323 -- delivers an ordered list of the local variables that will
324 -- be captured in the thunk for the RHS
325 -- The BCEnv argument tells which variables are in the local
326 -- environment: these are the ones that should be captured
328 -- The code that constructs the thunk, and the code that executes
329 -- it, have to agree about this layout
330 fvsToEnv p fvs = [v | v <- varSetElems fvs,
331 isId v, -- Could be a type variable
334 -- -----------------------------------------------------------------------------
339 { tickInfo_number :: Int -- the (module) unique number of the tick
340 , tickInfo_module :: Module -- the origin of the ticked expression
341 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
344 instance Outputable TickInfo where
345 ppr info = text "TickInfo" <+>
346 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
347 ppr (tickInfo_locals info))
349 -- Compile code to apply the given expression to the remaining args
350 -- on the stack, returning a HNF.
351 schemeE :: Word16 -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
354 | Just e' <- bcView e
357 -- Delegate tail-calls to schemeT.
358 schemeE d s p e@(AnnApp _ _)
361 schemeE d s p e@(AnnVar v)
362 | not (isUnLiftedType v_type)
363 = -- Lifted-type thing; push it in the normal way
367 = do -- Returning an unlifted value.
368 -- Heave it on the stack, SLIDE, and RETURN.
369 (push, szw) <- pushAtom d p (AnnVar v)
370 return (push -- value onto stack
371 `appOL` mkSLIDE szw (d-s) -- clear to sequel
372 `snocOL` RETURN_UBX v_rep) -- go
375 v_rep = typeCgRep v_type
377 schemeE d s p (AnnLit literal)
378 = do (push, szw) <- pushAtom d p (AnnLit literal)
379 let l_rep = typeCgRep (literalType literal)
380 return (push -- value onto stack
381 `appOL` mkSLIDE szw (d-s) -- clear to sequel
382 `snocOL` RETURN_UBX l_rep) -- go
384 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
385 | (AnnVar v, args_r_to_l) <- splitApp rhs,
386 Just data_con <- isDataConWorkId_maybe v,
387 dataConRepArity data_con == length args_r_to_l
388 = do -- Special case for a non-recursive let whose RHS is a
389 -- saturatred constructor application.
390 -- Just allocate the constructor and carry on
391 alloc_code <- mkConAppCode d s p data_con args_r_to_l
392 body_code <- schemeE (d+1) s (addToFM p x d) body
393 return (alloc_code `appOL` body_code)
395 -- General case for let. Generates correct, if inefficient, code in
397 schemeE d s p (AnnLet binds (_,body))
398 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
399 AnnRec xs_n_rhss -> unzip xs_n_rhss
400 n_binds = genericLength xs
402 fvss = map (fvsToEnv p' . fst) rhss
404 -- Sizes of free vars
405 sizes = map (\rhs_fvs -> sum (map (fromIntegral . idSizeW) rhs_fvs)) fvss
407 -- the arity of each rhs
408 arities = map (genericLength . fst . collect) rhss
410 -- This p', d' defn is safe because all the items being pushed
411 -- are ptrs, so all have size 1. d' and p' reflect the stack
412 -- after the closures have been allocated in the heap (but not
413 -- filled in), and pointers to them parked on the stack.
414 p' = addListToFM p (zipE xs (mkStackOffsets d (genericReplicate n_binds 1)))
416 zipE = zipEqual "schemeE"
418 -- ToDo: don't build thunks for things with no free variables
419 build_thunk _ [] size bco off arity
420 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
422 mkap | arity == 0 = MKAP
424 build_thunk dd (fv:fvs) size bco off arity = do
425 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
426 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
427 return (push_code `appOL` more_push_code)
429 alloc_code = toOL (zipWith mkAlloc sizes arities)
431 | is_tick = ALLOC_AP_NOUPD sz
432 | otherwise = ALLOC_AP sz
433 mkAlloc sz arity = ALLOC_PAP arity sz
435 is_tick = case binds of
436 AnnNonRec id _ -> occNameFS (getOccName id) == tickFS
439 compile_bind d' fvs x rhs size arity off = do
440 bco <- schemeR fvs (x,rhs)
441 build_thunk d' fvs size bco off arity
444 [ compile_bind d' fvs x rhs size arity n
445 | (fvs, x, rhs, size, arity, n) <-
446 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
449 body_code <- schemeE d' s p' body
450 thunk_codes <- sequence compile_binds
451 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
453 -- introduce a let binding for a ticked case expression. This rule
454 -- *should* only fire when the expression was not already let-bound
455 -- (the code gen for let bindings should take care of that). Todo: we
456 -- call exprFreeVars on a deAnnotated expression, this may not be the
457 -- best way to calculate the free vars but it seemed like the least
458 -- intrusive thing to do
459 schemeE d s p exp@(AnnCase {})
460 | Just (_tickInfo, _rhs) <- isTickedExp' exp
461 = if isUnLiftedType ty
463 -- If the result type is unlifted, then we must generate
464 -- let f = \s . case tick# of _ -> e
466 -- When we stop at the breakpoint, _result will have an unlifted
467 -- type and hence won't be bound in the environment, but the
468 -- breakpoint will otherwise work fine.
469 id <- newId (mkFunTy realWorldStatePrimTy ty)
470 st <- newId realWorldStatePrimTy
471 let letExp = AnnLet (AnnNonRec id (fvs, AnnLam st (emptyVarSet, exp)))
472 (emptyVarSet, (AnnApp (emptyVarSet, AnnVar id)
473 (emptyVarSet, AnnVar realWorldPrimId)))
477 -- Todo: is emptyVarSet correct on the next line?
478 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
480 where exp' = deAnnotate' exp
481 fvs = exprFreeVars exp'
484 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1, bind2], rhs)])
485 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
487 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
489 -- case .... of a { DEFAULT -> ... }
490 -- becuse the return convention for both are identical.
492 -- Note that it does not matter losing the void-rep thing from the
493 -- envt (it won't be bound now) because we never look such things up.
495 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
496 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
498 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
499 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
500 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
502 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1], rhs)])
503 | isUnboxedTupleCon dc
504 -- Similarly, convert
505 -- case .... of x { (# a #) -> ... }
507 -- case .... of a { DEFAULT -> ... }
508 = --trace "automagic mashing of case alts (# a #)" $
509 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
511 schemeE d s p (AnnCase scrut bndr _ alts)
512 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
515 = pprPanic "ByteCodeGen.schemeE: unhandled case"
516 (pprCoreExpr (deAnnotate' expr))
522 A ticked expression looks like this:
524 case tick<n> var1 ... varN of DEFAULT -> e
526 (*) <n> is the number of the tick, which is unique within a module
527 (*) var1 ... varN are the local variables in scope at the tick site
529 If we find a ticked expression we return:
531 Just ((n, [var1 ... varN]), e)
533 otherwise we return Nothing.
535 The idea is that the "case tick<n> ..." is really just an annotation on
536 the code. When we find such a thing, we pull out the useful information,
537 and then compile the code as if it was just the expression "e".
541 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
542 isTickedExp' (AnnCase scrut _bndr _type alts)
543 | Just tickInfo <- isTickedScrut scrut,
544 [(DEFAULT, _bndr, rhs)] <- alts
545 = Just (tickInfo, rhs)
547 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
550 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
551 = Just $ TickInfo { tickInfo_number = tickNumber
552 , tickInfo_module = modName
553 , tickInfo_locals = idsOfArgs args
555 | otherwise = Nothing
557 (f, args) = collectArgs $ deAnnotate expr
558 idsOfArgs :: [Expr Id] -> [Id]
559 idsOfArgs = catMaybes . map exprId
560 exprId :: Expr Id -> Maybe Id
561 exprId (Var id) = Just id
564 isTickedExp' _ = Nothing
566 -- Compile code to do a tail call. Specifically, push the fn,
567 -- slide the on-stack app back down to the sequel depth,
568 -- and enter. Four cases:
571 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
572 -- The int will be on the stack. Generate a code sequence
573 -- to convert it to the relevant constructor, SLIDE and ENTER.
575 -- 1. The fn denotes a ccall. Defer to generateCCall.
577 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
578 -- it simply as b -- since the representations are identical
579 -- (the VoidArg takes up zero stack space). Also, spot
580 -- (# b #) and treat it as b.
582 -- 3. Application of a constructor, by defn saturated.
583 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
584 -- then the ptrs, and then do PACK and RETURN.
586 -- 4. Otherwise, it must be a function call. Push the args
587 -- right to left, SLIDE and ENTER.
589 schemeT :: Word16 -- Stack depth
590 -> Sequel -- Sequel depth
591 -> BCEnv -- stack env
592 -> AnnExpr' Id VarSet
597 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
598 -- = panic "schemeT ?!?!"
600 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
604 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
605 = do (push, arg_words) <- pushAtom d p arg
606 tagToId_sequence <- implement_tagToId constr_names
607 return (push `appOL` tagToId_sequence
608 `appOL` mkSLIDE 1 (d+arg_words-s)
612 | Just (CCall ccall_spec) <- isFCallId_maybe fn
613 = generateCCall d s p ccall_spec fn args_r_to_l
615 -- Case 2: Constructor application
616 | Just con <- maybe_saturated_dcon,
617 isUnboxedTupleCon con
618 = case args_r_to_l of
619 [arg1,arg2] | isVoidArgAtom arg1 ->
620 unboxedTupleReturn d s p arg2
621 [arg1,arg2] | isVoidArgAtom arg2 ->
622 unboxedTupleReturn d s p arg1
623 _other -> unboxedTupleException
625 -- Case 3: Ordinary data constructor
626 | Just con <- maybe_saturated_dcon
627 = do alloc_con <- mkConAppCode d s p con args_r_to_l
628 return (alloc_con `appOL`
629 mkSLIDE 1 (d - s) `snocOL`
632 -- Case 4: Tail call of function
634 = doTailCall d s p fn args_r_to_l
637 -- Detect and extract relevant info for the tagToEnum kludge.
638 maybe_is_tagToEnum_call
639 = let extract_constr_Names ty
640 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
642 = map (getName . dataConWorkId) (tyConDataCons tyc)
643 -- NOTE: use the worker name, not the source name of
644 -- the DataCon. See DataCon.lhs for details.
646 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
649 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
650 -> case isPrimOpId_maybe v of
651 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
655 -- Extract the args (R->L) and fn
656 -- The function will necessarily be a variable,
657 -- because we are compiling a tail call
658 (AnnVar fn, args_r_to_l) = splitApp app
660 -- Only consider this to be a constructor application iff it is
661 -- saturated. Otherwise, we'll call the constructor wrapper.
662 n_args = length args_r_to_l
664 = case isDataConWorkId_maybe fn of
665 Just con | dataConRepArity con == n_args -> Just con
668 -- -----------------------------------------------------------------------------
669 -- Generate code to build a constructor application,
670 -- leaving it on top of the stack
672 mkConAppCode :: Word16 -> Sequel -> BCEnv
673 -> DataCon -- The data constructor
674 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
677 mkConAppCode _ _ _ con [] -- Nullary constructor
678 = ASSERT( isNullaryRepDataCon con )
679 return (unitOL (PUSH_G (getName (dataConWorkId con))))
680 -- Instead of doing a PACK, which would allocate a fresh
681 -- copy of this constructor, use the single shared version.
683 mkConAppCode orig_d _ p con args_r_to_l
684 = ASSERT( dataConRepArity con == length args_r_to_l )
685 do_pushery orig_d (non_ptr_args ++ ptr_args)
687 -- The args are already in reverse order, which is the way PACK
688 -- expects them to be. We must push the non-ptrs after the ptrs.
689 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
691 do_pushery d (arg:args)
692 = do (push, arg_words) <- pushAtom d p arg
693 more_push_code <- do_pushery (d+arg_words) args
694 return (push `appOL` more_push_code)
696 = return (unitOL (PACK con n_arg_words))
698 n_arg_words = d - orig_d
701 -- -----------------------------------------------------------------------------
702 -- Returning an unboxed tuple with one non-void component (the only
703 -- case we can handle).
705 -- Remember, we don't want to *evaluate* the component that is being
706 -- returned, even if it is a pointed type. We always just return.
709 :: Word16 -> Sequel -> BCEnv
710 -> AnnExpr' Id VarSet -> BcM BCInstrList
711 unboxedTupleReturn d s p arg = do
712 (push, sz) <- pushAtom d p arg
714 mkSLIDE sz (d-s) `snocOL`
715 RETURN_UBX (atomRep arg))
717 -- -----------------------------------------------------------------------------
718 -- Generate code for a tail-call
721 :: Word16 -> Sequel -> BCEnv
722 -> Id -> [AnnExpr' Id VarSet]
724 doTailCall init_d s p fn args
725 = do_pushes init_d args (map atomRep args)
727 do_pushes d [] reps = do
728 ASSERT( null reps ) return ()
729 (push_fn, sz) <- pushAtom d p (AnnVar fn)
730 ASSERT( sz == 1 ) return ()
731 return (push_fn `appOL` (
732 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
734 do_pushes d args reps = do
735 let (push_apply, n, rest_of_reps) = findPushSeq reps
736 (these_args, rest_of_args) = splitAt n args
737 (next_d, push_code) <- push_seq d these_args
738 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
739 -- ^^^ for the PUSH_APPLY_ instruction
740 return (push_code `appOL` (push_apply `consOL` instrs))
742 push_seq d [] = return (d, nilOL)
743 push_seq d (arg:args) = do
744 (push_code, sz) <- pushAtom d p arg
745 (final_d, more_push_code) <- push_seq (d+sz) args
746 return (final_d, push_code `appOL` more_push_code)
748 -- v. similar to CgStackery.findMatch, ToDo: merge
749 findPushSeq :: [CgRep] -> (BCInstr, Int, [CgRep])
750 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
751 = (PUSH_APPLY_PPPPPP, 6, rest)
752 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
753 = (PUSH_APPLY_PPPPP, 5, rest)
754 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
755 = (PUSH_APPLY_PPPP, 4, rest)
756 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
757 = (PUSH_APPLY_PPP, 3, rest)
758 findPushSeq (PtrArg: PtrArg: rest)
759 = (PUSH_APPLY_PP, 2, rest)
760 findPushSeq (PtrArg: rest)
761 = (PUSH_APPLY_P, 1, rest)
762 findPushSeq (VoidArg: rest)
763 = (PUSH_APPLY_V, 1, rest)
764 findPushSeq (NonPtrArg: rest)
765 = (PUSH_APPLY_N, 1, rest)
766 findPushSeq (FloatArg: rest)
767 = (PUSH_APPLY_F, 1, rest)
768 findPushSeq (DoubleArg: rest)
769 = (PUSH_APPLY_D, 1, rest)
770 findPushSeq (LongArg: rest)
771 = (PUSH_APPLY_L, 1, rest)
773 = panic "ByteCodeGen.findPushSeq"
775 -- -----------------------------------------------------------------------------
778 doCase :: Word16 -> Sequel -> BCEnv
779 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
780 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
782 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
784 -- Top of stack is the return itbl, as usual.
785 -- underneath it is the pointer to the alt_code BCO.
786 -- When an alt is entered, it assumes the returned value is
787 -- on top of the itbl.
790 -- An unlifted value gets an extra info table pushed on top
791 -- when it is returned.
792 unlifted_itbl_sizeW | isAlgCase = 0
795 -- depth of stack after the return value has been pushed
796 d_bndr = d + ret_frame_sizeW + fromIntegral (idSizeW bndr)
798 -- depth of stack after the extra info table for an unboxed return
799 -- has been pushed, if any. This is the stack depth at the
801 d_alts = d_bndr + unlifted_itbl_sizeW
803 -- Env in which to compile the alts, not including
804 -- any vars bound by the alts themselves
805 p_alts = addToFM p bndr (d_bndr - 1)
807 bndr_ty = idType bndr
808 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
810 -- given an alt, return a discr and code for it.
811 codeAlt (DEFAULT, _, (_,rhs))
812 = do rhs_code <- schemeE d_alts s p_alts rhs
813 return (NoDiscr, rhs_code)
815 codeAlt alt@(_, bndrs, (_,rhs))
816 -- primitive or nullary constructor alt: no need to UNPACK
817 | null real_bndrs = do
818 rhs_code <- schemeE d_alts s p_alts rhs
819 return (my_discr alt, rhs_code)
820 -- algebraic alt with some binders
823 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
824 ptr_sizes = map (fromIntegral . idSizeW) ptrs
825 nptrs_sizes = map (fromIntegral . idSizeW) nptrs
826 bind_sizes = ptr_sizes ++ nptrs_sizes
827 size = sum ptr_sizes + sum nptrs_sizes
828 -- the UNPACK instruction unpacks in reverse order...
829 p' = addListToFM p_alts
830 (zip (reverse (ptrs ++ nptrs))
831 (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.isTyVar) 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 MachFloat r -> DiscrF (fromRational r)
848 MachDouble r -> DiscrD (fromRational r)
849 MachChar i -> DiscrI (ord i)
850 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
853 | not isAlgCase = Nothing
855 = case [dc | (DataAlt dc, _, _) <- alts] of
857 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
859 -- the bitmap is relative to stack depth d, i.e. before the
860 -- BCO, info table and return value are pushed on.
861 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
862 -- except that here we build the bitmap from the known bindings of
863 -- things that are pointers, whereas in CgBindery the code builds the
864 -- bitmap from the free slots and unboxed bindings.
867 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
868 -- The bitmap must cover the portion of the stack up to the sequel only.
869 -- Previously we were building a bitmap for the whole depth (d), but we
870 -- really want a bitmap up to depth (d-s). This affects compilation of
871 -- case-of-case expressions, which is the only time we can be compiling a
872 -- case expression with s /= 0.
875 bitmap_size' = fromIntegral bitmap_size
876 bitmap = intsToReverseBitmap bitmap_size'{-size-}
877 (sortLe (<=) (filter (< bitmap_size') rel_slots))
880 rel_slots = map fromIntegral $ concat (map spread binds)
882 | isFollowableArg (idCgRep id) = [ rel_offset ]
884 where rel_offset = d - offset - 1
887 alt_stuff <- mapM codeAlt alts
888 alt_final <- mkMultiBranch maybe_ncons alt_stuff
891 alt_bco_name = getName bndr
892 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
893 0{-no arity-} bitmap_size bitmap True{-is alts-}
895 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
896 -- "\n bitmap = " ++ show bitmap) $ do
897 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
898 alt_bco' <- emitBc alt_bco
900 | isAlgCase = PUSH_ALTS alt_bco'
901 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
902 return (push_alts `consOL` scrut_code)
905 -- -----------------------------------------------------------------------------
906 -- Deal with a CCall.
908 -- Taggedly push the args onto the stack R->L,
909 -- deferencing ForeignObj#s and adjusting addrs to point to
910 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
911 -- (machine) code for the ccall, and create bytecodes to call that and
912 -- then return in the right way.
914 generateCCall :: Word16 -> Sequel -- stack and sequel depths
916 -> CCallSpec -- where to call
917 -> Id -- of target, for type info
918 -> [AnnExpr' Id VarSet] -- args (atoms)
921 generateCCall d0 s p (CCallSpec target cconv _) fn args_r_to_l
925 addr_sizeW = fromIntegral (cgRepSizeW NonPtrArg)
927 -- Get the args on the stack, with tags and suitably
928 -- dereferenced for the CCall. For each arg, return the
929 -- depth to the first word of the bits for that arg, and the
930 -- CgRep of what was actually pushed.
932 pargs _ [] = return []
934 = let arg_ty = repType (exprType (deAnnotate' a))
936 in case splitTyConApp_maybe arg_ty of
937 -- Don't push the FO; instead push the Addr# it
940 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
941 -> do rest <- pargs (d + addr_sizeW) az
942 code <- parg_ArrayishRep (fromIntegral arrPtrsHdrSize) d p a
943 return ((code,AddrRep):rest)
945 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
946 -> do rest <- pargs (d + addr_sizeW) az
947 code <- parg_ArrayishRep (fromIntegral arrWordsHdrSize) d p a
948 return ((code,AddrRep):rest)
950 -- Default case: push taggedly, but otherwise intact.
952 -> do (code_a, sz_a) <- pushAtom d p a
953 rest <- pargs (d+sz_a) az
954 return ((code_a, atomPrimRep a) : rest)
956 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
957 -- the stack but then advance it over the headers, so as to
958 -- point to the payload.
959 parg_ArrayishRep :: Word16 -> Word16 -> BCEnv -> AnnExpr' Id VarSet
961 parg_ArrayishRep hdrSize d p a
962 = do (push_fo, _) <- pushAtom d p a
963 -- The ptr points at the header. Advance it over the
964 -- header and then pretend this is an Addr#.
965 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
968 code_n_reps <- pargs d0 args_r_to_l
970 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
971 a_reps_sizeW = fromIntegral (sum (map primRepSizeW a_reps_pushed_r_to_l))
973 push_args = concatOL pushs_arg
974 d_after_args = d0 + a_reps_sizeW
976 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
977 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
979 = reverse (tail a_reps_pushed_r_to_l)
981 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
982 -- push_args is the code to do that.
983 -- d_after_args is the stack depth once the args are on.
985 -- Get the result rep.
986 (returns_void, r_rep)
987 = case maybe_getCCallReturnRep (idType fn) of
988 Nothing -> (True, VoidRep)
989 Just rr -> (False, rr)
991 Because the Haskell stack grows down, the a_reps refer to
992 lowest to highest addresses in that order. The args for the call
993 are on the stack. Now push an unboxed Addr# indicating
994 the C function to call. Then push a dummy placeholder for the
995 result. Finally, emit a CCALL insn with an offset pointing to the
996 Addr# just pushed, and a literal field holding the mallocville
997 address of the piece of marshalling code we generate.
998 So, just prior to the CCALL insn, the stack looks like this
999 (growing down, as usual):
1004 Addr# address_of_C_fn
1005 <placeholder-for-result#> (must be an unboxed type)
1007 The interpreter then calls the marshall code mentioned
1008 in the CCALL insn, passing it (& <placeholder-for-result#>),
1009 that is, the addr of the topmost word in the stack.
1010 When this returns, the placeholder will have been
1011 filled in. The placeholder is slid down to the sequel
1012 depth, and we RETURN.
1014 This arrangement makes it simple to do f-i-dynamic since the Addr#
1015 value is the first arg anyway.
1017 The marshalling code is generated specifically for this
1018 call site, and so knows exactly the (Haskell) stack
1019 offsets of the args, fn address and placeholder. It
1020 copies the args to the C stack, calls the stacked addr,
1021 and parks the result back in the placeholder. The interpreter
1022 calls it as a normal C call, assuming it has a signature
1023 void marshall_code ( StgWord* ptr_to_top_of_stack )
1025 -- resolve static address
1029 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
1031 -> do res <- ioToBc (lookupStaticPtr stdcall_adj_target)
1035 #ifdef mingw32_TARGET_OS
1036 | StdCallConv <- cconv
1037 = let size = a_reps_sizeW * wORD_SIZE in
1038 mkFastString (unpackFS target ++ '@':show size)
1044 (is_static, static_target_addr) <- get_target_info
1047 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1048 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1049 | is_static = a_reps_pushed_RAW
1050 | otherwise = if null a_reps_pushed_RAW
1051 then panic "ByteCodeGen.generateCCall: dyn with no args"
1052 else tail a_reps_pushed_RAW
1055 (push_Addr, d_after_Addr)
1057 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1058 d_after_args + addr_sizeW)
1059 | otherwise -- is already on the stack
1060 = (nilOL, d_after_args)
1062 -- Push the return placeholder. For a call returning nothing,
1063 -- this is a VoidArg (tag).
1064 r_sizeW = fromIntegral (primRepSizeW r_rep)
1065 d_after_r = d_after_Addr + r_sizeW
1066 r_lit = mkDummyLiteral r_rep
1067 push_r = (if returns_void
1069 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1071 -- generate the marshalling code we're going to call
1073 -- Offset of the next stack frame down the stack. The CCALL
1074 -- instruction needs to describe the chunk of stack containing
1075 -- the ccall args to the GC, so it needs to know how large it
1076 -- is. See comment in Interpreter.c with the CCALL instruction.
1077 stk_offset = d_after_r - s
1080 -- the only difference in libffi mode is that we prepare a cif
1081 -- describing the call type by calling libffi, and we attach the
1082 -- address of this to the CCALL instruction.
1083 token <- ioToBc $ prepForeignCall cconv a_reps r_rep
1084 let addr_of_marshaller = castPtrToFunPtr token
1086 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1089 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1091 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1092 `snocOL` RETURN_UBX (primRepToCgRep r_rep)
1094 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1097 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1100 -- Make a dummy literal, to be used as a placeholder for FFI return
1101 -- values on the stack.
1102 mkDummyLiteral :: PrimRep -> Literal
1106 WordRep -> MachWord 0
1107 AddrRep -> MachNullAddr
1108 DoubleRep -> MachDouble 0
1109 FloatRep -> MachFloat 0
1110 Int64Rep -> MachInt64 0
1111 Word64Rep -> MachWord64 0
1112 _ -> panic "mkDummyLiteral"
1116 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1117 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1120 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1122 -- Alternatively, for call-targets returning nothing, convert
1124 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1125 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1129 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
1130 maybe_getCCallReturnRep fn_ty
1131 = let (_a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1133 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1135 = case splitTyConApp_maybe (repType r_ty) of
1136 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
1138 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
1139 || r_reps == [VoidRep] )
1140 && isUnboxedTupleTyCon r_tycon
1141 && case maybe_r_rep_to_go of
1143 Just r_rep -> r_rep /= PtrRep
1144 -- if it was, it would be impossible
1145 -- to create a valid return value
1146 -- placeholder on the stack
1147 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1150 --trace (showSDoc (ppr (a_reps, r_reps))) $
1151 if ok then maybe_r_rep_to_go else blargh
1153 -- Compile code which expects an unboxed Int on the top of stack,
1154 -- (call it i), and pushes the i'th closure in the supplied list
1155 -- as a consequence.
1156 implement_tagToId :: [Name] -> BcM BCInstrList
1157 implement_tagToId names
1158 = ASSERT( notNull names )
1159 do labels <- getLabelsBc (genericLength names)
1160 label_fail <- getLabelBc
1161 label_exit <- getLabelBc
1162 let infos = zip4 labels (tail labels ++ [label_fail])
1164 steps = map (mkStep label_exit) infos
1165 return (concatOL steps
1167 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1169 mkStep l_exit (my_label, next_label, n, name_for_n)
1170 = toOL [LABEL my_label,
1171 TESTEQ_I n next_label,
1176 -- -----------------------------------------------------------------------------
1179 -- Push an atom onto the stack, returning suitable code & number of
1180 -- stack words used.
1182 -- The env p must map each variable to the highest- numbered stack
1183 -- slot for it. For example, if the stack has depth 4 and we
1184 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1185 -- the tag in stack[5], the stack will have depth 6, and p must map v
1186 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1187 -- depth 6 stack has valid words 0 .. 5.
1189 pushAtom :: Word16 -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Word16)
1192 | Just e' <- bcView e
1195 pushAtom d p (AnnVar v)
1196 | idCgRep v == VoidArg
1200 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1202 | Just primop <- isPrimOpId_maybe v
1203 = return (unitOL (PUSH_PRIMOP primop), 1)
1205 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1206 = let l = d - d_v + sz - 2
1207 in return (toOL (genericReplicate sz (PUSH_L l)), sz)
1208 -- d - d_v the number of words between the TOS
1209 -- and the 1st slot of the object
1211 -- d - d_v - 1 the offset from the TOS of the 1st slot
1213 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1216 -- Having found the last slot, we proceed to copy the right number of
1217 -- slots on to the top of the stack.
1219 | otherwise -- v must be a global variable
1221 return (unitOL (PUSH_G (getName v)), sz)
1225 sz = fromIntegral (idSizeW v)
1228 pushAtom _ _ (AnnLit lit)
1230 MachLabel _ _ _ -> code NonPtrArg
1231 MachWord _ -> code NonPtrArg
1232 MachInt _ -> code PtrArg
1233 MachFloat _ -> code FloatArg
1234 MachDouble _ -> code DoubleArg
1235 MachChar _ -> code NonPtrArg
1236 MachNullAddr -> code NonPtrArg
1237 MachStr s -> pushStr s
1238 l -> pprPanic "pushAtom" (ppr l)
1241 = let size_host_words = fromIntegral (cgRepSizeW rep)
1242 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1246 = let getMallocvilleAddr
1248 FastString _ n _ fp _ ->
1249 -- we could grab the Ptr from the ForeignPtr,
1250 -- but then we have no way to control its lifetime.
1251 -- In reality it'll probably stay alive long enoungh
1252 -- by virtue of the global FastString table, but
1253 -- to be on the safe side we copy the string into
1254 -- a malloc'd area of memory.
1255 do ptr <- ioToBc (mallocBytes (n+1))
1258 withForeignPtr fp $ \p -> do
1259 memcpy ptr p (fromIntegral n)
1260 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1264 addr <- getMallocvilleAddr
1265 -- Get the addr on the stack, untaggedly
1266 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1268 pushAtom d p (AnnCast e _)
1269 = pushAtom d p (snd e)
1272 = pprPanic "ByteCodeGen.pushAtom"
1273 (pprCoreExpr (deAnnotate (undefined, expr)))
1275 foreign import ccall unsafe "memcpy"
1276 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1279 -- -----------------------------------------------------------------------------
1280 -- Given a bunch of alts code and their discrs, do the donkey work
1281 -- of making a multiway branch using a switch tree.
1282 -- What a load of hassle!
1284 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1285 -- a hint; generates better code
1286 -- Nothing is always safe
1287 -> [(Discr, BCInstrList)]
1289 mkMultiBranch maybe_ncons raw_ways
1290 = let d_way = filter (isNoDiscr.fst) raw_ways
1292 (\w1 w2 -> leAlt (fst w1) (fst w2))
1293 (filter (not.isNoDiscr.fst) raw_ways)
1295 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1296 mkTree [] _range_lo _range_hi = return the_default
1298 mkTree [val] range_lo range_hi
1299 | range_lo `eqAlt` range_hi
1302 = do label_neq <- getLabelBc
1303 return (mkTestEQ (fst val) label_neq
1305 `appOL` unitOL (LABEL label_neq)
1306 `appOL` the_default))
1308 mkTree vals range_lo range_hi
1309 = let n = length vals `div` 2
1310 vals_lo = take n vals
1311 vals_hi = drop n vals
1312 v_mid = fst (head vals_hi)
1314 label_geq <- getLabelBc
1315 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1316 code_hi <- mkTree vals_hi v_mid range_hi
1317 return (mkTestLT v_mid label_geq
1319 `appOL` unitOL (LABEL label_geq)
1323 = case d_way of [] -> unitOL CASEFAIL
1325 _ -> panic "mkMultiBranch/the_default"
1327 -- None of these will be needed if there are no non-default alts
1328 (mkTestLT, mkTestEQ, init_lo, init_hi)
1330 = panic "mkMultiBranch: awesome foursome"
1332 = case fst (head notd_ways) of {
1333 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1334 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1337 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1338 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1341 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1342 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1345 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1346 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1348 DiscrP algMaxBound );
1349 NoDiscr -> panic "mkMultiBranch NoDiscr"
1352 (algMinBound, algMaxBound)
1353 = case maybe_ncons of
1354 -- XXX What happens when n == 0?
1355 Just n -> (0, fromIntegral n - 1)
1356 Nothing -> (minBound, maxBound)
1358 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1359 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1360 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1361 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1362 NoDiscr `eqAlt` NoDiscr = True
1365 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1366 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1367 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1368 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1369 NoDiscr `leAlt` NoDiscr = True
1372 isNoDiscr NoDiscr = True
1375 dec (DiscrI i) = DiscrI (i-1)
1376 dec (DiscrP i) = DiscrP (i-1)
1377 dec other = other -- not really right, but if you
1378 -- do cases on floating values, you'll get what you deserve
1380 -- same snotty comment applies to the following
1382 minD, maxD :: Double
1388 mkTree notd_ways init_lo init_hi
1391 -- -----------------------------------------------------------------------------
1392 -- Supporting junk for the compilation schemes
1394 -- Describes case alts
1402 instance Outputable Discr where
1403 ppr (DiscrI i) = int i
1404 ppr (DiscrF f) = text (show f)
1405 ppr (DiscrD d) = text (show d)
1406 ppr (DiscrP i) = ppr i
1407 ppr NoDiscr = text "DEF"
1410 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Word16
1411 lookupBCEnv_maybe = lookupFM
1413 idSizeW :: Id -> Int
1414 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1417 unboxedTupleException :: a
1418 unboxedTupleException
1421 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1422 " Possibly due to foreign import/export decls in source.\n"++
1423 " Workaround: use -fobject-code, or compile this module to .o separately."))
1426 mkSLIDE :: Word16 -> Word16 -> OrdList BCInstr
1427 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1429 splitApp :: AnnExpr' Var ann -> (AnnExpr' Var ann, [AnnExpr' Var ann])
1430 -- The arguments are returned in *right-to-left* order
1431 splitApp e | Just e' <- bcView e = splitApp e'
1432 splitApp (AnnApp (_,f) (_,a)) = case splitApp f of
1433 (f', as) -> (f', a:as)
1434 splitApp e = (e, [])
1437 bcView :: AnnExpr' Var ann -> Maybe (AnnExpr' Var ann)
1438 -- The "bytecode view" of a term discards
1439 -- a) type abstractions
1440 -- b) type applications
1443 -- Type lambdas *can* occur in random expressions,
1444 -- whereas value lambdas cannot; that is why they are nuked here
1445 bcView (AnnNote _ (_,e)) = Just e
1446 bcView (AnnCast (_,e) _) = Just e
1447 bcView (AnnLam v (_,e)) | isTyVar v = Just e
1448 bcView (AnnApp (_,e) (_, AnnType _)) = Just e
1451 isVoidArgAtom :: AnnExpr' Var ann -> Bool
1452 isVoidArgAtom e | Just e' <- bcView e = isVoidArgAtom e'
1453 isVoidArgAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1454 isVoidArgAtom _ = False
1456 atomPrimRep :: AnnExpr' Id ann -> PrimRep
1457 atomPrimRep e | Just e' <- bcView e = atomPrimRep e'
1458 atomPrimRep (AnnVar v) = typePrimRep (idType v)
1459 atomPrimRep (AnnLit l) = typePrimRep (literalType l)
1460 atomPrimRep other = pprPanic "atomPrimRep" (ppr (deAnnotate (undefined,other)))
1462 atomRep :: AnnExpr' Id ann -> CgRep
1463 atomRep e = primRepToCgRep (atomPrimRep e)
1465 isPtrAtom :: AnnExpr' Id ann -> Bool
1466 isPtrAtom e = atomRep e == PtrArg
1468 -- Let szsw be the sizes in words of some items pushed onto the stack,
1469 -- which has initial depth d'. Return the values which the stack environment
1470 -- should map these items to.
1471 mkStackOffsets :: Word16 -> [Word16] -> [Word16]
1472 mkStackOffsets original_depth szsw
1473 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1475 -- -----------------------------------------------------------------------------
1476 -- The bytecode generator's monad
1478 type BcPtr = Either ItblPtr (Ptr ())
1482 uniqSupply :: UniqSupply, -- for generating fresh variable names
1483 nextlabel :: Word16, -- for generating local labels
1484 malloced :: [BcPtr], -- thunks malloced for current BCO
1485 -- Should be free()d when it is GCd
1486 breakArray :: BreakArray -- array of breakpoint flags
1489 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1491 ioToBc :: IO a -> BcM a
1492 ioToBc io = BcM $ \st -> do
1496 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1497 runBc us modBreaks (BcM m)
1498 = m (BcM_State us 0 [] breakArray)
1500 breakArray = modBreaks_flags modBreaks
1502 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1503 thenBc (BcM expr) cont = BcM $ \st0 -> do
1504 (st1, q) <- expr st0
1509 thenBc_ :: BcM a -> BcM b -> BcM b
1510 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1511 (st1, _) <- expr st0
1512 (st2, r) <- cont st1
1515 returnBc :: a -> BcM a
1516 returnBc result = BcM $ \st -> (return (st, result))
1518 instance Monad BcM where
1523 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1525 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1527 recordMallocBc :: Ptr a -> BcM ()
1529 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1531 recordItblMallocBc :: ItblPtr -> BcM ()
1532 recordItblMallocBc a
1533 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1535 getLabelBc :: BcM Word16
1537 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1539 getLabelsBc :: Word16 -> BcM [Word16]
1541 = BcM $ \st -> let ctr = nextlabel st
1542 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1544 getBreakArray :: BcM BreakArray
1545 getBreakArray = BcM $ \st -> return (st, breakArray st)
1547 newUnique :: BcM Unique
1549 \st -> case splitUniqSupply (uniqSupply st) of
1550 (us1, us2) -> let newState = st { uniqSupply = us2 }
1551 in return (newState, uniqFromSupply us1)
1553 newId :: Type -> BcM Id
1556 return $ mkSysLocal tickFS uniq ty
1558 tickFS :: FastString
1559 tickFS = fsLit "ticked"