2 % (c) The University of Glasgow 2002-2006
5 ByteCodeGen: Generate bytecode from Core
8 module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where
10 #include "HsVersions.h"
53 import GHC.Exts ( Int(..), ByteArray# )
55 import Control.Monad ( when )
64 -- -----------------------------------------------------------------------------
65 -- Generating byte code for a complete module
67 byteCodeGen :: DynFlags
71 -> IO CompiledByteCode
72 byteCodeGen dflags binds tycs modBreaks
73 = do showPass dflags "ByteCodeGen"
75 let flatBinds = [ (bndr, freeVars rhs)
76 | (bndr, rhs) <- flattenBinds binds]
78 us <- mkSplitUniqSupply 'y'
79 (BcM_State _us _final_ctr mallocd _, proto_bcos)
80 <- runBc us modBreaks (mapM schemeTopBind flatBinds)
82 when (notNull mallocd)
83 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
85 dumpIfSet_dyn dflags Opt_D_dump_BCOs
86 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
88 assembleBCOs proto_bcos tycs
90 -- -----------------------------------------------------------------------------
91 -- Generating byte code for an expression
93 -- Returns: (the root BCO for this expression,
94 -- a list of auxilary BCOs resulting from compiling closures)
95 coreExprToBCOs :: DynFlags
98 coreExprToBCOs dflags expr
99 = do showPass dflags "ByteCodeGen"
101 -- create a totally bogus name for the top-level BCO; this
102 -- should be harmless, since it's never used for anything
103 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) (fsLit "ExprTopLevel")
104 invented_id = Id.mkLocalId invented_name (panic "invented_id's type")
106 -- the uniques are needed to generate fresh variables when we introduce new
107 -- let bindings for ticked expressions
108 us <- mkSplitUniqSupply 'y'
109 (BcM_State _us _final_ctr mallocd _ , proto_bco)
110 <- runBc us emptyModBreaks (schemeTopBind (invented_id, freeVars expr))
112 when (notNull mallocd)
113 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
115 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
117 assembleBCO proto_bco
120 -- -----------------------------------------------------------------------------
121 -- Compilation schema for the bytecode generator
123 type BCInstrList = OrdList BCInstr
125 type Sequel = Int -- back off to this depth before ENTER
127 -- Maps Ids to the offset from the stack _base_ so we don't have
128 -- to mess with it after each push/pop.
129 type BCEnv = FiniteMap Id Int -- To find vars on the stack
132 ppBCEnv :: BCEnv -> SDoc
135 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
138 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
139 cmp_snd x y = compare (snd x) (snd y)
142 -- Create a BCO and do a spot of peephole optimisation on the insns
147 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
151 -> Bool -- True <=> is a return point, rather than a function
154 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap is_ret mallocd_blocks
157 protoBCOInstrs = maybe_with_stack_check,
158 protoBCOBitmap = bitmap,
159 protoBCOBitmapSize = bitmap_size,
160 protoBCOArity = arity,
161 protoBCOExpr = origin,
162 protoBCOPtrs = mallocd_blocks
165 -- Overestimate the stack usage (in words) of this BCO,
166 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
167 -- stack check. (The interpreter always does a stack check
168 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
169 -- BCO anyway, so we only need to add an explicit on in the
170 -- (hopefully rare) cases when the (overestimated) stack use
171 -- exceeds iNTERP_STACK_CHECK_THRESH.
172 maybe_with_stack_check
173 | is_ret && stack_usage < aP_STACK_SPLIM = peep_d
174 -- don't do stack checks at return points,
175 -- everything is aggregated up to the top BCO
176 -- (which must be a function).
177 -- That is, unless the stack usage is >= AP_STACK_SPLIM,
179 | stack_usage >= iNTERP_STACK_CHECK_THRESH
180 = STKCHECK stack_usage : peep_d
182 = peep_d -- the supposedly common case
184 -- We assume that this sum doesn't wrap
185 stack_usage = sum (map bciStackUse peep_d)
187 -- Merge local pushes
188 peep_d = peep (fromOL instrs_ordlist)
190 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
191 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
192 peep (PUSH_L off1 : PUSH_L off2 : rest)
193 = PUSH_LL off1 (off2-1) : peep rest
199 argBits :: [CgRep] -> [Bool]
202 | isFollowableArg rep = False : argBits args
203 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
205 -- -----------------------------------------------------------------------------
208 -- Compile code for the right-hand side of a top-level binding
210 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
213 schemeTopBind (id, rhs)
214 | Just data_con <- isDataConWorkId_maybe id,
215 isNullaryRepDataCon data_con = do
216 -- Special case for the worker of a nullary data con.
217 -- It'll look like this: Nil = /\a -> Nil a
218 -- If we feed it into schemeR, we'll get
220 -- because mkConAppCode treats nullary constructor applications
221 -- by just re-using the single top-level definition. So
222 -- for the worker itself, we must allocate it directly.
223 -- ioToBc (putStrLn $ "top level BCO")
224 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
225 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
228 = schemeR [{- No free variables -}] (id, rhs)
231 -- -----------------------------------------------------------------------------
234 -- Compile code for a right-hand side, to give a BCO that,
235 -- when executed with the free variables and arguments on top of the stack,
236 -- will return with a pointer to the result on top of the stack, after
237 -- removing the free variables and arguments.
239 -- Park the resulting BCO in the monad. Also requires the
240 -- variable to which this value was bound, so as to give the
241 -- resulting BCO a name.
243 schemeR :: [Id] -- Free vars of the RHS, ordered as they
244 -- will appear in the thunk. Empty for
245 -- top-level things, which have no free vars.
246 -> (Id, AnnExpr Id VarSet)
247 -> BcM (ProtoBCO Name)
248 schemeR fvs (nm, rhs)
252 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
253 $$ pprCoreExpr (deAnnotate rhs)
259 = schemeR_wrk fvs nm rhs (collect [] rhs)
261 collect :: [Var] -> AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
262 collect xs (_, AnnNote _ e) = collect xs e
263 collect xs (_, AnnCast e _) = collect xs e
264 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
265 collect xs (_, not_lambda) = (reverse xs, not_lambda)
267 schemeR_wrk :: [Id] -> Id -> AnnExpr Id VarSet -> ([Var], AnnExpr' Var VarSet) -> BcM (ProtoBCO Name)
268 schemeR_wrk fvs nm original_body (args, body)
270 all_args = reverse args ++ fvs
271 arity = length all_args
272 -- all_args are the args in reverse order. We're compiling a function
273 -- \fv1..fvn x1..xn -> e
274 -- i.e. the fvs come first
276 szsw_args = map idSizeW all_args
277 szw_args = sum szsw_args
278 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
280 -- make the arg bitmap
281 bits = argBits (reverse (map idCgRep all_args))
282 bitmap_size = length bits
283 bitmap = mkBitmap bits
285 body_code <- schemeER_wrk szw_args p_init body
287 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
288 arity bitmap_size bitmap False{-not alts-})
290 -- introduce break instructions for ticked expressions
291 schemeER_wrk :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
293 | Just (tickInfo, (_annot, newRhs)) <- isTickedExp' rhs = do
294 code <- schemeE d 0 p newRhs
296 let idOffSets = getVarOffSets d p tickInfo
297 let tickNumber = tickInfo_number tickInfo
298 let breakInfo = BreakInfo
299 { breakInfo_module = tickInfo_module tickInfo
300 , breakInfo_number = tickNumber
301 , breakInfo_vars = idOffSets
302 , breakInfo_resty = exprType (deAnnotate' newRhs)
304 let breakInstr = case arr of (BA arr#) -> BRK_FUN arr# tickNumber breakInfo
305 return $ breakInstr `consOL` code
306 | otherwise = schemeE d 0 p rhs
308 getVarOffSets :: Int -> BCEnv -> TickInfo -> [(Id, Int)]
309 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
311 getOffSet :: Int -> BCEnv -> Id -> Maybe (Id, Int)
313 = case lookupBCEnv_maybe env id of
315 Just offset -> Just (id, d - offset)
317 fvsToEnv :: BCEnv -> VarSet -> [Id]
318 -- Takes the free variables of a right-hand side, and
319 -- delivers an ordered list of the local variables that will
320 -- be captured in the thunk for the RHS
321 -- The BCEnv argument tells which variables are in the local
322 -- environment: these are the ones that should be captured
324 -- The code that constructs the thunk, and the code that executes
325 -- it, have to agree about this layout
326 fvsToEnv p fvs = [v | v <- varSetElems fvs,
327 isId v, -- Could be a type variable
330 -- -----------------------------------------------------------------------------
335 { tickInfo_number :: Int -- the (module) unique number of the tick
336 , tickInfo_module :: Module -- the origin of the ticked expression
337 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
340 instance Outputable TickInfo where
341 ppr info = text "TickInfo" <+>
342 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
343 ppr (tickInfo_locals info))
345 -- Compile code to apply the given expression to the remaining args
346 -- on the stack, returning a HNF.
347 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
349 -- Delegate tail-calls to schemeT.
350 schemeE d s p e@(AnnApp _ _)
353 schemeE d s p e@(AnnVar v)
354 | not (isUnLiftedType v_type)
355 = -- Lifted-type thing; push it in the normal way
359 = do -- Returning an unlifted value.
360 -- Heave it on the stack, SLIDE, and RETURN.
361 (push, szw) <- pushAtom d p (AnnVar v)
362 return (push -- value onto stack
363 `appOL` mkSLIDE szw (d-s) -- clear to sequel
364 `snocOL` RETURN_UBX v_rep) -- go
367 v_rep = typeCgRep v_type
369 schemeE d s p (AnnLit literal)
370 = do (push, szw) <- pushAtom d p (AnnLit literal)
371 let l_rep = typeCgRep (literalType literal)
372 return (push -- value onto stack
373 `appOL` mkSLIDE szw (d-s) -- clear to sequel
374 `snocOL` RETURN_UBX l_rep) -- go
376 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
377 | (AnnVar v, args_r_to_l) <- splitApp rhs,
378 Just data_con <- isDataConWorkId_maybe v,
379 dataConRepArity data_con == length args_r_to_l
380 = do -- Special case for a non-recursive let whose RHS is a
381 -- saturatred constructor application.
382 -- Just allocate the constructor and carry on
383 alloc_code <- mkConAppCode d s p data_con args_r_to_l
384 body_code <- schemeE (d+1) s (addToFM p x d) body
385 return (alloc_code `appOL` body_code)
387 -- General case for let. Generates correct, if inefficient, code in
389 schemeE d s p (AnnLet binds (_,body))
390 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
391 AnnRec xs_n_rhss -> unzip xs_n_rhss
394 fvss = map (fvsToEnv p' . fst) rhss
396 -- Sizes of free vars
397 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
399 -- the arity of each rhs
400 arities = map (length . fst . collect []) rhss
402 -- This p', d' defn is safe because all the items being pushed
403 -- are ptrs, so all have size 1. d' and p' reflect the stack
404 -- after the closures have been allocated in the heap (but not
405 -- filled in), and pointers to them parked on the stack.
406 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
408 zipE = zipEqual "schemeE"
410 -- ToDo: don't build thunks for things with no free variables
411 build_thunk _ [] size bco off arity
412 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
414 mkap | arity == 0 = MKAP
416 build_thunk dd (fv:fvs) size bco off arity = do
417 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
418 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
419 return (push_code `appOL` more_push_code)
421 alloc_code = toOL (zipWith mkAlloc sizes arities)
423 | is_tick = ALLOC_AP_NOUPD sz
424 | otherwise = ALLOC_AP sz
425 mkAlloc sz arity = ALLOC_PAP arity sz
427 is_tick = case binds of
428 AnnNonRec id _ -> occNameFS (getOccName id) == tickFS
431 compile_bind d' fvs x rhs size arity off = do
432 bco <- schemeR fvs (x,rhs)
433 build_thunk d' fvs size bco off arity
436 [ compile_bind d' fvs x rhs size arity n
437 | (fvs, x, rhs, size, arity, n) <-
438 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
441 body_code <- schemeE d' s p' body
442 thunk_codes <- sequence compile_binds
443 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
445 -- introduce a let binding for a ticked case expression. This rule
446 -- *should* only fire when the expression was not already let-bound
447 -- (the code gen for let bindings should take care of that). Todo: we
448 -- call exprFreeVars on a deAnnotated expression, this may not be the
449 -- best way to calculate the free vars but it seemed like the least
450 -- intrusive thing to do
451 schemeE d s p exp@(AnnCase {})
452 | Just (_tickInfo, rhs) <- isTickedExp' exp
453 = if isUnLiftedType ty
454 then schemeE d s p (snd rhs)
457 -- Todo: is emptyVarSet correct on the next line?
458 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
460 where exp' = deAnnotate' exp
461 fvs = exprFreeVars exp'
464 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1, bind2], rhs)])
465 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
467 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
469 -- case .... of a { DEFAULT -> ... }
470 -- becuse the return convention for both are identical.
472 -- Note that it does not matter losing the void-rep thing from the
473 -- envt (it won't be bound now) because we never look such things up.
475 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
476 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
478 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
479 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
480 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
482 schemeE d s p (AnnCase scrut _ _ [(DataAlt dc, [bind1], rhs)])
483 | isUnboxedTupleCon dc
484 -- Similarly, convert
485 -- case .... of x { (# a #) -> ... }
487 -- case .... of a { DEFAULT -> ... }
488 = --trace "automagic mashing of case alts (# a #)" $
489 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
491 schemeE d s p (AnnCase scrut bndr _ alts)
492 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
494 schemeE d s p (AnnNote _ (_, body))
497 schemeE d s p (AnnCast (_, body) _)
501 = pprPanic "ByteCodeGen.schemeE: unhandled case"
502 (pprCoreExpr (deAnnotate' expr))
508 A ticked expression looks like this:
510 case tick<n> var1 ... varN of DEFAULT -> e
512 (*) <n> is the number of the tick, which is unique within a module
513 (*) var1 ... varN are the local variables in scope at the tick site
515 If we find a ticked expression we return:
517 Just ((n, [var1 ... varN]), e)
519 otherwise we return Nothing.
521 The idea is that the "case tick<n> ..." is really just an annotation on
522 the code. When we find such a thing, we pull out the useful information,
523 and then compile the code as if it was just the expression "e".
527 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
528 isTickedExp' (AnnCase scrut _bndr _type alts)
529 | Just tickInfo <- isTickedScrut scrut,
530 [(DEFAULT, _bndr, rhs)] <- alts
531 = Just (tickInfo, rhs)
533 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
536 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
537 = Just $ TickInfo { tickInfo_number = tickNumber
538 , tickInfo_module = modName
539 , tickInfo_locals = idsOfArgs args
541 | otherwise = Nothing
543 (f, args) = collectArgs $ deAnnotate expr
544 idsOfArgs :: [Expr Id] -> [Id]
545 idsOfArgs = catMaybes . map exprId
546 exprId :: Expr Id -> Maybe Id
547 exprId (Var id) = Just id
550 isTickedExp' _ = Nothing
552 -- Compile code to do a tail call. Specifically, push the fn,
553 -- slide the on-stack app back down to the sequel depth,
554 -- and enter. Four cases:
557 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
558 -- The int will be on the stack. Generate a code sequence
559 -- to convert it to the relevant constructor, SLIDE and ENTER.
561 -- 1. The fn denotes a ccall. Defer to generateCCall.
563 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
564 -- it simply as b -- since the representations are identical
565 -- (the VoidArg takes up zero stack space). Also, spot
566 -- (# b #) and treat it as b.
568 -- 3. Application of a constructor, by defn saturated.
569 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
570 -- then the ptrs, and then do PACK and RETURN.
572 -- 4. Otherwise, it must be a function call. Push the args
573 -- right to left, SLIDE and ENTER.
575 schemeT :: Int -- Stack depth
576 -> Sequel -- Sequel depth
577 -> BCEnv -- stack env
578 -> AnnExpr' Id VarSet
583 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
584 -- = panic "schemeT ?!?!"
586 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
590 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
591 = do (push, arg_words) <- pushAtom d p arg
592 tagToId_sequence <- implement_tagToId constr_names
593 return (push `appOL` tagToId_sequence
594 `appOL` mkSLIDE 1 (d+arg_words-s)
598 | Just (CCall ccall_spec) <- isFCallId_maybe fn
599 = generateCCall d s p ccall_spec fn args_r_to_l
601 -- Case 2: Constructor application
602 | Just con <- maybe_saturated_dcon,
603 isUnboxedTupleCon con
604 = case args_r_to_l of
605 [arg1,arg2] | isVoidArgAtom arg1 ->
606 unboxedTupleReturn d s p arg2
607 [arg1,arg2] | isVoidArgAtom arg2 ->
608 unboxedTupleReturn d s p arg1
609 _other -> unboxedTupleException
611 -- Case 3: Ordinary data constructor
612 | Just con <- maybe_saturated_dcon
613 = do alloc_con <- mkConAppCode d s p con args_r_to_l
614 return (alloc_con `appOL`
615 mkSLIDE 1 (d - s) `snocOL`
618 -- Case 4: Tail call of function
620 = doTailCall d s p fn args_r_to_l
623 -- Detect and extract relevant info for the tagToEnum kludge.
624 maybe_is_tagToEnum_call
625 = let extract_constr_Names ty
626 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
628 = map (getName . dataConWorkId) (tyConDataCons tyc)
629 -- NOTE: use the worker name, not the source name of
630 -- the DataCon. See DataCon.lhs for details.
632 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
635 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
636 -> case isPrimOpId_maybe v of
637 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
641 -- Extract the args (R->L) and fn
642 -- The function will necessarily be a variable,
643 -- because we are compiling a tail call
644 (AnnVar fn, args_r_to_l) = splitApp app
646 -- Only consider this to be a constructor application iff it is
647 -- saturated. Otherwise, we'll call the constructor wrapper.
648 n_args = length args_r_to_l
650 = case isDataConWorkId_maybe fn of
651 Just con | dataConRepArity con == n_args -> Just con
654 -- -----------------------------------------------------------------------------
655 -- Generate code to build a constructor application,
656 -- leaving it on top of the stack
658 mkConAppCode :: Int -> Sequel -> BCEnv
659 -> DataCon -- The data constructor
660 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
663 mkConAppCode _ _ _ con [] -- Nullary constructor
664 = ASSERT( isNullaryRepDataCon con )
665 return (unitOL (PUSH_G (getName (dataConWorkId con))))
666 -- Instead of doing a PACK, which would allocate a fresh
667 -- copy of this constructor, use the single shared version.
669 mkConAppCode orig_d _ p con args_r_to_l
670 = ASSERT( dataConRepArity con == length args_r_to_l )
671 do_pushery orig_d (non_ptr_args ++ ptr_args)
673 -- The args are already in reverse order, which is the way PACK
674 -- expects them to be. We must push the non-ptrs after the ptrs.
675 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
677 do_pushery d (arg:args)
678 = do (push, arg_words) <- pushAtom d p arg
679 more_push_code <- do_pushery (d+arg_words) args
680 return (push `appOL` more_push_code)
682 = return (unitOL (PACK con n_arg_words))
684 n_arg_words = d - orig_d
687 -- -----------------------------------------------------------------------------
688 -- Returning an unboxed tuple with one non-void component (the only
689 -- case we can handle).
691 -- Remember, we don't want to *evaluate* the component that is being
692 -- returned, even if it is a pointed type. We always just return.
695 :: Int -> Sequel -> BCEnv
696 -> AnnExpr' Id VarSet -> BcM BCInstrList
697 unboxedTupleReturn d s p arg = do
698 (push, sz) <- pushAtom d p arg
700 mkSLIDE sz (d-s) `snocOL`
701 RETURN_UBX (atomRep arg))
703 -- -----------------------------------------------------------------------------
704 -- Generate code for a tail-call
707 :: Int -> Sequel -> BCEnv
708 -> Id -> [AnnExpr' Id VarSet]
710 doTailCall init_d s p fn args
711 = do_pushes init_d args (map atomRep args)
713 do_pushes d [] reps = do
714 ASSERT( null reps ) return ()
715 (push_fn, sz) <- pushAtom d p (AnnVar fn)
716 ASSERT( sz == 1 ) return ()
717 return (push_fn `appOL` (
718 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
720 do_pushes d args reps = do
721 let (push_apply, n, rest_of_reps) = findPushSeq reps
722 (these_args, rest_of_args) = splitAt n args
723 (next_d, push_code) <- push_seq d these_args
724 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
725 -- ^^^ for the PUSH_APPLY_ instruction
726 return (push_code `appOL` (push_apply `consOL` instrs))
728 push_seq d [] = return (d, nilOL)
729 push_seq d (arg:args) = do
730 (push_code, sz) <- pushAtom d p arg
731 (final_d, more_push_code) <- push_seq (d+sz) args
732 return (final_d, push_code `appOL` more_push_code)
734 -- v. similar to CgStackery.findMatch, ToDo: merge
735 findPushSeq :: [CgRep] -> (BCInstr, Int, [CgRep])
736 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
737 = (PUSH_APPLY_PPPPPP, 6, rest)
738 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
739 = (PUSH_APPLY_PPPPP, 5, rest)
740 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
741 = (PUSH_APPLY_PPPP, 4, rest)
742 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
743 = (PUSH_APPLY_PPP, 3, rest)
744 findPushSeq (PtrArg: PtrArg: rest)
745 = (PUSH_APPLY_PP, 2, rest)
746 findPushSeq (PtrArg: rest)
747 = (PUSH_APPLY_P, 1, rest)
748 findPushSeq (VoidArg: rest)
749 = (PUSH_APPLY_V, 1, rest)
750 findPushSeq (NonPtrArg: rest)
751 = (PUSH_APPLY_N, 1, rest)
752 findPushSeq (FloatArg: rest)
753 = (PUSH_APPLY_F, 1, rest)
754 findPushSeq (DoubleArg: rest)
755 = (PUSH_APPLY_D, 1, rest)
756 findPushSeq (LongArg: rest)
757 = (PUSH_APPLY_L, 1, rest)
759 = panic "ByteCodeGen.findPushSeq"
761 -- -----------------------------------------------------------------------------
764 doCase :: Int -> Sequel -> BCEnv
765 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
766 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
768 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
770 -- Top of stack is the return itbl, as usual.
771 -- underneath it is the pointer to the alt_code BCO.
772 -- When an alt is entered, it assumes the returned value is
773 -- on top of the itbl.
776 -- An unlifted value gets an extra info table pushed on top
777 -- when it is returned.
778 unlifted_itbl_sizeW | isAlgCase = 0
781 -- depth of stack after the return value has been pushed
782 d_bndr = d + ret_frame_sizeW + idSizeW bndr
784 -- depth of stack after the extra info table for an unboxed return
785 -- has been pushed, if any. This is the stack depth at the
787 d_alts = d_bndr + unlifted_itbl_sizeW
789 -- Env in which to compile the alts, not including
790 -- any vars bound by the alts themselves
791 p_alts = addToFM p bndr (d_bndr - 1)
793 bndr_ty = idType bndr
794 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
796 -- given an alt, return a discr and code for it.
797 codeAlt (DEFAULT, _, (_,rhs))
798 = do rhs_code <- schemeE d_alts s p_alts rhs
799 return (NoDiscr, rhs_code)
801 codeAlt alt@(_, bndrs, (_,rhs))
802 -- primitive or nullary constructor alt: no need to UNPACK
803 | null real_bndrs = do
804 rhs_code <- schemeE d_alts s p_alts rhs
805 return (my_discr alt, rhs_code)
806 -- algebraic alt with some binders
809 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
810 ptr_sizes = map idSizeW ptrs
811 nptrs_sizes = map idSizeW nptrs
812 bind_sizes = ptr_sizes ++ nptrs_sizes
813 size = sum ptr_sizes + sum nptrs_sizes
814 -- the UNPACK instruction unpacks in reverse order...
815 p' = addListToFM p_alts
816 (zip (reverse (ptrs ++ nptrs))
817 (mkStackOffsets d_alts (reverse bind_sizes)))
820 rhs_code <- schemeE (d_alts+size) s p' rhs
821 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
823 real_bndrs = filter (not.isTyVar) bndrs
825 my_discr (DEFAULT, _, _) = NoDiscr {-shouldn't really happen-}
826 my_discr (DataAlt dc, _, _)
827 | isUnboxedTupleCon dc
828 = unboxedTupleException
830 = DiscrP (dataConTag dc - fIRST_TAG)
831 my_discr (LitAlt l, _, _)
832 = case l of MachInt i -> DiscrI (fromInteger i)
833 MachFloat r -> DiscrF (fromRational r)
834 MachDouble r -> DiscrD (fromRational r)
835 MachChar i -> DiscrI (ord i)
836 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
839 | not isAlgCase = Nothing
841 = case [dc | (DataAlt dc, _, _) <- alts] of
843 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
845 -- the bitmap is relative to stack depth d, i.e. before the
846 -- BCO, info table and return value are pushed on.
847 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
848 -- except that here we build the bitmap from the known bindings of
849 -- things that are pointers, whereas in CgBindery the code builds the
850 -- bitmap from the free slots and unboxed bindings.
853 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
854 -- The bitmap must cover the portion of the stack up to the sequel only.
855 -- Previously we were building a bitmap for the whole depth (d), but we
856 -- really want a bitmap up to depth (d-s). This affects compilation of
857 -- case-of-case expressions, which is the only time we can be compiling a
858 -- case expression with s /= 0.
860 bitmap = intsToReverseBitmap bitmap_size{-size-}
861 (sortLe (<=) (filter (< bitmap_size) rel_slots))
864 rel_slots = concat (map spread binds)
866 | isFollowableArg (idCgRep id) = [ rel_offset ]
868 where rel_offset = d - offset - 1
871 alt_stuff <- mapM codeAlt alts
872 alt_final <- mkMultiBranch maybe_ncons alt_stuff
875 alt_bco_name = getName bndr
876 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
877 0{-no arity-} bitmap_size bitmap True{-is alts-}
879 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
880 -- "\n bitmap = " ++ show bitmap) $ do
881 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
882 alt_bco' <- emitBc alt_bco
884 | isAlgCase = PUSH_ALTS alt_bco'
885 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
886 return (push_alts `consOL` scrut_code)
889 -- -----------------------------------------------------------------------------
890 -- Deal with a CCall.
892 -- Taggedly push the args onto the stack R->L,
893 -- deferencing ForeignObj#s and adjusting addrs to point to
894 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
895 -- (machine) code for the ccall, and create bytecodes to call that and
896 -- then return in the right way.
898 generateCCall :: Int -> Sequel -- stack and sequel depths
900 -> CCallSpec -- where to call
901 -> Id -- of target, for type info
902 -> [AnnExpr' Id VarSet] -- args (atoms)
905 generateCCall d0 s p (CCallSpec target cconv _) fn args_r_to_l
908 addr_sizeW = cgRepSizeW NonPtrArg
910 -- Get the args on the stack, with tags and suitably
911 -- dereferenced for the CCall. For each arg, return the
912 -- depth to the first word of the bits for that arg, and the
913 -- CgRep of what was actually pushed.
915 pargs _ [] = return []
917 = let arg_ty = repType (exprType (deAnnotate' a))
919 in case splitTyConApp_maybe arg_ty of
920 -- Don't push the FO; instead push the Addr# it
923 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
924 -> do rest <- pargs (d + addr_sizeW) az
925 code <- parg_ArrayishRep arrPtrsHdrSize d p a
926 return ((code,AddrRep):rest)
928 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
929 -> do rest <- pargs (d + addr_sizeW) az
930 code <- parg_ArrayishRep arrWordsHdrSize d p a
931 return ((code,AddrRep):rest)
933 -- Default case: push taggedly, but otherwise intact.
935 -> do (code_a, sz_a) <- pushAtom d p a
936 rest <- pargs (d+sz_a) az
937 return ((code_a, atomPrimRep a) : rest)
939 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
940 -- the stack but then advance it over the headers, so as to
941 -- point to the payload.
942 parg_ArrayishRep hdrSize d p a
943 = do (push_fo, _) <- pushAtom d p a
944 -- The ptr points at the header. Advance it over the
945 -- header and then pretend this is an Addr#.
946 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
949 code_n_reps <- pargs d0 args_r_to_l
951 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
952 a_reps_sizeW = sum (map primRepSizeW a_reps_pushed_r_to_l)
954 push_args = concatOL pushs_arg
955 d_after_args = d0 + a_reps_sizeW
957 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
958 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
960 = reverse (tail a_reps_pushed_r_to_l)
962 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
963 -- push_args is the code to do that.
964 -- d_after_args is the stack depth once the args are on.
966 -- Get the result rep.
967 (returns_void, r_rep)
968 = case maybe_getCCallReturnRep (idType fn) of
969 Nothing -> (True, VoidRep)
970 Just rr -> (False, rr)
972 Because the Haskell stack grows down, the a_reps refer to
973 lowest to highest addresses in that order. The args for the call
974 are on the stack. Now push an unboxed Addr# indicating
975 the C function to call. Then push a dummy placeholder for the
976 result. Finally, emit a CCALL insn with an offset pointing to the
977 Addr# just pushed, and a literal field holding the mallocville
978 address of the piece of marshalling code we generate.
979 So, just prior to the CCALL insn, the stack looks like this
980 (growing down, as usual):
985 Addr# address_of_C_fn
986 <placeholder-for-result#> (must be an unboxed type)
988 The interpreter then calls the marshall code mentioned
989 in the CCALL insn, passing it (& <placeholder-for-result#>),
990 that is, the addr of the topmost word in the stack.
991 When this returns, the placeholder will have been
992 filled in. The placeholder is slid down to the sequel
993 depth, and we RETURN.
995 This arrangement makes it simple to do f-i-dynamic since the Addr#
996 value is the first arg anyway.
998 The marshalling code is generated specifically for this
999 call site, and so knows exactly the (Haskell) stack
1000 offsets of the args, fn address and placeholder. It
1001 copies the args to the C stack, calls the stacked addr,
1002 and parks the result back in the placeholder. The interpreter
1003 calls it as a normal C call, assuming it has a signature
1004 void marshall_code ( StgWord* ptr_to_top_of_stack )
1006 -- resolve static address
1010 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
1012 -> do res <- ioToBc (lookupStaticPtr stdcall_adj_target)
1016 #ifdef mingw32_TARGET_OS
1017 | StdCallConv <- cconv
1018 = let size = a_reps_sizeW * wORD_SIZE in
1019 mkFastString (unpackFS target ++ '@':show size)
1025 (is_static, static_target_addr) <- get_target_info
1028 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1029 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1030 | is_static = a_reps_pushed_RAW
1031 | otherwise = if null a_reps_pushed_RAW
1032 then panic "ByteCodeGen.generateCCall: dyn with no args"
1033 else tail a_reps_pushed_RAW
1036 (push_Addr, d_after_Addr)
1038 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1039 d_after_args + addr_sizeW)
1040 | otherwise -- is already on the stack
1041 = (nilOL, d_after_args)
1043 -- Push the return placeholder. For a call returning nothing,
1044 -- this is a VoidArg (tag).
1045 r_sizeW = primRepSizeW r_rep
1046 d_after_r = d_after_Addr + r_sizeW
1047 r_lit = mkDummyLiteral r_rep
1048 push_r = (if returns_void
1050 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1052 -- generate the marshalling code we're going to call
1054 -- Offset of the next stack frame down the stack. The CCALL
1055 -- instruction needs to describe the chunk of stack containing
1056 -- the ccall args to the GC, so it needs to know how large it
1057 -- is. See comment in Interpreter.c with the CCALL instruction.
1058 stk_offset = d_after_r - s
1061 -- the only difference in libffi mode is that we prepare a cif
1062 -- describing the call type by calling libffi, and we attach the
1063 -- address of this to the CCALL instruction.
1064 token <- ioToBc $ prepForeignCall cconv a_reps r_rep
1065 let addr_of_marshaller = castPtrToFunPtr token
1067 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1070 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1072 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1073 `snocOL` RETURN_UBX (primRepToCgRep r_rep)
1075 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1078 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1081 -- Make a dummy literal, to be used as a placeholder for FFI return
1082 -- values on the stack.
1083 mkDummyLiteral :: PrimRep -> Literal
1087 WordRep -> MachWord 0
1088 AddrRep -> MachNullAddr
1089 DoubleRep -> MachDouble 0
1090 FloatRep -> MachFloat 0
1091 Int64Rep -> MachInt64 0
1092 Word64Rep -> MachWord64 0
1093 _ -> panic "mkDummyLiteral"
1097 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1098 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1101 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1103 -- Alternatively, for call-targets returning nothing, convert
1105 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1106 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1110 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
1111 maybe_getCCallReturnRep fn_ty
1112 = let (_a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1114 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1116 = case splitTyConApp_maybe (repType r_ty) of
1117 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
1119 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
1120 || r_reps == [VoidRep] )
1121 && isUnboxedTupleTyCon r_tycon
1122 && case maybe_r_rep_to_go of
1124 Just r_rep -> r_rep /= PtrRep
1125 -- if it was, it would be impossible
1126 -- to create a valid return value
1127 -- placeholder on the stack
1128 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1131 --trace (showSDoc (ppr (a_reps, r_reps))) $
1132 if ok then maybe_r_rep_to_go else blargh
1134 -- Compile code which expects an unboxed Int on the top of stack,
1135 -- (call it i), and pushes the i'th closure in the supplied list
1136 -- as a consequence.
1137 implement_tagToId :: [Name] -> BcM BCInstrList
1138 implement_tagToId names
1139 = ASSERT( notNull names )
1140 do labels <- getLabelsBc (length names)
1141 label_fail <- getLabelBc
1142 label_exit <- getLabelBc
1143 let infos = zip4 labels (tail labels ++ [label_fail])
1145 steps = map (mkStep label_exit) infos
1146 return (concatOL steps
1148 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1150 mkStep l_exit (my_label, next_label, n, name_for_n)
1151 = toOL [LABEL my_label,
1152 TESTEQ_I n next_label,
1157 -- -----------------------------------------------------------------------------
1160 -- Push an atom onto the stack, returning suitable code & number of
1161 -- stack words used.
1163 -- The env p must map each variable to the highest- numbered stack
1164 -- slot for it. For example, if the stack has depth 4 and we
1165 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1166 -- the tag in stack[5], the stack will have depth 6, and p must map v
1167 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1168 -- depth 6 stack has valid words 0 .. 5.
1170 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1172 pushAtom d p (AnnApp f (_, AnnType _))
1173 = pushAtom d p (snd f)
1175 pushAtom d p (AnnNote _ e)
1176 = pushAtom d p (snd e)
1178 pushAtom d p (AnnLam x e)
1180 = pushAtom d p (snd e)
1182 pushAtom d p (AnnVar v)
1184 | idCgRep v == VoidArg
1188 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1190 | Just primop <- isPrimOpId_maybe v
1191 = return (unitOL (PUSH_PRIMOP primop), 1)
1193 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1194 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1195 -- d - d_v the number of words between the TOS
1196 -- and the 1st slot of the object
1198 -- d - d_v - 1 the offset from the TOS of the 1st slot
1200 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1203 -- Having found the last slot, we proceed to copy the right number of
1204 -- slots on to the top of the stack.
1206 | otherwise -- v must be a global variable
1208 return (unitOL (PUSH_G (getName v)), sz)
1214 pushAtom _ _ (AnnLit lit)
1216 MachLabel _ _ -> code NonPtrArg
1217 MachWord _ -> code NonPtrArg
1218 MachInt _ -> code PtrArg
1219 MachFloat _ -> code FloatArg
1220 MachDouble _ -> code DoubleArg
1221 MachChar _ -> code NonPtrArg
1222 MachNullAddr -> code NonPtrArg
1223 MachStr s -> pushStr s
1224 l -> pprPanic "pushAtom" (ppr l)
1227 = let size_host_words = cgRepSizeW rep
1228 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1232 = let getMallocvilleAddr
1234 FastString _ n _ fp _ ->
1235 -- we could grab the Ptr from the ForeignPtr,
1236 -- but then we have no way to control its lifetime.
1237 -- In reality it'll probably stay alive long enoungh
1238 -- by virtue of the global FastString table, but
1239 -- to be on the safe side we copy the string into
1240 -- a malloc'd area of memory.
1241 do ptr <- ioToBc (mallocBytes (n+1))
1244 withForeignPtr fp $ \p -> do
1245 memcpy ptr p (fromIntegral n)
1246 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1250 addr <- getMallocvilleAddr
1251 -- Get the addr on the stack, untaggedly
1252 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1254 pushAtom d p (AnnCast e _)
1255 = pushAtom d p (snd e)
1258 = pprPanic "ByteCodeGen.pushAtom"
1259 (pprCoreExpr (deAnnotate (undefined, expr)))
1261 foreign import ccall unsafe "memcpy"
1262 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1265 -- -----------------------------------------------------------------------------
1266 -- Given a bunch of alts code and their discrs, do the donkey work
1267 -- of making a multiway branch using a switch tree.
1268 -- What a load of hassle!
1270 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1271 -- a hint; generates better code
1272 -- Nothing is always safe
1273 -> [(Discr, BCInstrList)]
1275 mkMultiBranch maybe_ncons raw_ways
1276 = let d_way = filter (isNoDiscr.fst) raw_ways
1278 (\w1 w2 -> leAlt (fst w1) (fst w2))
1279 (filter (not.isNoDiscr.fst) raw_ways)
1281 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1282 mkTree [] _range_lo _range_hi = return the_default
1284 mkTree [val] range_lo range_hi
1285 | range_lo `eqAlt` range_hi
1288 = do label_neq <- getLabelBc
1289 return (mkTestEQ (fst val) label_neq
1291 `appOL` unitOL (LABEL label_neq)
1292 `appOL` the_default))
1294 mkTree vals range_lo range_hi
1295 = let n = length vals `div` 2
1296 vals_lo = take n vals
1297 vals_hi = drop n vals
1298 v_mid = fst (head vals_hi)
1300 label_geq <- getLabelBc
1301 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1302 code_hi <- mkTree vals_hi v_mid range_hi
1303 return (mkTestLT v_mid label_geq
1305 `appOL` unitOL (LABEL label_geq)
1309 = case d_way of [] -> unitOL CASEFAIL
1311 _ -> panic "mkMultiBranch/the_default"
1313 -- None of these will be needed if there are no non-default alts
1314 (mkTestLT, mkTestEQ, init_lo, init_hi)
1316 = panic "mkMultiBranch: awesome foursome"
1318 = case fst (head notd_ways) of {
1319 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1320 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1323 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1324 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1327 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1328 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1331 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1332 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1334 DiscrP algMaxBound );
1335 NoDiscr -> panic "mkMultiBranch NoDiscr"
1338 (algMinBound, algMaxBound)
1339 = case maybe_ncons of
1340 Just n -> (0, n - 1)
1341 Nothing -> (minBound, maxBound)
1343 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1344 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1345 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1346 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1347 NoDiscr `eqAlt` NoDiscr = True
1350 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1351 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1352 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1353 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1354 NoDiscr `leAlt` NoDiscr = True
1357 isNoDiscr NoDiscr = True
1360 dec (DiscrI i) = DiscrI (i-1)
1361 dec (DiscrP i) = DiscrP (i-1)
1362 dec other = other -- not really right, but if you
1363 -- do cases on floating values, you'll get what you deserve
1365 -- same snotty comment applies to the following
1367 minD, maxD :: Double
1373 mkTree notd_ways init_lo init_hi
1376 -- -----------------------------------------------------------------------------
1377 -- Supporting junk for the compilation schemes
1379 -- Describes case alts
1387 instance Outputable Discr where
1388 ppr (DiscrI i) = int i
1389 ppr (DiscrF f) = text (show f)
1390 ppr (DiscrD d) = text (show d)
1391 ppr (DiscrP i) = int i
1392 ppr NoDiscr = text "DEF"
1395 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1396 lookupBCEnv_maybe = lookupFM
1398 idSizeW :: Id -> Int
1399 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1402 unboxedTupleException :: a
1403 unboxedTupleException
1406 ("Error: bytecode compiler can't handle unboxed tuples.\n"++
1407 " Possibly due to foreign import/export decls in source.\n"++
1408 " Workaround: use -fobject-code, or compile this module to .o separately."))
1411 mkSLIDE :: Int -> Int -> OrdList BCInstr
1412 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1414 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1415 -- The arguments are returned in *right-to-left* order
1416 splitApp (AnnApp (_,f) (_,a))
1417 | isTypeAtom a = splitApp f
1418 | otherwise = case splitApp f of
1419 (f', as) -> (f', a:as)
1420 splitApp (AnnNote _ (_,e)) = splitApp e
1421 splitApp (AnnCast (_,e) _) = splitApp e
1422 splitApp e = (e, [])
1425 isTypeAtom :: AnnExpr' id ann -> Bool
1426 isTypeAtom (AnnType _) = True
1427 isTypeAtom _ = False
1429 isVoidArgAtom :: AnnExpr' id ann -> Bool
1430 isVoidArgAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1431 isVoidArgAtom (AnnNote _ (_,e)) = isVoidArgAtom e
1432 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1433 isVoidArgAtom _ = False
1435 atomPrimRep :: AnnExpr' Id ann -> PrimRep
1436 atomPrimRep (AnnVar v) = typePrimRep (idType v)
1437 atomPrimRep (AnnLit l) = typePrimRep (literalType l)
1438 atomPrimRep (AnnNote _ b) = atomPrimRep (snd b)
1439 atomPrimRep (AnnApp f (_, AnnType _)) = atomPrimRep (snd f)
1440 atomPrimRep (AnnLam x e) | isTyVar x = atomPrimRep (snd e)
1441 atomPrimRep (AnnCast b _) = atomPrimRep (snd b)
1442 atomPrimRep other = pprPanic "atomPrimRep" (ppr (deAnnotate (undefined,other)))
1444 atomRep :: AnnExpr' Id ann -> CgRep
1445 atomRep e = primRepToCgRep (atomPrimRep e)
1447 isPtrAtom :: AnnExpr' Id ann -> Bool
1448 isPtrAtom e = atomRep e == PtrArg
1450 -- Let szsw be the sizes in words of some items pushed onto the stack,
1451 -- which has initial depth d'. Return the values which the stack environment
1452 -- should map these items to.
1453 mkStackOffsets :: Int -> [Int] -> [Int]
1454 mkStackOffsets original_depth szsw
1455 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1457 -- -----------------------------------------------------------------------------
1458 -- The bytecode generator's monad
1460 type BcPtr = Either ItblPtr (Ptr ())
1464 uniqSupply :: UniqSupply, -- for generating fresh variable names
1465 nextlabel :: Int, -- for generating local labels
1466 malloced :: [BcPtr], -- thunks malloced for current BCO
1467 -- Should be free()d when it is GCd
1468 breakArray :: BreakArray -- array of breakpoint flags
1471 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1473 ioToBc :: IO a -> BcM a
1474 ioToBc io = BcM $ \st -> do
1478 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1479 runBc us modBreaks (BcM m)
1480 = m (BcM_State us 0 [] breakArray)
1482 breakArray = modBreaks_flags modBreaks
1484 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1485 thenBc (BcM expr) cont = BcM $ \st0 -> do
1486 (st1, q) <- expr st0
1491 thenBc_ :: BcM a -> BcM b -> BcM b
1492 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1493 (st1, _) <- expr st0
1494 (st2, r) <- cont st1
1497 returnBc :: a -> BcM a
1498 returnBc result = BcM $ \st -> (return (st, result))
1500 instance Monad BcM where
1505 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1507 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1509 recordMallocBc :: Ptr a -> BcM ()
1511 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1513 recordItblMallocBc :: ItblPtr -> BcM ()
1514 recordItblMallocBc a
1515 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1517 getLabelBc :: BcM Int
1519 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1521 getLabelsBc :: Int -> BcM [Int]
1523 = BcM $ \st -> let ctr = nextlabel st
1524 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1526 getBreakArray :: BcM BreakArray
1527 getBreakArray = BcM $ \st -> return (st, breakArray st)
1529 newUnique :: BcM Unique
1531 \st -> case splitUniqSupply (uniqSupply st) of
1532 (us1, us2) -> let newState = st { uniqSupply = us2 }
1533 in return (newState, uniqFromSupply us1)
1535 newId :: Type -> BcM Id
1538 return $ mkSysLocal tickFS uniq ty
1540 tickFS :: FastString
1541 tickFS = fsLit "ticked"