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"
50 import Data.List ( intersperse, sortBy, zip4, zip6, partition )
51 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8,
52 withForeignPtr, castFunPtrToPtr, nullPtr, plusPtr )
54 import Control.Exception ( throwDyn )
56 import GHC.Exts ( Int(..), ByteArray# )
58 import Control.Monad ( when )
59 import Data.Char ( ord, chr )
67 -- -----------------------------------------------------------------------------
68 -- Generating byte code for a complete module
70 byteCodeGen :: DynFlags
74 -> IO CompiledByteCode
75 byteCodeGen dflags binds tycs modBreaks
76 = do showPass dflags "ByteCodeGen"
78 let flatBinds = [ (bndr, freeVars rhs)
79 | (bndr, rhs) <- flattenBinds binds]
81 us <- mkSplitUniqSupply 'y'
82 (BcM_State _us final_ctr mallocd _, proto_bcos)
83 <- runBc us modBreaks (mapM schemeTopBind flatBinds)
85 when (notNull mallocd)
86 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
88 dumpIfSet_dyn dflags Opt_D_dump_BCOs
89 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
91 assembleBCOs proto_bcos tycs
93 -- -----------------------------------------------------------------------------
94 -- Generating byte code for an expression
96 -- Returns: (the root BCO for this expression,
97 -- a list of auxilary BCOs resulting from compiling closures)
98 coreExprToBCOs :: DynFlags
101 coreExprToBCOs dflags expr
102 = do showPass dflags "ByteCodeGen"
104 -- create a totally bogus name for the top-level BCO; this
105 -- should be harmless, since it's never used for anything
106 let invented_name = mkSystemVarName (mkPseudoUniqueE 0) FSLIT("ExprTopLevel")
107 invented_id = Id.mkLocalId invented_name (panic "invented_id's type")
109 -- the uniques are needed to generate fresh variables when we introduce new
110 -- let bindings for ticked expressions
111 us <- mkSplitUniqSupply 'y'
112 (BcM_State _us final_ctr mallocd _ , proto_bco)
113 <- runBc us emptyModBreaks (schemeTopBind (invented_id, freeVars expr))
115 when (notNull mallocd)
116 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
118 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
120 assembleBCO proto_bco
123 -- -----------------------------------------------------------------------------
124 -- Compilation schema for the bytecode generator
126 type BCInstrList = OrdList BCInstr
128 type Sequel = Int -- back off to this depth before ENTER
130 -- Maps Ids to the offset from the stack _base_ so we don't have
131 -- to mess with it after each push/pop.
132 type BCEnv = FiniteMap Id Int -- To find vars on the stack
134 ppBCEnv :: BCEnv -> SDoc
137 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
140 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idCgRep var)
141 cmp_snd x y = compare (snd x) (snd y)
143 -- Create a BCO and do a spot of peephole optimisation on the insns
148 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
152 -> Bool -- True <=> is a return point, rather than a function
155 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap is_ret mallocd_blocks
158 protoBCOInstrs = maybe_with_stack_check,
159 protoBCOBitmap = bitmap,
160 protoBCOBitmapSize = bitmap_size,
161 protoBCOArity = arity,
162 protoBCOExpr = origin,
163 protoBCOPtrs = mallocd_blocks
166 -- Overestimate the stack usage (in words) of this BCO,
167 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
168 -- stack check. (The interpreter always does a stack check
169 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
170 -- BCO anyway, so we only need to add an explicit on in the
171 -- (hopefully rare) cases when the (overestimated) stack use
172 -- exceeds iNTERP_STACK_CHECK_THRESH.
173 maybe_with_stack_check
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 | stack_overest >= iNTERP_STACK_CHECK_THRESH
179 = STKCHECK stack_overest : peep_d
181 = peep_d -- the supposedly common case
183 -- We assume that this sum doesn't wrap
184 stack_overest = sum (map bciStackUse peep_d)
186 -- Merge local pushes
187 peep_d = peep (fromOL instrs_ordlist)
189 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
190 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
191 peep (PUSH_L off1 : PUSH_L off2 : rest)
192 = PUSH_LL off1 (off2-1) : peep rest
198 argBits :: [CgRep] -> [Bool]
201 | isFollowableArg rep = False : argBits args
202 | otherwise = take (cgRepSizeW rep) (repeat True) ++ argBits args
204 -- -----------------------------------------------------------------------------
207 -- Compile code for the right-hand side of a top-level binding
209 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
212 schemeTopBind (id, rhs)
213 | Just data_con <- isDataConWorkId_maybe id,
214 isNullaryRepDataCon data_con = do
215 -- Special case for the worker of a nullary data con.
216 -- It'll look like this: Nil = /\a -> Nil a
217 -- If we feed it into schemeR, we'll get
219 -- because mkConAppCode treats nullary constructor applications
220 -- by just re-using the single top-level definition. So
221 -- for the worker itself, we must allocate it directly.
222 -- ioToBc (putStrLn $ "top level BCO")
223 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
224 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
227 = schemeR [{- No free variables -}] (id, rhs)
230 -- -----------------------------------------------------------------------------
233 -- Compile code for a right-hand side, to give a BCO that,
234 -- when executed with the free variables and arguments on top of the stack,
235 -- will return with a pointer to the result on top of the stack, after
236 -- removing the free variables and arguments.
238 -- Park the resulting BCO in the monad. Also requires the
239 -- variable to which this value was bound, so as to give the
240 -- resulting BCO a name.
242 schemeR :: [Id] -- Free vars of the RHS, ordered as they
243 -- will appear in the thunk. Empty for
244 -- top-level things, which have no free vars.
245 -> (Id, AnnExpr Id VarSet)
246 -> BcM (ProtoBCO Name)
247 schemeR fvs (nm, rhs)
251 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
252 $$ pprCoreExpr (deAnnotate rhs)
258 = schemeR_wrk fvs nm rhs (collect [] rhs)
260 collect :: [Var] -> AnnExpr Id VarSet -> ([Var], AnnExpr' Id VarSet)
261 collect xs (_, AnnNote note e) = collect xs e
262 collect xs (_, AnnCast e _) = collect xs e
263 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
264 collect xs (_, not_lambda) = (reverse xs, not_lambda)
266 schemeR_wrk :: [Id] -> Id -> AnnExpr Id VarSet -> ([Var], AnnExpr' Var VarSet) -> BcM (ProtoBCO Name)
267 schemeR_wrk fvs nm original_body (args, body)
269 all_args = reverse args ++ fvs
270 arity = length all_args
271 -- all_args are the args in reverse order. We're compiling a function
272 -- \fv1..fvn x1..xn -> e
273 -- i.e. the fvs come first
275 szsw_args = map idSizeW all_args
276 szw_args = sum szsw_args
277 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
279 -- make the arg bitmap
280 bits = argBits (reverse (map idCgRep all_args))
281 bitmap_size = length bits
282 bitmap = mkBitmap bits
284 body_code <- schemeER_wrk szw_args p_init body
286 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
287 arity bitmap_size bitmap False{-not alts-})
289 -- introduce break instructions for ticked expressions
290 schemeER_wrk :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
292 | Just (tickInfo, (_annot, newRhs)) <- isTickedExp' rhs = do
293 code <- schemeE d 0 p newRhs
295 let idOffSets = getVarOffSets d p tickInfo
296 let tickNumber = tickInfo_number tickInfo
297 let breakInfo = BreakInfo
298 { breakInfo_module = tickInfo_module tickInfo
299 , breakInfo_number = tickNumber
300 , breakInfo_vars = idOffSets
301 , breakInfo_resty = exprType (deAnnotate' newRhs)
303 let breakInstr = case arr of (BA arr#) -> BRK_FUN arr# tickNumber breakInfo
304 return $ breakInstr `consOL` code
305 | otherwise = schemeE d 0 p rhs
307 getVarOffSets :: Int -> BCEnv -> TickInfo -> [(Id, Int)]
308 getVarOffSets d p = catMaybes . map (getOffSet d p) . tickInfo_locals
310 getOffSet :: Int -> BCEnv -> Id -> Maybe (Id, Int)
312 = case lookupBCEnv_maybe env id of
314 Just offset -> Just (id, d - offset)
316 fvsToEnv :: BCEnv -> VarSet -> [Id]
317 -- Takes the free variables of a right-hand side, and
318 -- delivers an ordered list of the local variables that will
319 -- be captured in the thunk for the RHS
320 -- The BCEnv argument tells which variables are in the local
321 -- environment: these are the ones that should be captured
323 -- The code that constructs the thunk, and the code that executes
324 -- it, have to agree about this layout
325 fvsToEnv p fvs = [v | v <- varSetElems fvs,
326 isId v, -- Could be a type variable
329 -- -----------------------------------------------------------------------------
334 { tickInfo_number :: Int -- the (module) unique number of the tick
335 , tickInfo_module :: Module -- the origin of the ticked expression
336 , tickInfo_locals :: [Id] -- the local vars in scope at the ticked expression
339 instance Outputable TickInfo where
340 ppr info = text "TickInfo" <+>
341 parens (int (tickInfo_number info) <+> ppr (tickInfo_module info) <+>
342 ppr (tickInfo_locals info))
344 -- Compile code to apply the given expression to the remaining args
345 -- on the stack, returning a HNF.
346 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
348 -- Delegate tail-calls to schemeT.
349 schemeE d s p e@(AnnApp f a)
352 schemeE d s p e@(AnnVar v)
353 | not (isUnLiftedType v_type)
354 = -- Lifted-type thing; push it in the normal way
358 = do -- Returning an unlifted value.
359 -- Heave it on the stack, SLIDE, and RETURN.
360 (push, szw) <- pushAtom d p (AnnVar v)
361 return (push -- value onto stack
362 `appOL` mkSLIDE szw (d-s) -- clear to sequel
363 `snocOL` RETURN_UBX v_rep) -- go
366 v_rep = typeCgRep v_type
368 schemeE d s p (AnnLit literal)
369 = do (push, szw) <- pushAtom d p (AnnLit literal)
370 let l_rep = typeCgRep (literalType literal)
371 return (push -- value onto stack
372 `appOL` mkSLIDE szw (d-s) -- clear to sequel
373 `snocOL` RETURN_UBX l_rep) -- go
375 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
376 | (AnnVar v, args_r_to_l) <- splitApp rhs,
377 Just data_con <- isDataConWorkId_maybe v,
378 dataConRepArity data_con == length args_r_to_l
379 = do -- Special case for a non-recursive let whose RHS is a
380 -- saturatred constructor application.
381 -- Just allocate the constructor and carry on
382 alloc_code <- mkConAppCode d s p data_con args_r_to_l
383 body_code <- schemeE (d+1) s (addToFM p x d) body
384 return (alloc_code `appOL` body_code)
386 -- General case for let. Generates correct, if inefficient, code in
388 schemeE d s p (AnnLet binds (_,body))
389 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
390 AnnRec xs_n_rhss -> unzip xs_n_rhss
393 fvss = map (fvsToEnv p' . fst) rhss
395 -- Sizes of free vars
396 sizes = map (\rhs_fvs -> sum (map idSizeW rhs_fvs)) fvss
398 -- the arity of each rhs
399 arities = map (length . fst . collect []) rhss
401 -- This p', d' defn is safe because all the items being pushed
402 -- are ptrs, so all have size 1. d' and p' reflect the stack
403 -- after the closures have been allocated in the heap (but not
404 -- filled in), and pointers to them parked on the stack.
405 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
407 zipE = zipEqual "schemeE"
409 -- ToDo: don't build thunks for things with no free variables
410 build_thunk dd [] size bco off arity
411 = return (PUSH_BCO bco `consOL` unitOL (mkap (off+size) size))
413 mkap | arity == 0 = MKAP
415 build_thunk dd (fv:fvs) size bco off arity = do
416 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
417 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off arity
418 return (push_code `appOL` more_push_code)
420 alloc_code = toOL (zipWith mkAlloc sizes arities)
421 where mkAlloc sz 0 = ALLOC_AP sz
422 mkAlloc sz arity = ALLOC_PAP arity sz
424 compile_bind d' fvs x rhs size arity off = do
425 bco <- schemeR fvs (x,rhs)
426 build_thunk d' fvs size bco off arity
429 [ compile_bind d' fvs x rhs size arity n
430 | (fvs, x, rhs, size, arity, n) <-
431 zip6 fvss xs rhss sizes arities [n_binds, n_binds-1 .. 1]
434 body_code <- schemeE d' s p' body
435 thunk_codes <- sequence compile_binds
436 return (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
438 -- introduce a let binding for a ticked case expression. This rule *should* only fire when the
439 -- expression was not already let-bound (the code gen for let bindings should take care of that).
440 -- Todo: we call exprFreeVars on a deAnnotated expression, this may not be the best way
441 -- to calculate the free vars but it seemed like the least intrusive thing to do
442 schemeE d s p exp@(AnnCase {})
443 | Just (tickInfo, _exp) <- isTickedExp' exp = do
444 let fvs = exprFreeVars $ deAnnotate' exp
445 let ty = exprType $ deAnnotate' exp
447 -- Todo: is emptyVarSet correct on the next line?
448 let letExp = AnnLet (AnnNonRec id (fvs, exp)) (emptyVarSet, AnnVar id)
451 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1, bind2], rhs)])
452 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind1)
454 -- case .... of x { (# VoidArg'd-thing, a #) -> ... }
456 -- case .... of a { DEFAULT -> ... }
457 -- becuse the return convention for both are identical.
459 -- Note that it does not matter losing the void-rep thing from the
460 -- envt (it won't be bound now) because we never look such things up.
462 = --trace "automagic mashing of case alts (# VoidArg, a #)" $
463 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
465 | isUnboxedTupleCon dc, VoidArg <- typeCgRep (idType bind2)
466 = --trace "automagic mashing of case alts (# a, VoidArg #)" $
467 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
469 schemeE d s p (AnnCase scrut bndr _ [(DataAlt dc, [bind1], rhs)])
470 | isUnboxedTupleCon dc
471 -- Similarly, convert
472 -- case .... of x { (# a #) -> ... }
474 -- case .... of a { DEFAULT -> ... }
475 = --trace "automagic mashing of case alts (# a #)" $
476 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
478 schemeE d s p (AnnCase scrut bndr _ alts)
479 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
481 schemeE d s p (AnnNote note (_, body))
484 schemeE d s p (AnnCast (_, body) _)
488 = pprPanic "ByteCodeGen.schemeE: unhandled case"
489 (pprCoreExpr (deAnnotate' other))
495 A ticked expression looks like this:
497 case tick<n> var1 ... varN of DEFAULT -> e
499 (*) <n> is the number of the tick, which is unique within a module
500 (*) var1 ... varN are the local variables in scope at the tick site
502 If we find a ticked expression we return:
504 Just ((n, [var1 ... varN]), e)
506 otherwise we return Nothing.
508 The idea is that the "case tick<n> ..." is really just an annotation on
509 the code. When we find such a thing, we pull out the useful information,
510 and then compile the code as if it was just the expression "e".
514 isTickedExp :: AnnExpr Id a -> Maybe (TickInfo, AnnExpr Id a)
515 isTickedExp (annot, expr) = isTickedExp' expr
517 isTickedExp' :: AnnExpr' Id a -> Maybe (TickInfo, AnnExpr Id a)
518 isTickedExp' (AnnCase scrut _bndr _type alts)
519 | Just tickInfo <- isTickedScrut scrut,
520 [(DEFAULT, _bndr, rhs)] <- alts
521 = Just (tickInfo, rhs)
523 isTickedScrut :: (AnnExpr Id a) -> Maybe TickInfo
526 Just (TickBox modName tickNumber) <- isTickBoxOp_maybe id
527 = Just $ TickInfo { tickInfo_number = tickNumber
528 , tickInfo_module = modName
529 , tickInfo_locals = idsOfArgs args
531 | otherwise = Nothing
533 (f, args) = collectArgs $ deAnnotate expr
534 idsOfArgs :: [Expr Id] -> [Id]
535 idsOfArgs = catMaybes . map exprId
536 exprId :: Expr Id -> Maybe Id
537 exprId (Var id) = Just id
538 exprId other = Nothing
540 isTickedExp' other = Nothing
542 -- Compile code to do a tail call. Specifically, push the fn,
543 -- slide the on-stack app back down to the sequel depth,
544 -- and enter. Four cases:
547 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
548 -- The int will be on the stack. Generate a code sequence
549 -- to convert it to the relevant constructor, SLIDE and ENTER.
551 -- 1. The fn denotes a ccall. Defer to generateCCall.
553 -- 2. (Another nasty hack). Spot (# a::VoidArg, b #) and treat
554 -- it simply as b -- since the representations are identical
555 -- (the VoidArg takes up zero stack space). Also, spot
556 -- (# b #) and treat it as b.
558 -- 3. Application of a constructor, by defn saturated.
559 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
560 -- then the ptrs, and then do PACK and RETURN.
562 -- 4. Otherwise, it must be a function call. Push the args
563 -- right to left, SLIDE and ENTER.
565 schemeT :: Int -- Stack depth
566 -> Sequel -- Sequel depth
567 -> BCEnv -- stack env
568 -> AnnExpr' Id VarSet
573 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
574 -- = panic "schemeT ?!?!"
576 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
580 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
581 = do (push, arg_words) <- pushAtom d p arg
582 tagToId_sequence <- implement_tagToId constr_names
583 return (push `appOL` tagToId_sequence
584 `appOL` mkSLIDE 1 (d+arg_words-s)
588 | Just (CCall ccall_spec) <- isFCallId_maybe fn
589 = generateCCall d s p ccall_spec fn args_r_to_l
591 -- Case 2: Constructor application
592 | Just con <- maybe_saturated_dcon,
593 isUnboxedTupleCon con
594 = case args_r_to_l of
595 [arg1,arg2] | isVoidArgAtom arg1 ->
596 unboxedTupleReturn d s p arg2
597 [arg1,arg2] | isVoidArgAtom arg2 ->
598 unboxedTupleReturn d s p arg1
599 _other -> unboxedTupleException
601 -- Case 3: Ordinary data constructor
602 | Just con <- maybe_saturated_dcon
603 = do alloc_con <- mkConAppCode d s p con args_r_to_l
604 return (alloc_con `appOL`
605 mkSLIDE 1 (d - s) `snocOL`
608 -- Case 4: Tail call of function
610 = doTailCall d s p fn args_r_to_l
613 -- Detect and extract relevant info for the tagToEnum kludge.
614 maybe_is_tagToEnum_call
615 = let extract_constr_Names ty
616 | Just (tyc, []) <- splitTyConApp_maybe (repType ty),
618 = map (getName . dataConWorkId) (tyConDataCons tyc)
619 -- NOTE: use the worker name, not the source name of
620 -- the DataCon. See DataCon.lhs for details.
622 = panic "maybe_is_tagToEnum_call.extract_constr_Ids"
625 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
626 -> case isPrimOpId_maybe v of
627 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
631 -- Extract the args (R->L) and fn
632 -- The function will necessarily be a variable,
633 -- because we are compiling a tail call
634 (AnnVar fn, args_r_to_l) = splitApp app
636 -- Only consider this to be a constructor application iff it is
637 -- saturated. Otherwise, we'll call the constructor wrapper.
638 n_args = length args_r_to_l
640 = case isDataConWorkId_maybe fn of
641 Just con | dataConRepArity con == n_args -> Just con
644 -- -----------------------------------------------------------------------------
645 -- Generate code to build a constructor application,
646 -- leaving it on top of the stack
648 mkConAppCode :: Int -> Sequel -> BCEnv
649 -> DataCon -- The data constructor
650 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
653 mkConAppCode orig_d s p con [] -- Nullary constructor
654 = ASSERT( isNullaryRepDataCon con )
655 return (unitOL (PUSH_G (getName (dataConWorkId con))))
656 -- Instead of doing a PACK, which would allocate a fresh
657 -- copy of this constructor, use the single shared version.
659 mkConAppCode orig_d s p con args_r_to_l
660 = ASSERT( dataConRepArity con == length args_r_to_l )
661 do_pushery orig_d (non_ptr_args ++ ptr_args)
663 -- The args are already in reverse order, which is the way PACK
664 -- expects them to be. We must push the non-ptrs after the ptrs.
665 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
667 do_pushery d (arg:args)
668 = do (push, arg_words) <- pushAtom d p arg
669 more_push_code <- do_pushery (d+arg_words) args
670 return (push `appOL` more_push_code)
672 = return (unitOL (PACK con n_arg_words))
674 n_arg_words = d - orig_d
677 -- -----------------------------------------------------------------------------
678 -- Returning an unboxed tuple with one non-void component (the only
679 -- case we can handle).
681 -- Remember, we don't want to *evaluate* the component that is being
682 -- returned, even if it is a pointed type. We always just return.
685 :: Int -> Sequel -> BCEnv
686 -> AnnExpr' Id VarSet -> BcM BCInstrList
687 unboxedTupleReturn d s p arg = do
688 (push, sz) <- pushAtom d p arg
690 mkSLIDE sz (d-s) `snocOL`
691 RETURN_UBX (atomRep arg))
693 -- -----------------------------------------------------------------------------
694 -- Generate code for a tail-call
697 :: Int -> Sequel -> BCEnv
698 -> Id -> [AnnExpr' Id VarSet]
700 doTailCall init_d s p fn args
701 = do_pushes init_d args (map atomRep args)
703 do_pushes d [] reps = do
704 ASSERT( null reps ) return ()
705 (push_fn, sz) <- pushAtom d p (AnnVar fn)
706 ASSERT( sz == 1 ) return ()
707 return (push_fn `appOL` (
708 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
710 do_pushes d args reps = do
711 let (push_apply, n, rest_of_reps) = findPushSeq reps
712 (these_args, rest_of_args) = splitAt n args
713 (next_d, push_code) <- push_seq d these_args
714 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
715 -- ^^^ for the PUSH_APPLY_ instruction
716 return (push_code `appOL` (push_apply `consOL` instrs))
718 push_seq d [] = return (d, nilOL)
719 push_seq d (arg:args) = do
720 (push_code, sz) <- pushAtom d p arg
721 (final_d, more_push_code) <- push_seq (d+sz) args
722 return (final_d, push_code `appOL` more_push_code)
724 -- v. similar to CgStackery.findMatch, ToDo: merge
725 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
726 = (PUSH_APPLY_PPPPPP, 6, rest)
727 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: PtrArg: rest)
728 = (PUSH_APPLY_PPPPP, 5, rest)
729 findPushSeq (PtrArg: PtrArg: PtrArg: PtrArg: rest)
730 = (PUSH_APPLY_PPPP, 4, rest)
731 findPushSeq (PtrArg: PtrArg: PtrArg: rest)
732 = (PUSH_APPLY_PPP, 3, rest)
733 findPushSeq (PtrArg: PtrArg: rest)
734 = (PUSH_APPLY_PP, 2, rest)
735 findPushSeq (PtrArg: rest)
736 = (PUSH_APPLY_P, 1, rest)
737 findPushSeq (VoidArg: rest)
738 = (PUSH_APPLY_V, 1, rest)
739 findPushSeq (NonPtrArg: rest)
740 = (PUSH_APPLY_N, 1, rest)
741 findPushSeq (FloatArg: rest)
742 = (PUSH_APPLY_F, 1, rest)
743 findPushSeq (DoubleArg: rest)
744 = (PUSH_APPLY_D, 1, rest)
745 findPushSeq (LongArg: rest)
746 = (PUSH_APPLY_L, 1, rest)
748 = panic "ByteCodeGen.findPushSeq"
750 -- -----------------------------------------------------------------------------
753 doCase :: Int -> Sequel -> BCEnv
754 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
755 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
757 doCase d s p (_,scrut) bndr alts is_unboxed_tuple
759 -- Top of stack is the return itbl, as usual.
760 -- underneath it is the pointer to the alt_code BCO.
761 -- When an alt is entered, it assumes the returned value is
762 -- on top of the itbl.
765 -- An unlifted value gets an extra info table pushed on top
766 -- when it is returned.
767 unlifted_itbl_sizeW | isAlgCase = 0
770 -- depth of stack after the return value has been pushed
771 d_bndr = d + ret_frame_sizeW + idSizeW bndr
773 -- depth of stack after the extra info table for an unboxed return
774 -- has been pushed, if any. This is the stack depth at the
776 d_alts = d_bndr + unlifted_itbl_sizeW
778 -- Env in which to compile the alts, not including
779 -- any vars bound by the alts themselves
780 p_alts = addToFM p bndr (d_bndr - 1)
782 bndr_ty = idType bndr
783 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
785 -- given an alt, return a discr and code for it.
786 codeAlt alt@(DEFAULT, _, (_,rhs))
787 = do rhs_code <- schemeE d_alts s p_alts rhs
788 return (NoDiscr, rhs_code)
790 codeAlt alt@(discr, bndrs, (_,rhs))
791 -- primitive or nullary constructor alt: no need to UNPACK
792 | null real_bndrs = do
793 rhs_code <- schemeE d_alts s p_alts rhs
794 return (my_discr alt, rhs_code)
795 -- algebraic alt with some binders
796 | ASSERT(isAlgCase) otherwise =
798 (ptrs,nptrs) = partition (isFollowableArg.idCgRep) real_bndrs
799 ptr_sizes = map idSizeW ptrs
800 nptrs_sizes = map idSizeW nptrs
801 bind_sizes = ptr_sizes ++ nptrs_sizes
802 size = sum ptr_sizes + sum nptrs_sizes
803 -- the UNPACK instruction unpacks in reverse order...
804 p' = addListToFM p_alts
805 (zip (reverse (ptrs ++ nptrs))
806 (mkStackOffsets d_alts (reverse bind_sizes)))
808 rhs_code <- schemeE (d_alts+size) s p' rhs
809 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
811 real_bndrs = filter (not.isTyVar) bndrs
813 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
814 my_discr (DataAlt dc, binds, rhs)
815 | isUnboxedTupleCon dc
816 = unboxedTupleException
818 = DiscrP (dataConTag dc - fIRST_TAG)
819 my_discr (LitAlt l, binds, rhs)
820 = case l of MachInt i -> DiscrI (fromInteger i)
821 MachFloat r -> DiscrF (fromRational r)
822 MachDouble r -> DiscrD (fromRational r)
823 MachChar i -> DiscrI (ord i)
824 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
827 | not isAlgCase = Nothing
829 = case [dc | (DataAlt dc, _, _) <- alts] of
831 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
833 -- the bitmap is relative to stack depth d, i.e. before the
834 -- BCO, info table and return value are pushed on.
835 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
836 -- except that here we build the bitmap from the known bindings of
837 -- things that are pointers, whereas in CgBindery the code builds the
838 -- bitmap from the free slots and unboxed bindings.
841 -- NOTE [7/12/2006] bug #1013, testcase ghci/should_run/ghci002.
842 -- The bitmap must cover the portion of the stack up to the sequel only.
843 -- Previously we were building a bitmap for the whole depth (d), but we
844 -- really want a bitmap up to depth (d-s). This affects compilation of
845 -- case-of-case expressions, which is the only time we can be compiling a
846 -- case expression with s /= 0.
848 bitmap = intsToReverseBitmap bitmap_size{-size-}
849 (sortLe (<=) (filter (< bitmap_size) rel_slots))
852 rel_slots = concat (map spread binds)
854 | isFollowableArg (idCgRep id) = [ rel_offset ]
856 where rel_offset = d - offset - 1
859 alt_stuff <- mapM codeAlt alts
860 alt_final <- mkMultiBranch maybe_ncons alt_stuff
863 alt_bco_name = getName bndr
864 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
865 0{-no arity-} bitmap_size bitmap True{-is alts-}
867 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
868 -- "\n bitmap = " ++ show bitmap) $ do
869 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
870 alt_bco' <- emitBc alt_bco
872 | isAlgCase = PUSH_ALTS alt_bco'
873 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typeCgRep bndr_ty)
874 return (push_alts `consOL` scrut_code)
877 -- -----------------------------------------------------------------------------
878 -- Deal with a CCall.
880 -- Taggedly push the args onto the stack R->L,
881 -- deferencing ForeignObj#s and adjusting addrs to point to
882 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
883 -- (machine) code for the ccall, and create bytecodes to call that and
884 -- then return in the right way.
886 generateCCall :: Int -> Sequel -- stack and sequel depths
888 -> CCallSpec -- where to call
889 -> Id -- of target, for type info
890 -> [AnnExpr' Id VarSet] -- args (atoms)
893 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
896 addr_sizeW = cgRepSizeW NonPtrArg
898 -- Get the args on the stack, with tags and suitably
899 -- dereferenced for the CCall. For each arg, return the
900 -- depth to the first word of the bits for that arg, and the
901 -- CgRep of what was actually pushed.
903 pargs d [] = return []
905 = let arg_ty = repType (exprType (deAnnotate' a))
907 in case splitTyConApp_maybe arg_ty of
908 -- Don't push the FO; instead push the Addr# it
911 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
912 -> do rest <- pargs (d + addr_sizeW) az
913 code <- parg_ArrayishRep arrPtrsHdrSize d p a
914 return ((code,NonPtrArg):rest)
916 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
917 -> do rest <- pargs (d + addr_sizeW) az
918 code <- parg_ArrayishRep arrWordsHdrSize d p a
919 return ((code,NonPtrArg):rest)
921 -- Default case: push taggedly, but otherwise intact.
923 -> do (code_a, sz_a) <- pushAtom d p a
924 rest <- pargs (d+sz_a) az
925 return ((code_a, atomRep a) : rest)
927 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
928 -- the stack but then advance it over the headers, so as to
929 -- point to the payload.
930 parg_ArrayishRep hdrSize d p a
931 = do (push_fo, _) <- pushAtom d p a
932 -- The ptr points at the header. Advance it over the
933 -- header and then pretend this is an Addr#.
934 return (push_fo `snocOL` SWIZZLE 0 hdrSize)
937 code_n_reps <- pargs d0 args_r_to_l
939 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
941 push_args = concatOL pushs_arg
942 d_after_args = d0 + sum (map cgRepSizeW a_reps_pushed_r_to_l)
944 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidArg
945 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
947 = reverse (tail a_reps_pushed_r_to_l)
949 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
950 -- push_args is the code to do that.
951 -- d_after_args is the stack depth once the args are on.
953 -- Get the result rep.
954 (returns_void, r_rep)
955 = case maybe_getCCallReturnRep (idType fn) of
956 Nothing -> (True, VoidArg)
957 Just rr -> (False, rr)
959 Because the Haskell stack grows down, the a_reps refer to
960 lowest to highest addresses in that order. The args for the call
961 are on the stack. Now push an unboxed Addr# indicating
962 the C function to call. Then push a dummy placeholder for the
963 result. Finally, emit a CCALL insn with an offset pointing to the
964 Addr# just pushed, and a literal field holding the mallocville
965 address of the piece of marshalling code we generate.
966 So, just prior to the CCALL insn, the stack looks like this
967 (growing down, as usual):
972 Addr# address_of_C_fn
973 <placeholder-for-result#> (must be an unboxed type)
975 The interpreter then calls the marshall code mentioned
976 in the CCALL insn, passing it (& <placeholder-for-result#>),
977 that is, the addr of the topmost word in the stack.
978 When this returns, the placeholder will have been
979 filled in. The placeholder is slid down to the sequel
980 depth, and we RETURN.
982 This arrangement makes it simple to do f-i-dynamic since the Addr#
983 value is the first arg anyway.
985 The marshalling code is generated specifically for this
986 call site, and so knows exactly the (Haskell) stack
987 offsets of the args, fn address and placeholder. It
988 copies the args to the C stack, calls the stacked addr,
989 and parks the result back in the placeholder. The interpreter
990 calls it as a normal C call, assuming it has a signature
991 void marshall_code ( StgWord* ptr_to_top_of_stack )
993 -- resolve static address
997 -> return (False, panic "ByteCodeGen.generateCCall(dyn)")
999 -> do res <- ioToBc (lookupStaticPtr target)
1002 (is_static, static_target_addr) <- get_target_info
1005 -- Get the arg reps, zapping the leading Addr# in the dynamic case
1006 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
1007 | is_static = a_reps_pushed_RAW
1008 | otherwise = if null a_reps_pushed_RAW
1009 then panic "ByteCodeGen.generateCCall: dyn with no args"
1010 else tail a_reps_pushed_RAW
1013 (push_Addr, d_after_Addr)
1015 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
1016 d_after_args + addr_sizeW)
1017 | otherwise -- is already on the stack
1018 = (nilOL, d_after_args)
1020 -- Push the return placeholder. For a call returning nothing,
1021 -- this is a VoidArg (tag).
1022 r_sizeW = cgRepSizeW r_rep
1023 d_after_r = d_after_Addr + r_sizeW
1024 r_lit = mkDummyLiteral r_rep
1025 push_r = (if returns_void
1027 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
1029 -- generate the marshalling code we're going to call
1032 arg1_offW = r_sizeW + addr_sizeW
1033 args_offW = map (arg1_offW +)
1034 (init (scanl (+) 0 (map cgRepSizeW a_reps)))
1036 addr_of_marshaller <- ioToBc (mkMarshalCode cconv
1037 (r_offW, r_rep) addr_offW
1038 (zip args_offW a_reps))
1039 recordItblMallocBc (ItblPtr (castFunPtrToPtr addr_of_marshaller))
1041 -- Offset of the next stack frame down the stack. The CCALL
1042 -- instruction needs to describe the chunk of stack containing
1043 -- the ccall args to the GC, so it needs to know how large it
1044 -- is. See comment in Interpreter.c with the CCALL instruction.
1045 stk_offset = d_after_r - s
1048 do_call = unitOL (CCALL stk_offset (castFunPtrToPtr addr_of_marshaller))
1050 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
1051 `snocOL` RETURN_UBX r_rep
1053 --trace (show (arg1_offW, args_offW , (map cgRepSizeW a_reps) )) $
1056 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
1060 -- Make a dummy literal, to be used as a placeholder for FFI return
1061 -- values on the stack.
1062 mkDummyLiteral :: CgRep -> Literal
1065 NonPtrArg -> MachWord 0
1066 DoubleArg -> MachDouble 0
1067 FloatArg -> MachFloat 0
1068 LongArg -> MachWord64 0
1069 _ -> moan64 "mkDummyLiteral" (ppr pr)
1073 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1074 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
1077 -- and check that an unboxed pair is returned wherein the first arg is VoidArg'd.
1079 -- Alternatively, for call-targets returning nothing, convert
1081 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1082 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1086 maybe_getCCallReturnRep :: Type -> Maybe CgRep
1087 maybe_getCCallReturnRep fn_ty
1088 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1090 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1092 = case splitTyConApp_maybe (repType r_ty) of
1093 (Just (tyc, tys)) -> (tyc, map typeCgRep tys)
1095 ok = ( ( r_reps `lengthIs` 2 && VoidArg == head r_reps)
1096 || r_reps == [VoidArg] )
1097 && isUnboxedTupleTyCon r_tycon
1098 && case maybe_r_rep_to_go of
1100 Just r_rep -> r_rep /= PtrArg
1101 -- if it was, it would be impossible
1102 -- to create a valid return value
1103 -- placeholder on the stack
1104 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1107 --trace (showSDoc (ppr (a_reps, r_reps))) $
1108 if ok then maybe_r_rep_to_go else blargh
1110 -- Compile code which expects an unboxed Int on the top of stack,
1111 -- (call it i), and pushes the i'th closure in the supplied list
1112 -- as a consequence.
1113 implement_tagToId :: [Name] -> BcM BCInstrList
1114 implement_tagToId names
1115 = ASSERT( notNull names )
1116 do labels <- getLabelsBc (length names)
1117 label_fail <- getLabelBc
1118 label_exit <- getLabelBc
1119 let infos = zip4 labels (tail labels ++ [label_fail])
1121 steps = map (mkStep label_exit) infos
1122 return (concatOL steps
1124 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1126 mkStep l_exit (my_label, next_label, n, name_for_n)
1127 = toOL [LABEL my_label,
1128 TESTEQ_I n next_label,
1133 -- -----------------------------------------------------------------------------
1136 -- Push an atom onto the stack, returning suitable code & number of
1137 -- stack words used.
1139 -- The env p must map each variable to the highest- numbered stack
1140 -- slot for it. For example, if the stack has depth 4 and we
1141 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1142 -- the tag in stack[5], the stack will have depth 6, and p must map v
1143 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1144 -- depth 6 stack has valid words 0 .. 5.
1146 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1148 pushAtom d p (AnnApp f (_, AnnType _))
1149 = pushAtom d p (snd f)
1151 pushAtom d p (AnnNote note e)
1152 = pushAtom d p (snd e)
1154 pushAtom d p (AnnLam x e)
1156 = pushAtom d p (snd e)
1158 pushAtom d p (AnnVar v)
1160 | idCgRep v == VoidArg
1164 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1166 | Just primop <- isPrimOpId_maybe v
1167 = return (unitOL (PUSH_PRIMOP primop), 1)
1169 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1170 = return (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1171 -- d - d_v the number of words between the TOS
1172 -- and the 1st slot of the object
1174 -- d - d_v - 1 the offset from the TOS of the 1st slot
1176 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1179 -- Having found the last slot, we proceed to copy the right number of
1180 -- slots on to the top of the stack.
1182 | otherwise -- v must be a global variable
1184 return (unitOL (PUSH_G (getName v)), sz)
1190 pushAtom d p (AnnLit lit)
1192 MachLabel fs _ -> code NonPtrArg
1193 MachWord w -> code NonPtrArg
1194 MachInt i -> code PtrArg
1195 MachFloat r -> code FloatArg
1196 MachDouble r -> code DoubleArg
1197 MachChar c -> code NonPtrArg
1198 MachStr s -> pushStr s
1201 = let size_host_words = cgRepSizeW rep
1202 in return (unitOL (PUSH_UBX (Left lit) size_host_words),
1206 = let getMallocvilleAddr
1208 FastString _ n _ fp _ ->
1209 -- we could grab the Ptr from the ForeignPtr,
1210 -- but then we have no way to control its lifetime.
1211 -- In reality it'll probably stay alive long enoungh
1212 -- by virtue of the global FastString table, but
1213 -- to be on the safe side we copy the string into
1214 -- a malloc'd area of memory.
1215 do ptr <- ioToBc (mallocBytes (n+1))
1218 withForeignPtr fp $ \p -> do
1219 memcpy ptr p (fromIntegral n)
1220 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1224 addr <- getMallocvilleAddr
1225 -- Get the addr on the stack, untaggedly
1226 return (unitOL (PUSH_UBX (Right addr) 1), 1)
1228 pushAtom d p (AnnCast e _)
1229 = pushAtom d p (snd e)
1232 = pprPanic "ByteCodeGen.pushAtom"
1233 (pprCoreExpr (deAnnotate (undefined, other)))
1235 foreign import ccall unsafe "memcpy"
1236 memcpy :: Ptr a -> Ptr b -> CSize -> IO ()
1239 -- -----------------------------------------------------------------------------
1240 -- Given a bunch of alts code and their discrs, do the donkey work
1241 -- of making a multiway branch using a switch tree.
1242 -- What a load of hassle!
1244 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1245 -- a hint; generates better code
1246 -- Nothing is always safe
1247 -> [(Discr, BCInstrList)]
1249 mkMultiBranch maybe_ncons raw_ways
1250 = let d_way = filter (isNoDiscr.fst) raw_ways
1252 (\w1 w2 -> leAlt (fst w1) (fst w2))
1253 (filter (not.isNoDiscr.fst) raw_ways)
1255 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1256 mkTree [] range_lo range_hi = return the_default
1258 mkTree [val] range_lo range_hi
1259 | range_lo `eqAlt` range_hi
1262 = do label_neq <- getLabelBc
1263 return (mkTestEQ (fst val) label_neq
1265 `appOL` unitOL (LABEL label_neq)
1266 `appOL` the_default))
1268 mkTree vals range_lo range_hi
1269 = let n = length vals `div` 2
1270 vals_lo = take n vals
1271 vals_hi = drop n vals
1272 v_mid = fst (head vals_hi)
1274 label_geq <- getLabelBc
1275 code_lo <- mkTree vals_lo range_lo (dec v_mid)
1276 code_hi <- mkTree vals_hi v_mid range_hi
1277 return (mkTestLT v_mid label_geq
1279 `appOL` unitOL (LABEL label_geq)
1283 = case d_way of [] -> unitOL CASEFAIL
1286 -- None of these will be needed if there are no non-default alts
1287 (mkTestLT, mkTestEQ, init_lo, init_hi)
1289 = panic "mkMultiBranch: awesome foursome"
1291 = case fst (head notd_ways) of {
1292 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1293 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1296 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1297 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1300 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1301 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1304 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1305 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1307 DiscrP algMaxBound )
1310 (algMinBound, algMaxBound)
1311 = case maybe_ncons of
1312 Just n -> (0, n - 1)
1313 Nothing -> (minBound, maxBound)
1315 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1316 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1317 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1318 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1319 NoDiscr `eqAlt` NoDiscr = True
1322 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1323 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1324 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1325 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1326 NoDiscr `leAlt` NoDiscr = True
1329 isNoDiscr NoDiscr = True
1332 dec (DiscrI i) = DiscrI (i-1)
1333 dec (DiscrP i) = DiscrP (i-1)
1334 dec other = other -- not really right, but if you
1335 -- do cases on floating values, you'll get what you deserve
1337 -- same snotty comment applies to the following
1339 minD, maxD :: Double
1345 mkTree notd_ways init_lo init_hi
1348 -- -----------------------------------------------------------------------------
1349 -- Supporting junk for the compilation schemes
1351 -- Describes case alts
1359 instance Outputable Discr where
1360 ppr (DiscrI i) = int i
1361 ppr (DiscrF f) = text (show f)
1362 ppr (DiscrD d) = text (show d)
1363 ppr (DiscrP i) = int i
1364 ppr NoDiscr = text "DEF"
1367 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1368 lookupBCEnv_maybe = lookupFM
1370 idSizeW :: Id -> Int
1371 idSizeW id = cgRepSizeW (typeCgRep (idType id))
1373 unboxedTupleException :: a
1374 unboxedTupleException
1377 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1378 "\tto foreign import/export decls in source. Workaround:\n" ++
1379 "\tcompile this module to a .o file, then restart session."))
1382 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1385 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1386 -- The arguments are returned in *right-to-left* order
1387 splitApp (AnnApp (_,f) (_,a))
1388 | isTypeAtom a = splitApp f
1389 | otherwise = case splitApp f of
1390 (f', as) -> (f', a:as)
1391 splitApp (AnnNote n (_,e)) = splitApp e
1392 splitApp (AnnCast (_,e) _) = splitApp e
1393 splitApp e = (e, [])
1396 isTypeAtom :: AnnExpr' id ann -> Bool
1397 isTypeAtom (AnnType _) = True
1398 isTypeAtom _ = False
1400 isVoidArgAtom :: AnnExpr' id ann -> Bool
1401 isVoidArgAtom (AnnVar v) = typeCgRep (idType v) == VoidArg
1402 isVoidArgAtom (AnnNote n (_,e)) = isVoidArgAtom e
1403 isVoidArgAtom (AnnCast (_,e) _) = isVoidArgAtom e
1404 isVoidArgAtom _ = False
1406 atomRep :: AnnExpr' Id ann -> CgRep
1407 atomRep (AnnVar v) = typeCgRep (idType v)
1408 atomRep (AnnLit l) = typeCgRep (literalType l)
1409 atomRep (AnnNote n b) = atomRep (snd b)
1410 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1411 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1412 atomRep (AnnCast b _) = atomRep (snd b)
1413 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1415 isPtrAtom :: AnnExpr' Id ann -> Bool
1416 isPtrAtom e = atomRep e == PtrArg
1418 -- Let szsw be the sizes in words of some items pushed onto the stack,
1419 -- which has initial depth d'. Return the values which the stack environment
1420 -- should map these items to.
1421 mkStackOffsets :: Int -> [Int] -> [Int]
1422 mkStackOffsets original_depth szsw
1423 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1425 -- -----------------------------------------------------------------------------
1426 -- The bytecode generator's monad
1428 type BcPtr = Either ItblPtr (Ptr ())
1432 uniqSupply :: UniqSupply, -- for generating fresh variable names
1433 nextlabel :: Int, -- for generating local labels
1434 malloced :: [BcPtr], -- thunks malloced for current BCO
1435 -- Should be free()d when it is GCd
1436 breakArray :: BreakArray -- array of breakpoint flags
1439 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1441 ioToBc :: IO a -> BcM a
1442 ioToBc io = BcM $ \st -> do
1446 runBc :: UniqSupply -> ModBreaks -> BcM r -> IO (BcM_State, r)
1447 runBc us modBreaks (BcM m)
1448 = m (BcM_State us 0 [] breakArray)
1450 breakArray = modBreaks_flags modBreaks
1452 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1453 thenBc (BcM expr) cont = BcM $ \st0 -> do
1454 (st1, q) <- expr st0
1459 thenBc_ :: BcM a -> BcM b -> BcM b
1460 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1461 (st1, q) <- expr st0
1462 (st2, r) <- cont st1
1465 returnBc :: a -> BcM a
1466 returnBc result = BcM $ \st -> (return (st, result))
1468 instance Monad BcM where
1473 emitBc :: ([BcPtr] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1475 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1477 recordMallocBc :: Ptr a -> BcM ()
1479 = BcM $ \st -> return (st{malloced = Right (castPtr a) : malloced st}, ())
1481 recordItblMallocBc :: ItblPtr -> BcM ()
1482 recordItblMallocBc a
1483 = BcM $ \st -> return (st{malloced = Left a : malloced st}, ())
1485 getLabelBc :: BcM Int
1487 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1489 getLabelsBc :: Int -> BcM [Int]
1491 = BcM $ \st -> let ctr = nextlabel st
1492 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])
1494 getBreakArray :: BcM BreakArray
1495 getBreakArray = BcM $ \st -> return (st, breakArray st)
1497 newUnique :: BcM Unique
1499 \st -> case splitUniqSupply (uniqSupply st) of
1500 (us1, us2) -> let newState = st { uniqSupply = us2 }
1501 in return (newState, uniqFromSupply us1)
1503 newId :: Type -> BcM Id
1506 return $ mkSysLocal FSLIT("ticked") uniq ty