2 % (c) The University of Glasgow 2002
4 \section[ByteCodeGen]{Generate bytecode from Core}
7 module ByteCodeGen ( UnlinkedBCO, byteCodeGen, coreExprToBCOs ) where
9 #include "HsVersions.h"
12 import ByteCodeFFI ( mkMarshalCode, moan64 )
13 import ByteCodeAsm ( CompiledByteCode(..), UnlinkedBCO,
14 assembleBCO, assembleBCOs, iNTERP_STACK_CHECK_THRESH )
15 import ByteCodeLink ( lookupStaticPtr )
18 import Name ( Name, getName, mkSystemName )
21 import ForeignCall ( ForeignCall(..), CCallTarget(..), CCallSpec(..) )
22 import HscTypes ( ModGuts(..), ModGuts,
23 TypeEnv, typeEnvTyCons, typeEnvClasses )
24 import CoreUtils ( exprType )
26 import PprCore ( pprCoreExpr )
27 import Literal ( Literal(..), literalPrimRep )
29 import PrimOp ( PrimOp(..) )
30 import CoreFVs ( freeVars )
31 import Type ( typePrimRep, isUnLiftedType, splitTyConApp_maybe,
33 import DataCon ( DataCon, dataConTag, fIRST_TAG, dataConTyCon,
34 isUnboxedTupleCon, isNullaryDataCon, dataConWorkId,
36 import TyCon ( tyConFamilySize, isDataTyCon, tyConDataCons,
37 isFunTyCon, isUnboxedTupleTyCon )
38 import Class ( Class, classTyCon )
39 import Type ( Type, repType, splitFunTys, dropForAlls )
41 import DataCon ( dataConRepArity )
42 import Var ( isTyVar )
43 import VarSet ( VarSet, varSetElems )
44 import TysPrim ( foreignObjPrimTyCon,
45 arrayPrimTyCon, mutableArrayPrimTyCon,
46 byteArrayPrimTyCon, mutableByteArrayPrimTyCon
48 import PrimRep ( isFollowableRep )
49 import CmdLineOpts ( DynFlags, DynFlag(..) )
50 import ErrUtils ( showPass, dumpIfSet_dyn )
51 import Unique ( mkPseudoUnique3 )
52 import FastString ( FastString(..), unpackFS )
53 import Panic ( GhcException(..) )
54 import PprType ( pprType )
55 import SMRep ( arrWordsHdrSize, arrPtrsHdrSize )
57 import Constants ( wORD_SIZE )
58 import BasicTypes ( TopLevelFlag(..), isTopLevel, isNotTopLevel )
60 import Data.List ( intersperse, sortBy, zip4, zip5, partition )
61 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8 )
62 import Foreign.C ( CInt )
63 import Control.Exception ( throwDyn )
65 import GHC.Exts ( Int(..), ByteArray# )
67 import Control.Monad ( when, mapAndUnzipM )
68 import Data.Char ( ord )
71 -- -----------------------------------------------------------------------------
72 -- Generating byte code for a complete module
74 byteCodeGen :: DynFlags
77 -> IO CompiledByteCode
78 byteCodeGen dflags binds type_env
79 = do showPass dflags "ByteCodeGen"
80 let local_tycons = typeEnvTyCons type_env
81 local_classes = typeEnvClasses type_env
82 tycs = local_tycons ++ map classTyCon local_classes
84 let flatBinds = [ (bndr, freeVars rhs)
85 | (bndr, rhs) <- flattenBinds binds]
87 (BcM_State final_ctr mallocd, proto_bcos)
88 <- runBc (mapM schemeTopBind flatBinds)
90 when (notNull mallocd)
91 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
93 dumpIfSet_dyn dflags Opt_D_dump_BCOs
94 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
96 assembleBCOs proto_bcos tycs
98 -- -----------------------------------------------------------------------------
99 -- Generating byte code for an expression
101 -- Returns: (the root BCO for this expression,
102 -- a list of auxilary BCOs resulting from compiling closures)
103 coreExprToBCOs :: DynFlags
106 coreExprToBCOs dflags expr
107 = do showPass dflags "ByteCodeGen"
109 -- create a totally bogus name for the top-level BCO; this
110 -- should be harmless, since it's never used for anything
111 let invented_name = mkSystemName (mkPseudoUnique3 0) FSLIT("ExprTopLevel")
112 invented_id = mkLocalId invented_name (panic "invented_id's type")
114 (BcM_State final_ctr mallocd, proto_bco)
115 <- runBc (schemeTopBind (invented_id, freeVars expr))
117 when (notNull mallocd)
118 (panic "ByteCodeGen.coreExprToBCOs: missing final emitBc?")
120 dumpIfSet_dyn dflags Opt_D_dump_BCOs "Proto-BCOs" (ppr proto_bco)
122 assembleBCO proto_bco
125 -- -----------------------------------------------------------------------------
126 -- Compilation schema for the bytecode generator
128 type BCInstrList = OrdList BCInstr
130 type Sequel = Int -- back off to this depth before ENTER
132 -- Maps Ids to the offset from the stack _base_ so we don't have
133 -- to mess with it after each push/pop.
134 type BCEnv = FiniteMap Id Int -- To find vars on the stack
136 ppBCEnv :: BCEnv -> SDoc
139 $$ nest 4 (vcat (map pp_one (sortBy cmp_snd (fmToList p))))
142 pp_one (var, offset) = int offset <> colon <+> ppr var <+> ppr (idPrimRep var)
143 cmp_snd x y = compare (snd x) (snd y)
145 -- Create a BCO and do a spot of peephole optimisation on the insns
150 -> Either [AnnAlt Id VarSet] (AnnExpr Id VarSet)
154 -> Bool -- True <=> is a return point, rather than a function
157 mkProtoBCO nm instrs_ordlist origin arity bitmap_size bitmap
158 is_ret mallocd_blocks
161 protoBCOInstrs = maybe_with_stack_check,
162 protoBCOBitmap = bitmap,
163 protoBCOBitmapSize = bitmap_size,
164 protoBCOArity = arity,
165 protoBCOExpr = origin,
166 protoBCOPtrs = mallocd_blocks
169 -- Overestimate the stack usage (in words) of this BCO,
170 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
171 -- stack check. (The interpreter always does a stack check
172 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
173 -- BCO anyway, so we only need to add an explicit on in the
174 -- (hopefully rare) cases when the (overestimated) stack use
175 -- exceeds iNTERP_STACK_CHECK_THRESH.
176 maybe_with_stack_check
178 -- don't do stack checks at return points;
179 -- everything is aggregated up to the top BCO
180 -- (which must be a function)
181 | stack_overest >= 65535
182 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
184 | stack_overest >= iNTERP_STACK_CHECK_THRESH
185 = STKCHECK stack_overest : peep_d
187 = peep_d -- the supposedly common case
189 stack_overest = sum (map bciStackUse peep_d)
191 -- Merge local pushes
192 peep_d = peep (fromOL instrs_ordlist)
194 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
195 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
196 peep (PUSH_L off1 : PUSH_L off2 : rest)
197 = PUSH_LL off1 (off2-1) : peep rest
203 argBits :: [PrimRep] -> [Bool]
206 | isFollowableRep rep = False : argBits args
207 | otherwise = take (getPrimRepSize rep) (repeat True) ++ argBits args
209 mkBitmap :: [Bool] -> [StgWord]
211 mkBitmap stuff = chunkToLiveness chunk : mkBitmap rest
212 where (chunk, rest) = splitAt wORD_SIZE_IN_BITS stuff
214 chunkToLiveness :: [Bool] -> StgWord
215 chunkToLiveness chunk =
216 foldr (.|.) 0 [ 1 `shiftL` n | (True,n) <- zip chunk [0..] ]
218 -- make a bitmap where the slots specified are the *zeros* in the bitmap.
219 -- eg. [1,2,4], size 4 ==> 0x8 (we leave any bits outside the size as zero,
220 -- just to make the bitmap easier to read).
221 intsToBitmap :: Int -> [Int] -> [StgWord]
222 intsToBitmap size slots{- must be sorted -}
225 (foldr xor init (map (1 `shiftL`) these)) :
226 intsToBitmap (size - wORD_SIZE_IN_BITS)
227 (map (\x -> x - wORD_SIZE_IN_BITS) rest)
228 where (these,rest) = span (<wORD_SIZE_IN_BITS) slots
230 | size >= wORD_SIZE_IN_BITS = complement 0
231 | otherwise = (1 `shiftL` size) - 1
233 wORD_SIZE_IN_BITS = wORD_SIZE * 8 :: Int
235 -- -----------------------------------------------------------------------------
238 -- Compile code for the right-hand side of a top-level binding
240 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
243 schemeTopBind (id, rhs)
244 | Just data_con <- isDataConWorkId_maybe id,
245 isNullaryDataCon data_con
246 = -- Special case for the worker of a nullary data con.
247 -- It'll look like this: Nil = /\a -> Nil a
248 -- If we feed it into schemeR, we'll get
250 -- because mkConAppCode treats nullary constructor applications
251 -- by just re-using the single top-level definition. So
252 -- for the worker itself, we must allocate it directly.
253 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
254 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
257 = schemeR [{- No free variables -}] (id, rhs)
259 -- -----------------------------------------------------------------------------
262 -- Compile code for a right-hand side, to give a BCO that,
263 -- when executed with the free variables and arguments on top of the stack,
264 -- will return with a pointer to the result on top of the stack, after
265 -- removing the free variables and arguments.
267 -- Park the resulting BCO in the monad. Also requires the
268 -- variable to which this value was bound, so as to give the
269 -- resulting BCO a name.
271 schemeR :: [Id] -- Free vars of the RHS, ordered as they
272 -- will appear in the thunk. Empty for
273 -- top-level things, which have no free vars.
274 -> (Id, AnnExpr Id VarSet)
275 -> BcM (ProtoBCO Name)
276 schemeR fvs (nm, rhs)
280 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
281 $$ pprCoreExpr (deAnnotate rhs)
287 = schemeR_wrk fvs nm rhs (collect [] rhs)
289 collect xs (_, AnnNote note e) = collect xs e
290 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
291 collect xs (_, not_lambda) = (reverse xs, not_lambda)
293 schemeR_wrk fvs nm original_body (args, body)
295 all_args = reverse args ++ fvs
296 arity = length all_args
297 -- all_args are the args in reverse order. We're compiling a function
298 -- \fv1..fvn x1..xn -> e
299 -- i.e. the fvs come first
301 szsw_args = map idSizeW all_args
302 szw_args = sum szsw_args
303 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
305 -- make the arg bitmap
306 bits = argBits (reverse (map idPrimRep all_args))
307 bitmap_size = length bits
308 bitmap = mkBitmap bits
310 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
311 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
312 arity bitmap_size bitmap False{-not alts-})
315 fvsToEnv :: BCEnv -> VarSet -> [Id]
316 -- Takes the free variables of a right-hand side, and
317 -- delivers an ordered list of the local variables that will
318 -- be captured in the thunk for the RHS
319 -- The BCEnv argument tells which variables are in the local
320 -- environment: these are the ones that should be captured
322 -- The code that constructs the thunk, and the code that executes
323 -- it, have to agree about this layout
324 fvsToEnv p fvs = [v | v <- varSetElems fvs,
325 isId v, -- Could be a type variable
328 -- -----------------------------------------------------------------------------
331 -- Compile code to apply the given expression to the remaining args
332 -- on the stack, returning a HNF.
333 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
335 -- Delegate tail-calls to schemeT.
336 schemeE d s p e@(AnnApp f a)
339 schemeE d s p e@(AnnVar v)
340 | not (isUnLiftedType v_type)
341 = -- Lifted-type thing; push it in the normal way
345 = -- Returning an unlifted value.
346 -- Heave it on the stack, SLIDE, and RETURN.
347 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
348 returnBc (push -- value onto stack
349 `appOL` mkSLIDE szw (d-s) -- clear to sequel
350 `snocOL` RETURN_UBX v_rep) -- go
353 v_rep = typePrimRep v_type
355 schemeE d s p (AnnLit literal)
356 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
357 let l_rep = literalPrimRep literal
358 in returnBc (push -- value onto stack
359 `appOL` mkSLIDE szw (d-s) -- clear to sequel
360 `snocOL` RETURN_UBX l_rep) -- go
363 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
364 | (AnnVar v, args_r_to_l) <- splitApp rhs,
365 Just data_con <- isDataConWorkId_maybe v,
366 dataConRepArity data_con == length args_r_to_l
367 = -- Special case for a non-recursive let whose RHS is a
368 -- saturatred constructor application.
369 -- Just allocate the constructor and carry on
370 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
371 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
372 returnBc (alloc_code `appOL` body_code)
374 -- General case for let. Generates correct, if inefficient, code in
376 schemeE d s p (AnnLet binds (_,body))
377 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
378 AnnRec xs_n_rhss -> unzip xs_n_rhss
381 fvss = map (fvsToEnv p' . fst) rhss
383 -- Sizes of free vars, + 1 for the fn
384 sizes = map (\rhs_fvs -> 1 + sum (map idSizeW rhs_fvs)) fvss
386 -- the arity of each rhs
387 arities = map (length . fst . collect []) rhss
389 -- This p', d' defn is safe because all the items being pushed
390 -- are ptrs, so all have size 1. d' and p' reflect the stack
391 -- after the closures have been allocated in the heap (but not
392 -- filled in), and pointers to them parked on the stack.
393 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
395 zipE = zipEqual "schemeE"
397 -- ToDo: don't build thunks for things with no free variables
398 build_thunk dd [] size bco off
399 = returnBc (PUSH_BCO bco
400 `consOL` unitOL (MKAP (off+size-1) size))
401 build_thunk dd (fv:fvs) size bco off = do
402 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
403 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off
404 returnBc (push_code `appOL` more_push_code)
406 alloc_code = toOL (zipWith mkAlloc sizes arities)
407 where mkAlloc sz 0 = ALLOC_AP sz
408 mkAlloc sz arity = ALLOC_PAP arity sz
410 compile_bind d' fvs x rhs size off = do
411 bco <- schemeR fvs (x,rhs)
412 build_thunk d' fvs size bco off
415 [ compile_bind d' fvs x rhs size n
416 | (fvs, x, rhs, size, n) <-
417 zip5 fvss xs rhss sizes [n_binds, n_binds-1 .. 1]
420 body_code <- schemeE d' s p' body
421 thunk_codes <- sequence compile_binds
422 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
426 schemeE d s p (AnnCase scrut bndr [(DataAlt dc, [bind1, bind2], rhs)])
427 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind1)
429 -- case .... of x { (# VoidRep'd-thing, a #) -> ... }
431 -- case .... of a { DEFAULT -> ... }
432 -- becuse the return convention for both are identical.
434 -- Note that it does not matter losing the void-rep thing from the
435 -- envt (it won't be bound now) because we never look such things up.
437 = --trace "automagic mashing of case alts (# VoidRep, a #)" $
438 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
440 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind2)
441 = --trace "automagic mashing of case alts (# a, VoidRep #)" $
442 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
444 schemeE d s p (AnnCase scrut bndr [(DataAlt dc, [bind1], rhs)])
445 | isUnboxedTupleCon dc
446 -- Similarly, convert
447 -- case .... of x { (# a #) -> ... }
449 -- case .... of a { DEFAULT -> ... }
450 = --trace "automagic mashing of case alts (# a #)" $
451 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
453 schemeE d s p (AnnCase scrut bndr alts)
454 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
456 schemeE d s p (AnnNote note (_, body))
460 = pprPanic "ByteCodeGen.schemeE: unhandled case"
461 (pprCoreExpr (deAnnotate' other))
464 -- Compile code to do a tail call. Specifically, push the fn,
465 -- slide the on-stack app back down to the sequel depth,
466 -- and enter. Four cases:
469 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
470 -- The int will be on the stack. Generate a code sequence
471 -- to convert it to the relevant constructor, SLIDE and ENTER.
473 -- 1. The fn denotes a ccall. Defer to generateCCall.
475 -- 2. (Another nasty hack). Spot (# a::VoidRep, b #) and treat
476 -- it simply as b -- since the representations are identical
477 -- (the VoidRep takes up zero stack space). Also, spot
478 -- (# b #) and treat it as b.
480 -- 3. Application of a constructor, by defn saturated.
481 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
482 -- then the ptrs, and then do PACK and RETURN.
484 -- 4. Otherwise, it must be a function call. Push the args
485 -- right to left, SLIDE and ENTER.
487 schemeT :: Int -- Stack depth
488 -> Sequel -- Sequel depth
489 -> BCEnv -- stack env
490 -> AnnExpr' Id VarSet
495 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
496 -- = panic "schemeT ?!?!"
498 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
502 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
503 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
504 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
505 returnBc (push `appOL` tagToId_sequence
506 `appOL` mkSLIDE 1 (d+arg_words-s)
510 | Just (CCall ccall_spec) <- isFCallId_maybe fn
511 = generateCCall d s p ccall_spec fn args_r_to_l
513 -- Case 2: Constructor application
514 | Just con <- maybe_saturated_dcon,
515 isUnboxedTupleCon con
516 = case args_r_to_l of
517 [arg1,arg2] | isVoidRepAtom arg1 ->
518 unboxedTupleReturn d s p arg2
519 [arg1,arg2] | isVoidRepAtom arg2 ->
520 unboxedTupleReturn d s p arg1
521 _other -> unboxedTupleException
523 -- Case 3: Ordinary data constructor
524 | Just con <- maybe_saturated_dcon
525 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
526 returnBc (alloc_con `appOL`
527 mkSLIDE 1 (d - s) `snocOL`
530 -- Case 4: Tail call of function
532 = doTailCall d s p fn args_r_to_l
535 -- Detect and extract relevant info for the tagToEnum kludge.
536 maybe_is_tagToEnum_call
537 = let extract_constr_Names ty
538 = case splitTyConApp_maybe (repType ty) of
539 (Just (tyc, [])) | isDataTyCon tyc
540 -> map getName (tyConDataCons tyc)
541 other -> panic "maybe_is_tagToEnum_call.extract_constr_Ids"
544 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
545 -> case isPrimOpId_maybe v of
546 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
550 -- Extract the args (R->L) and fn
551 -- The function will necessarily be a variable,
552 -- because we are compiling a tail call
553 (AnnVar fn, args_r_to_l) = splitApp app
555 -- Only consider this to be a constructor application iff it is
556 -- saturated. Otherwise, we'll call the constructor wrapper.
557 n_args = length args_r_to_l
559 = case isDataConWorkId_maybe fn of
560 Just con | dataConRepArity con == n_args -> Just con
563 -- -----------------------------------------------------------------------------
564 -- Generate code to build a constructor application,
565 -- leaving it on top of the stack
567 mkConAppCode :: Int -> Sequel -> BCEnv
568 -> DataCon -- The data constructor
569 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
572 mkConAppCode orig_d s p con [] -- Nullary constructor
573 = ASSERT( isNullaryDataCon con )
574 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
575 -- Instead of doing a PACK, which would allocate a fresh
576 -- copy of this constructor, use the single shared version.
578 mkConAppCode orig_d s p con args_r_to_l
579 = ASSERT( dataConRepArity con == length args_r_to_l )
580 do_pushery orig_d (non_ptr_args ++ ptr_args)
582 -- The args are already in reverse order, which is the way PACK
583 -- expects them to be. We must push the non-ptrs after the ptrs.
584 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
586 do_pushery d (arg:args)
587 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
588 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
589 returnBc (push `appOL` more_push_code)
591 = returnBc (unitOL (PACK con n_arg_words))
593 n_arg_words = d - orig_d
596 -- -----------------------------------------------------------------------------
597 -- Returning an unboxed tuple with one non-void component (the only
598 -- case we can handle).
600 -- Remember, we don't want to *evaluate* the component that is being
601 -- returned, even if it is a pointed type. We always just return.
604 :: Int -> Sequel -> BCEnv
605 -> AnnExpr' Id VarSet -> BcM BCInstrList
606 unboxedTupleReturn d s p arg = do
607 (push, sz) <- pushAtom d p arg
608 returnBc (push `appOL`
609 mkSLIDE sz (d-s) `snocOL`
610 RETURN_UBX (atomRep arg))
612 -- -----------------------------------------------------------------------------
613 -- Generate code for a tail-call
616 :: Int -> Sequel -> BCEnv
617 -> Id -> [AnnExpr' Id VarSet]
619 doTailCall init_d s p fn args
620 = do_pushes init_d args (map (primRepToArgRep.atomRep) args)
622 do_pushes d [] reps = do
624 (push_fn, sz) <- pushAtom d p (AnnVar fn)
626 returnBc (push_fn `appOL` (
627 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
629 do_pushes d args reps = do
630 let (push_apply, n, rest_of_reps) = findPushSeq reps
631 (these_args, rest_of_args) = splitAt n args
632 (next_d, push_code) <- push_seq d these_args
633 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
634 -- ^^^ for the PUSH_APPLY_ instruction
635 returnBc (push_code `appOL` (push_apply `consOL` instrs))
637 push_seq d [] = return (d, nilOL)
638 push_seq d (arg:args) = do
639 (push_code, sz) <- pushAtom d p arg
640 (final_d, more_push_code) <- push_seq (d+sz) args
641 return (final_d, push_code `appOL` more_push_code)
643 -- v. similar to CgStackery.findMatch, ToDo: merge
644 findPushSeq (RepP: RepP: RepP: RepP: RepP: RepP: RepP: rest)
645 = (PUSH_APPLY_PPPPPPP, 7, rest)
646 findPushSeq (RepP: RepP: RepP: RepP: RepP: RepP: rest)
647 = (PUSH_APPLY_PPPPPP, 6, rest)
648 findPushSeq (RepP: RepP: RepP: RepP: RepP: rest)
649 = (PUSH_APPLY_PPPPP, 5, rest)
650 findPushSeq (RepP: RepP: RepP: RepP: rest)
651 = (PUSH_APPLY_PPPP, 4, rest)
652 findPushSeq (RepP: RepP: RepP: rest)
653 = (PUSH_APPLY_PPP, 3, rest)
654 findPushSeq (RepP: RepP: rest)
655 = (PUSH_APPLY_PP, 2, rest)
656 findPushSeq (RepP: rest)
657 = (PUSH_APPLY_P, 1, rest)
658 findPushSeq (RepV: rest)
659 = (PUSH_APPLY_V, 1, rest)
660 findPushSeq (RepN: rest)
661 = (PUSH_APPLY_N, 1, rest)
662 findPushSeq (RepF: rest)
663 = (PUSH_APPLY_F, 1, rest)
664 findPushSeq (RepD: rest)
665 = (PUSH_APPLY_D, 1, rest)
666 findPushSeq (RepL: rest)
667 = (PUSH_APPLY_L, 1, rest)
669 = panic "ByteCodeGen.findPushSeq"
671 -- -----------------------------------------------------------------------------
674 doCase :: Int -> Sequel -> BCEnv
675 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
676 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
678 doCase d s p (_,scrut)
679 bndr alts is_unboxed_tuple
681 -- Top of stack is the return itbl, as usual.
682 -- underneath it is the pointer to the alt_code BCO.
683 -- When an alt is entered, it assumes the returned value is
684 -- on top of the itbl.
687 -- An unlifted value gets an extra info table pushed on top
688 -- when it is returned.
689 unlifted_itbl_sizeW | isAlgCase = 0
692 -- depth of stack after the return value has been pushed
693 d_bndr = d + ret_frame_sizeW + idSizeW bndr
695 -- depth of stack after the extra info table for an unboxed return
696 -- has been pushed, if any. This is the stack depth at the
698 d_alts = d_bndr + unlifted_itbl_sizeW
700 -- Env in which to compile the alts, not including
701 -- any vars bound by the alts themselves
702 p_alts = addToFM p bndr (d_bndr - 1)
704 bndr_ty = idType bndr
705 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
707 -- given an alt, return a discr and code for it.
708 codeALt alt@(DEFAULT, _, (_,rhs))
709 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
710 returnBc (NoDiscr, rhs_code)
711 codeAlt alt@(discr, bndrs, (_,rhs))
712 -- primitive or nullary constructor alt: no need to UNPACK
713 | null real_bndrs = do
714 rhs_code <- schemeE d_alts s p_alts rhs
715 returnBc (my_discr alt, rhs_code)
716 -- algebraic alt with some binders
717 | ASSERT(isAlgCase) otherwise =
719 (ptrs,nptrs) = partition (isFollowableRep.idPrimRep) real_bndrs
720 ptr_sizes = map idSizeW ptrs
721 nptrs_sizes = map idSizeW nptrs
722 bind_sizes = ptr_sizes ++ nptrs_sizes
723 size = sum ptr_sizes + sum nptrs_sizes
724 -- the UNPACK instruction unpacks in reverse order...
725 p' = addListToFM p_alts
726 (zip (reverse (ptrs ++ nptrs))
727 (mkStackOffsets d_alts (reverse bind_sizes)))
729 rhs_code <- schemeE (d_alts+size) s p' rhs
730 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
732 real_bndrs = filter (not.isTyVar) bndrs
735 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
736 my_discr (DataAlt dc, binds, rhs)
737 | isUnboxedTupleCon dc
738 = unboxedTupleException
740 = DiscrP (dataConTag dc - fIRST_TAG)
741 my_discr (LitAlt l, binds, rhs)
742 = case l of MachInt i -> DiscrI (fromInteger i)
743 MachFloat r -> DiscrF (fromRational r)
744 MachDouble r -> DiscrD (fromRational r)
745 MachChar i -> DiscrI i
746 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
749 | not isAlgCase = Nothing
751 = case [dc | (DataAlt dc, _, _) <- alts] of
753 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
755 -- the bitmap is relative to stack depth d, i.e. before the
756 -- BCO, info table and return value are pushed on.
757 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
758 -- except that here we build the bitmap from the known bindings of
759 -- things that are pointers, whereas in CgBindery the code builds the
760 -- bitmap from the free slots and unboxed bindings.
762 bitmap = intsToBitmap d{-size-} (sortLt (<) rel_slots)
765 rel_slots = concat (map spread binds)
767 | isFollowableRep (idPrimRep id) = [ rel_offset ]
769 where rel_offset = d - offset - 1
772 alt_stuff <- mapM codeAlt alts
773 alt_final <- mkMultiBranch maybe_ncons alt_stuff
775 alt_bco_name = getName bndr
776 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
777 0{-no arity-} d{-bitmap size-} bitmap True{-is alts-}
779 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
780 -- "\n bitmap = " ++ show bitmap) $ do
781 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
782 alt_bco' <- emitBc alt_bco
784 | isAlgCase = PUSH_ALTS alt_bco'
785 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typePrimRep bndr_ty)
786 returnBc (push_alts `consOL` scrut_code)
789 -- -----------------------------------------------------------------------------
790 -- Deal with a CCall.
792 -- Taggedly push the args onto the stack R->L,
793 -- deferencing ForeignObj#s and adjusting addrs to point to
794 -- payloads in Ptr/Byte arrays. Then, generate the marshalling
795 -- (machine) code for the ccall, and create bytecodes to call that and
796 -- then return in the right way.
798 generateCCall :: Int -> Sequel -- stack and sequel depths
800 -> CCallSpec -- where to call
801 -> Id -- of target, for type info
802 -> [AnnExpr' Id VarSet] -- args (atoms)
805 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
808 addr_sizeW = getPrimRepSize AddrRep
810 -- Get the args on the stack, with tags and suitably
811 -- dereferenced for the CCall. For each arg, return the
812 -- depth to the first word of the bits for that arg, and the
813 -- PrimRep of what was actually pushed.
815 pargs d [] = returnBc []
817 = let arg_ty = repType (exprType (deAnnotate' a))
819 in case splitTyConApp_maybe arg_ty of
820 -- Don't push the FO; instead push the Addr# it
823 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
824 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
825 parg_ArrayishRep arrPtrsHdrSize d p a
827 returnBc ((code,AddrRep):rest)
829 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
830 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
831 parg_ArrayishRep arrWordsHdrSize d p a
833 returnBc ((code,AddrRep):rest)
835 -- Default case: push taggedly, but otherwise intact.
837 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
838 pargs (d+sz_a) az `thenBc` \ rest ->
839 returnBc ((code_a, atomRep a) : rest)
841 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
842 -- the stack but then advance it over the headers, so as to
843 -- point to the payload.
844 parg_ArrayishRep hdrSizeW d p a
845 = pushAtom d p a `thenBc` \ (push_fo, _) ->
846 -- The ptr points at the header. Advance it over the
847 -- header and then pretend this is an Addr#.
848 returnBc (push_fo `snocOL`
849 SWIZZLE 0 (hdrSizeW * getPrimRepSize WordRep
853 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
855 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
857 push_args = concatOL pushs_arg
858 d_after_args = d0 + sum (map getPrimRepSize a_reps_pushed_r_to_l)
860 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
861 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
863 = reverse (tail a_reps_pushed_r_to_l)
865 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
866 -- push_args is the code to do that.
867 -- d_after_args is the stack depth once the args are on.
869 -- Get the result rep.
870 (returns_void, r_rep)
871 = case maybe_getCCallReturnRep (idType fn) of
872 Nothing -> (True, VoidRep)
873 Just rr -> (False, rr)
875 Because the Haskell stack grows down, the a_reps refer to
876 lowest to highest addresses in that order. The args for the call
877 are on the stack. Now push an unboxed Addr# indicating
878 the C function to call. Then push a dummy placeholder for the
879 result. Finally, emit a CCALL insn with an offset pointing to the
880 Addr# just pushed, and a literal field holding the mallocville
881 address of the piece of marshalling code we generate.
882 So, just prior to the CCALL insn, the stack looks like this
883 (growing down, as usual):
888 Addr# address_of_C_fn
889 <placeholder-for-result#> (must be an unboxed type)
891 The interpreter then calls the marshall code mentioned
892 in the CCALL insn, passing it (& <placeholder-for-result#>),
893 that is, the addr of the topmost word in the stack.
894 When this returns, the placeholder will have been
895 filled in. The placeholder is slid down to the sequel
896 depth, and we RETURN.
898 This arrangement makes it simple to do f-i-dynamic since the Addr#
899 value is the first arg anyway.
901 The marshalling code is generated specifically for this
902 call site, and so knows exactly the (Haskell) stack
903 offsets of the args, fn address and placeholder. It
904 copies the args to the C stack, calls the stacked addr,
905 and parks the result back in the placeholder. The interpreter
906 calls it as a normal C call, assuming it has a signature
907 void marshall_code ( StgWord* ptr_to_top_of_stack )
909 -- resolve static address
913 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
915 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
918 -> pprPanic "ByteCodeGen.generateCCall: casm" (ppr ccall_spec)
920 get_target_info `thenBc` \ (is_static, static_target_addr) ->
923 -- Get the arg reps, zapping the leading Addr# in the dynamic case
924 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
925 | is_static = a_reps_pushed_RAW
926 | otherwise = if null a_reps_pushed_RAW
927 then panic "ByteCodeGen.generateCCall: dyn with no args"
928 else tail a_reps_pushed_RAW
931 (push_Addr, d_after_Addr)
933 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
934 d_after_args + addr_sizeW)
935 | otherwise -- is already on the stack
936 = (nilOL, d_after_args)
938 -- Push the return placeholder. For a call returning nothing,
939 -- this is a VoidRep (tag).
940 r_sizeW = getPrimRepSize r_rep
941 d_after_r = d_after_Addr + r_sizeW
942 r_lit = mkDummyLiteral r_rep
943 push_r = (if returns_void
945 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
947 -- generate the marshalling code we're going to call
950 arg1_offW = r_sizeW + addr_sizeW
951 args_offW = map (arg1_offW +)
952 (init (scanl (+) 0 (map getPrimRepSize a_reps)))
954 ioToBc (mkMarshalCode cconv
955 (r_offW, r_rep) addr_offW
956 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
957 recordMallocBc addr_of_marshaller `thenBc_`
959 -- Offset of the next stack frame down the stack. The CCALL
960 -- instruction needs to describe the chunk of stack containing
961 -- the ccall args to the GC, so it needs to know how large it
962 -- is. See comment in Interpreter.c with the CCALL instruction.
963 stk_offset = d_after_r - s
966 do_call = unitOL (CCALL stk_offset (castPtr addr_of_marshaller))
968 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
969 `snocOL` RETURN_UBX r_rep
971 --trace (show (arg1_offW, args_offW , (map getPrimRepSize a_reps) )) $
974 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
978 -- Make a dummy literal, to be used as a placeholder for FFI return
979 -- values on the stack.
980 mkDummyLiteral :: PrimRep -> Literal
983 CharRep -> MachChar 0
985 WordRep -> MachWord 0
986 DoubleRep -> MachDouble 0
987 FloatRep -> MachFloat 0
988 AddrRep | getPrimRepSize AddrRep == getPrimRepSize WordRep -> MachWord 0
989 _ -> moan64 "mkDummyLiteral" (ppr pr)
993 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
994 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
997 -- and check that an unboxed pair is returned wherein the first arg is VoidRep'd.
999 -- Alternatively, for call-targets returning nothing, convert
1001 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1002 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1006 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
1007 maybe_getCCallReturnRep fn_ty
1008 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1010 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1012 = case splitTyConApp_maybe (repType r_ty) of
1013 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
1015 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
1016 || r_reps == [VoidRep] )
1017 && isUnboxedTupleTyCon r_tycon
1018 && case maybe_r_rep_to_go of
1020 Just r_rep -> r_rep /= PtrRep
1021 -- if it was, it would be impossible
1022 -- to create a valid return value
1023 -- placeholder on the stack
1024 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1027 --trace (showSDoc (ppr (a_reps, r_reps))) $
1028 if ok then maybe_r_rep_to_go else blargh
1030 -- Compile code which expects an unboxed Int on the top of stack,
1031 -- (call it i), and pushes the i'th closure in the supplied list
1032 -- as a consequence.
1033 implement_tagToId :: [Name] -> BcM BCInstrList
1034 implement_tagToId names
1035 = ASSERT( notNull names )
1036 getLabelsBc (length names) `thenBc` \ labels ->
1037 getLabelBc `thenBc` \ label_fail ->
1038 getLabelBc `thenBc` \ label_exit ->
1039 zip4 labels (tail labels ++ [label_fail])
1040 [0 ..] names `bind` \ infos ->
1041 map (mkStep label_exit) infos `bind` \ steps ->
1042 returnBc (concatOL steps
1044 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1046 mkStep l_exit (my_label, next_label, n, name_for_n)
1047 = toOL [LABEL my_label,
1048 TESTEQ_I n next_label,
1053 -- -----------------------------------------------------------------------------
1056 -- Push an atom onto the stack, returning suitable code & number of
1057 -- stack words used.
1059 -- The env p must map each variable to the highest- numbered stack
1060 -- slot for it. For example, if the stack has depth 4 and we
1061 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1062 -- the tag in stack[5], the stack will have depth 6, and p must map v
1063 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1064 -- depth 6 stack has valid words 0 .. 5.
1066 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1068 pushAtom d p (AnnApp f (_, AnnType _))
1069 = pushAtom d p (snd f)
1071 pushAtom d p (AnnNote note e)
1072 = pushAtom d p (snd e)
1074 pushAtom d p (AnnLam x e)
1076 = pushAtom d p (snd e)
1078 pushAtom d p (AnnVar v)
1080 | idPrimRep v == VoidRep
1081 = returnBc (nilOL, 0)
1084 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1086 | Just primop <- isPrimOpId_maybe v
1087 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1089 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1090 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1091 -- d - d_v the number of words between the TOS
1092 -- and the 1st slot of the object
1094 -- d - d_v - 1 the offset from the TOS of the 1st slot
1096 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1099 -- Having found the last slot, we proceed to copy the right number of
1100 -- slots on to the top of the stack.
1102 | otherwise -- v must be a global variable
1104 returnBc (unitOL (PUSH_G (getName v)), sz)
1110 pushAtom d p (AnnLit lit)
1112 MachLabel fs -> code CodePtrRep
1113 MachWord w -> code WordRep
1114 MachInt i -> code IntRep
1115 MachFloat r -> code FloatRep
1116 MachDouble r -> code DoubleRep
1117 MachChar c -> code CharRep
1118 MachStr s -> pushStr s
1121 = let size_host_words = getPrimRepSize rep
1122 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1126 = let getMallocvilleAddr
1128 FastString _ l ba ->
1129 -- sigh, a string in the heap is no good to us.
1130 -- We need a static C pointer, since the type of
1131 -- a string literal is Addr#. So, copy the string
1132 -- into C land and remember the pointer so we can
1135 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1136 in ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1137 recordMallocBc ptr `thenBc_`
1139 do memcpy ptr ba (fromIntegral n)
1140 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1143 other -> panic "ByteCodeGen.pushAtom.pushStr"
1145 getMallocvilleAddr `thenBc` \ addr ->
1146 -- Get the addr on the stack, untaggedly
1147 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1150 = pprPanic "ByteCodeGen.pushAtom"
1151 (pprCoreExpr (deAnnotate (undefined, other)))
1153 foreign import ccall unsafe "memcpy"
1154 memcpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1157 -- -----------------------------------------------------------------------------
1158 -- Given a bunch of alts code and their discrs, do the donkey work
1159 -- of making a multiway branch using a switch tree.
1160 -- What a load of hassle!
1162 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1163 -- a hint; generates better code
1164 -- Nothing is always safe
1165 -> [(Discr, BCInstrList)]
1167 mkMultiBranch maybe_ncons raw_ways
1168 = let d_way = filter (isNoDiscr.fst) raw_ways
1169 notd_ways = naturalMergeSortLe
1170 (\w1 w2 -> leAlt (fst w1) (fst w2))
1171 (filter (not.isNoDiscr.fst) raw_ways)
1173 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1174 mkTree [] range_lo range_hi = returnBc the_default
1176 mkTree [val] range_lo range_hi
1177 | range_lo `eqAlt` range_hi
1178 = returnBc (snd val)
1180 = getLabelBc `thenBc` \ label_neq ->
1181 returnBc (mkTestEQ (fst val) label_neq
1183 `appOL` unitOL (LABEL label_neq)
1184 `appOL` the_default))
1186 mkTree vals range_lo range_hi
1187 = let n = length vals `div` 2
1188 vals_lo = take n vals
1189 vals_hi = drop n vals
1190 v_mid = fst (head vals_hi)
1192 getLabelBc `thenBc` \ label_geq ->
1193 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1194 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1195 returnBc (mkTestLT v_mid label_geq
1197 `appOL` unitOL (LABEL label_geq)
1201 = case d_way of [] -> unitOL CASEFAIL
1204 -- None of these will be needed if there are no non-default alts
1205 (mkTestLT, mkTestEQ, init_lo, init_hi)
1207 = panic "mkMultiBranch: awesome foursome"
1209 = case fst (head notd_ways) of {
1210 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1211 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1214 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1215 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1218 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1219 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1222 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1223 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1225 DiscrP algMaxBound )
1228 (algMinBound, algMaxBound)
1229 = case maybe_ncons of
1230 Just n -> (0, n - 1)
1231 Nothing -> (minBound, maxBound)
1233 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1234 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1235 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1236 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1237 NoDiscr `eqAlt` NoDiscr = True
1240 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1241 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1242 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1243 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1244 NoDiscr `leAlt` NoDiscr = True
1247 isNoDiscr NoDiscr = True
1250 dec (DiscrI i) = DiscrI (i-1)
1251 dec (DiscrP i) = DiscrP (i-1)
1252 dec other = other -- not really right, but if you
1253 -- do cases on floating values, you'll get what you deserve
1255 -- same snotty comment applies to the following
1257 minD, maxD :: Double
1263 mkTree notd_ways init_lo init_hi
1266 -- -----------------------------------------------------------------------------
1267 -- Supporting junk for the compilation schemes
1269 -- Describes case alts
1277 instance Outputable Discr where
1278 ppr (DiscrI i) = int i
1279 ppr (DiscrF f) = text (show f)
1280 ppr (DiscrD d) = text (show d)
1281 ppr (DiscrP i) = int i
1282 ppr NoDiscr = text "DEF"
1285 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1286 lookupBCEnv_maybe = lookupFM
1288 idSizeW :: Id -> Int
1289 idSizeW id = getPrimRepSize (typePrimRep (idType id))
1291 unboxedTupleException :: a
1292 unboxedTupleException
1295 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1296 "\tto foreign import/export decls in source. Workaround:\n" ++
1297 "\tcompile this module to a .o file, then restart session."))
1300 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1303 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1304 -- The arguments are returned in *right-to-left* order
1305 splitApp (AnnApp (_,f) (_,a))
1306 | isTypeAtom a = splitApp f
1307 | otherwise = case splitApp f of
1308 (f', as) -> (f', a:as)
1309 splitApp (AnnNote n (_,e)) = splitApp e
1310 splitApp e = (e, [])
1313 isTypeAtom :: AnnExpr' id ann -> Bool
1314 isTypeAtom (AnnType _) = True
1315 isTypeAtom _ = False
1317 isVoidRepAtom :: AnnExpr' id ann -> Bool
1318 isVoidRepAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1319 isVoidRepAtom (AnnNote n (_,e)) = isVoidRepAtom e
1320 isVoidRepAtom _ = False
1322 atomRep :: AnnExpr' Id ann -> PrimRep
1323 atomRep (AnnVar v) = typePrimRep (idType v)
1324 atomRep (AnnLit l) = literalPrimRep l
1325 atomRep (AnnNote n b) = atomRep (snd b)
1326 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1327 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1328 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1330 isPtrAtom :: AnnExpr' Id ann -> Bool
1331 isPtrAtom e = isFollowableRep (atomRep e)
1333 -- Let szsw be the sizes in words of some items pushed onto the stack,
1334 -- which has initial depth d'. Return the values which the stack environment
1335 -- should map these items to.
1336 mkStackOffsets :: Int -> [Int] -> [Int]
1337 mkStackOffsets original_depth szsw
1338 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1340 -- -----------------------------------------------------------------------------
1341 -- The bytecode generator's monad
1345 nextlabel :: Int, -- for generating local labels
1346 malloced :: [Ptr ()] } -- ptrs malloced for current BCO
1347 -- Should be free()d when it is GCd
1349 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1351 ioToBc :: IO a -> BcM a
1352 ioToBc io = BcM $ \st -> do
1356 runBc :: BcM r -> IO (BcM_State, r)
1357 runBc (BcM m) = m (BcM_State 0 [])
1359 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1360 thenBc (BcM expr) cont = BcM $ \st0 -> do
1361 (st1, q) <- expr st0
1366 thenBc_ :: BcM a -> BcM b -> BcM b
1367 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1368 (st1, q) <- expr st0
1369 (st2, r) <- cont st1
1372 returnBc :: a -> BcM a
1373 returnBc result = BcM $ \st -> (return (st, result))
1375 instance Monad BcM where
1380 emitBc :: ([Ptr ()] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1382 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1384 recordMallocBc :: Ptr a -> BcM ()
1386 = BcM $ \st -> return (st{malloced = castPtr a : malloced st}, ())
1388 getLabelBc :: BcM Int
1390 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1392 getLabelsBc :: Int -> BcM [Int]
1394 = BcM $ \st -> let ctr = nextlabel st
1395 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])