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, typeEnvTyCons, typeEnvClasses )
23 import CoreUtils ( exprType )
25 import PprCore ( pprCoreExpr )
26 import Literal ( Literal(..), literalPrimRep )
28 import PrimOp ( PrimOp(..) )
29 import CoreFVs ( freeVars )
30 import Type ( typePrimRep, isUnLiftedType, splitTyConApp_maybe,
32 import DataCon ( DataCon, dataConTag, fIRST_TAG, dataConTyCon,
33 isUnboxedTupleCon, isNullaryDataCon, dataConWorkId,
35 import TyCon ( tyConFamilySize, isDataTyCon, tyConDataCons,
36 isFunTyCon, isUnboxedTupleTyCon )
37 import Class ( Class, classTyCon )
38 import Type ( Type, repType, splitFunTys, dropForAlls )
40 import DataCon ( dataConRepArity )
41 import Var ( isTyVar )
42 import VarSet ( VarSet, varSetElems )
43 import TysPrim ( foreignObjPrimTyCon,
44 arrayPrimTyCon, mutableArrayPrimTyCon,
45 byteArrayPrimTyCon, mutableByteArrayPrimTyCon
47 import PrimRep ( isFollowableRep )
48 import CmdLineOpts ( DynFlags, DynFlag(..) )
49 import ErrUtils ( showPass, dumpIfSet_dyn )
50 import Unique ( mkPseudoUnique3 )
51 import FastString ( FastString(..), unpackFS )
52 import Panic ( GhcException(..) )
53 import PprType ( pprType )
54 import SMRep ( arrWordsHdrSize, arrPtrsHdrSize )
56 import Constants ( wORD_SIZE )
57 import BasicTypes ( TopLevelFlag(..), isTopLevel, isNotTopLevel )
59 import Data.List ( intersperse, sortBy, zip4, zip5, partition )
60 import Foreign ( Ptr, castPtr, mallocBytes, pokeByteOff, Word8 )
61 import Foreign.C ( CInt )
62 import Control.Exception ( throwDyn )
64 import GHC.Exts ( Int(..), ByteArray# )
66 import Control.Monad ( when, mapAndUnzipM )
67 import Data.Char ( ord )
70 -- -----------------------------------------------------------------------------
71 -- Generating byte code for a complete module
73 byteCodeGen :: DynFlags
75 -> IO CompiledByteCode
76 byteCodeGen dflags (ModGuts { mg_binds = binds, mg_types = type_env })
77 = do showPass dflags "ByteCodeGen"
78 let local_tycons = typeEnvTyCons type_env
79 local_classes = typeEnvClasses type_env
80 tycs = local_tycons ++ map classTyCon local_classes
82 let flatBinds = [ (bndr, freeVars rhs)
83 | (bndr, rhs) <- flattenBinds binds]
85 (BcM_State final_ctr mallocd, proto_bcos)
86 <- runBc (mapM schemeTopBind flatBinds)
88 when (notNull mallocd)
89 (panic "ByteCodeGen.byteCodeGen: missing final emitBc?")
91 dumpIfSet_dyn dflags Opt_D_dump_BCOs
92 "Proto-BCOs" (vcat (intersperse (char ' ') (map ppr proto_bcos)))
94 assembleBCOs proto_bcos tycs
96 -- -----------------------------------------------------------------------------
97 -- Generating byte code for an expression
99 -- Returns: (the root BCO for this expression,
100 -- a list of auxilary BCOs resulting from compiling closures)
101 coreExprToBCOs :: DynFlags
104 coreExprToBCOs dflags expr
105 = do showPass dflags "ByteCodeGen"
107 -- create a totally bogus name for the top-level BCO; this
108 -- should be harmless, since it's never used for anything
109 let invented_name = mkSystemName (mkPseudoUnique3 0) FSLIT("ExprTopLevel")
110 invented_id = mkLocalId invented_name (panic "invented_id's type")
112 (BcM_State final_ctr mallocd, proto_bco)
113 <- runBc (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 (idPrimRep 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
156 is_ret mallocd_blocks
159 protoBCOInstrs = maybe_with_stack_check,
160 protoBCOBitmap = bitmap,
161 protoBCOBitmapSize = bitmap_size,
162 protoBCOArity = arity,
163 protoBCOExpr = origin,
164 protoBCOPtrs = mallocd_blocks
167 -- Overestimate the stack usage (in words) of this BCO,
168 -- and if >= iNTERP_STACK_CHECK_THRESH, add an explicit
169 -- stack check. (The interpreter always does a stack check
170 -- for iNTERP_STACK_CHECK_THRESH words at the start of each
171 -- BCO anyway, so we only need to add an explicit on in the
172 -- (hopefully rare) cases when the (overestimated) stack use
173 -- exceeds iNTERP_STACK_CHECK_THRESH.
174 maybe_with_stack_check
176 -- don't do stack checks at return points;
177 -- everything is aggregated up to the top BCO
178 -- (which must be a function)
179 | stack_overest >= 65535
180 = pprPanic "mkProtoBCO: stack use won't fit in 16 bits"
182 | stack_overest >= iNTERP_STACK_CHECK_THRESH
183 = STKCHECK stack_overest : peep_d
185 = peep_d -- the supposedly common case
187 stack_overest = sum (map bciStackUse peep_d)
189 -- Merge local pushes
190 peep_d = peep (fromOL instrs_ordlist)
192 peep (PUSH_L off1 : PUSH_L off2 : PUSH_L off3 : rest)
193 = PUSH_LLL off1 (off2-1) (off3-2) : peep rest
194 peep (PUSH_L off1 : PUSH_L off2 : rest)
195 = PUSH_LL off1 (off2-1) : peep rest
201 argBits :: [PrimRep] -> [Bool]
204 | isFollowableRep rep = False : argBits args
205 | otherwise = take (getPrimRepSize rep) (repeat True) ++ argBits args
207 mkBitmap :: [Bool] -> [StgWord]
209 mkBitmap stuff = chunkToLiveness chunk : mkBitmap rest
210 where (chunk, rest) = splitAt wORD_SIZE_IN_BITS stuff
212 chunkToLiveness :: [Bool] -> StgWord
213 chunkToLiveness chunk =
214 foldr (.|.) 0 [ 1 `shiftL` n | (True,n) <- zip chunk [0..] ]
216 -- make a bitmap where the slots specified are the *zeros* in the bitmap.
217 -- eg. [1,2,4], size 4 ==> 0x8 (we leave any bits outside the size as zero,
218 -- just to make the bitmap easier to read).
219 intsToBitmap :: Int -> [Int] -> [StgWord]
220 intsToBitmap size slots{- must be sorted -}
223 (foldr xor init (map (1 `shiftL`) these)) :
224 intsToBitmap (size - wORD_SIZE_IN_BITS)
225 (map (\x -> x - wORD_SIZE_IN_BITS) rest)
226 where (these,rest) = span (<wORD_SIZE_IN_BITS) slots
228 | size >= wORD_SIZE_IN_BITS = complement 0
229 | otherwise = (1 `shiftL` size) - 1
231 wORD_SIZE_IN_BITS = wORD_SIZE * 8 :: Int
233 -- -----------------------------------------------------------------------------
236 -- Compile code for the right-hand side of a top-level binding
238 schemeTopBind :: (Id, AnnExpr Id VarSet) -> BcM (ProtoBCO Name)
241 schemeTopBind (id, rhs)
242 | Just data_con <- isDataConWorkId_maybe id,
243 isNullaryDataCon data_con
244 = -- Special case for the worker of a nullary data con.
245 -- It'll look like this: Nil = /\a -> Nil a
246 -- If we feed it into schemeR, we'll get
248 -- because mkConAppCode treats nullary constructor applications
249 -- by just re-using the single top-level definition. So
250 -- for the worker itself, we must allocate it directly.
251 emitBc (mkProtoBCO (getName id) (toOL [PACK data_con 0, ENTER])
252 (Right rhs) 0 0 [{-no bitmap-}] False{-not alts-})
255 = schemeR [{- No free variables -}] (id, rhs)
257 -- -----------------------------------------------------------------------------
260 -- Compile code for a right-hand side, to give a BCO that,
261 -- when executed with the free variables and arguments on top of the stack,
262 -- will return with a pointer to the result on top of the stack, after
263 -- removing the free variables and arguments.
265 -- Park the resulting BCO in the monad. Also requires the
266 -- variable to which this value was bound, so as to give the
267 -- resulting BCO a name.
269 schemeR :: [Id] -- Free vars of the RHS, ordered as they
270 -- will appear in the thunk. Empty for
271 -- top-level things, which have no free vars.
272 -> (Id, AnnExpr Id VarSet)
273 -> BcM (ProtoBCO Name)
274 schemeR fvs (nm, rhs)
278 $$ (ppr.filter (not.isTyVar).varSetElems.fst) rhs
279 $$ pprCoreExpr (deAnnotate rhs)
285 = schemeR_wrk fvs nm rhs (collect [] rhs)
287 collect xs (_, AnnNote note e) = collect xs e
288 collect xs (_, AnnLam x e) = collect (if isTyVar x then xs else (x:xs)) e
289 collect xs (_, not_lambda) = (reverse xs, not_lambda)
291 schemeR_wrk fvs nm original_body (args, body)
293 all_args = reverse args ++ fvs
294 arity = length all_args
295 -- all_args are the args in reverse order. We're compiling a function
296 -- \fv1..fvn x1..xn -> e
297 -- i.e. the fvs come first
299 szsw_args = map idSizeW all_args
300 szw_args = sum szsw_args
301 p_init = listToFM (zip all_args (mkStackOffsets 0 szsw_args))
303 -- make the arg bitmap
304 bits = argBits (reverse (map idPrimRep all_args))
305 bitmap_size = length bits
306 bitmap = mkBitmap bits
308 schemeE szw_args 0 p_init body `thenBc` \ body_code ->
309 emitBc (mkProtoBCO (getName nm) body_code (Right original_body)
310 arity bitmap_size bitmap False{-not alts-})
313 fvsToEnv :: BCEnv -> VarSet -> [Id]
314 -- Takes the free variables of a right-hand side, and
315 -- delivers an ordered list of the local variables that will
316 -- be captured in the thunk for the RHS
317 -- The BCEnv argument tells which variables are in the local
318 -- environment: these are the ones that should be captured
320 -- The code that constructs the thunk, and the code that executes
321 -- it, have to agree about this layout
322 fvsToEnv p fvs = [v | v <- varSetElems fvs,
323 isId v, -- Could be a type variable
326 -- -----------------------------------------------------------------------------
329 -- Compile code to apply the given expression to the remaining args
330 -- on the stack, returning a HNF.
331 schemeE :: Int -> Sequel -> BCEnv -> AnnExpr' Id VarSet -> BcM BCInstrList
333 -- Delegate tail-calls to schemeT.
334 schemeE d s p e@(AnnApp f a)
337 schemeE d s p e@(AnnVar v)
338 | not (isUnLiftedType v_type)
339 = -- Lifted-type thing; push it in the normal way
343 = -- Returning an unlifted value.
344 -- Heave it on the stack, SLIDE, and RETURN.
345 pushAtom d p (AnnVar v) `thenBc` \ (push, szw) ->
346 returnBc (push -- value onto stack
347 `appOL` mkSLIDE szw (d-s) -- clear to sequel
348 `snocOL` RETURN_UBX v_rep) -- go
351 v_rep = typePrimRep v_type
353 schemeE d s p (AnnLit literal)
354 = pushAtom d p (AnnLit literal) `thenBc` \ (push, szw) ->
355 let l_rep = literalPrimRep literal
356 in returnBc (push -- value onto stack
357 `appOL` mkSLIDE szw (d-s) -- clear to sequel
358 `snocOL` RETURN_UBX l_rep) -- go
361 schemeE d s p (AnnLet (AnnNonRec x (_,rhs)) (_,body))
362 | (AnnVar v, args_r_to_l) <- splitApp rhs,
363 Just data_con <- isDataConWorkId_maybe v,
364 dataConRepArity data_con == length args_r_to_l
365 = -- Special case for a non-recursive let whose RHS is a
366 -- saturatred constructor application.
367 -- Just allocate the constructor and carry on
368 mkConAppCode d s p data_con args_r_to_l `thenBc` \ alloc_code ->
369 schemeE (d+1) s (addToFM p x d) body `thenBc` \ body_code ->
370 returnBc (alloc_code `appOL` body_code)
372 -- General case for let. Generates correct, if inefficient, code in
374 schemeE d s p (AnnLet binds (_,body))
375 = let (xs,rhss) = case binds of AnnNonRec x rhs -> ([x],[rhs])
376 AnnRec xs_n_rhss -> unzip xs_n_rhss
379 fvss = map (fvsToEnv p' . fst) rhss
381 -- Sizes of free vars, + 1 for the fn
382 sizes = map (\rhs_fvs -> 1 + sum (map idSizeW rhs_fvs)) fvss
384 -- the arity of each rhs
385 arities = map (length . fst . collect []) rhss
387 -- This p', d' defn is safe because all the items being pushed
388 -- are ptrs, so all have size 1. d' and p' reflect the stack
389 -- after the closures have been allocated in the heap (but not
390 -- filled in), and pointers to them parked on the stack.
391 p' = addListToFM p (zipE xs (mkStackOffsets d (nOfThem n_binds 1)))
393 zipE = zipEqual "schemeE"
395 -- ToDo: don't build thunks for things with no free variables
396 build_thunk dd [] size bco off
397 = returnBc (PUSH_BCO bco
398 `consOL` unitOL (MKAP (off+size-1) size))
399 build_thunk dd (fv:fvs) size bco off = do
400 (push_code, pushed_szw) <- pushAtom dd p' (AnnVar fv)
401 more_push_code <- build_thunk (dd+pushed_szw) fvs size bco off
402 returnBc (push_code `appOL` more_push_code)
404 alloc_code = toOL (zipWith mkAlloc sizes arities)
405 where mkAlloc sz 0 = ALLOC_AP sz
406 mkAlloc sz arity = ALLOC_PAP arity sz
408 compile_bind d' fvs x rhs size off = do
409 bco <- schemeR fvs (x,rhs)
410 build_thunk d' fvs size bco off
413 [ compile_bind d' fvs x rhs size n
414 | (fvs, x, rhs, size, n) <-
415 zip5 fvss xs rhss sizes [n_binds, n_binds-1 .. 1]
418 body_code <- schemeE d' s p' body
419 thunk_codes <- sequence compile_binds
420 returnBc (alloc_code `appOL` concatOL thunk_codes `appOL` body_code)
424 schemeE d s p (AnnCase scrut bndr [(DataAlt dc, [bind1, bind2], rhs)])
425 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind1)
427 -- case .... of x { (# VoidRep'd-thing, a #) -> ... }
429 -- case .... of a { DEFAULT -> ... }
430 -- becuse the return convention for both are identical.
432 -- Note that it does not matter losing the void-rep thing from the
433 -- envt (it won't be bound now) because we never look such things up.
435 = --trace "automagic mashing of case alts (# VoidRep, a #)" $
436 doCase d s p scrut bind2 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
438 | isUnboxedTupleCon dc && VoidRep == typePrimRep (idType bind2)
439 = --trace "automagic mashing of case alts (# a, VoidRep #)" $
440 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
442 schemeE d s p (AnnCase scrut bndr [(DataAlt dc, [bind1], rhs)])
443 | isUnboxedTupleCon dc
444 -- Similarly, convert
445 -- case .... of x { (# a #) -> ... }
447 -- case .... of a { DEFAULT -> ... }
448 = --trace "automagic mashing of case alts (# a #)" $
449 doCase d s p scrut bind1 [(DEFAULT, [], rhs)] True{-unboxed tuple-}
451 schemeE d s p (AnnCase scrut bndr alts)
452 = doCase d s p scrut bndr alts False{-not an unboxed tuple-}
454 schemeE d s p (AnnNote note (_, body))
458 = pprPanic "ByteCodeGen.schemeE: unhandled case"
459 (pprCoreExpr (deAnnotate' other))
462 -- Compile code to do a tail call. Specifically, push the fn,
463 -- slide the on-stack app back down to the sequel depth,
464 -- and enter. Four cases:
467 -- An application "GHC.Prim.tagToEnum# <type> unboxed-int".
468 -- The int will be on the stack. Generate a code sequence
469 -- to convert it to the relevant constructor, SLIDE and ENTER.
471 -- 1. The fn denotes a ccall. Defer to generateCCall.
473 -- 2. (Another nasty hack). Spot (# a::VoidRep, b #) and treat
474 -- it simply as b -- since the representations are identical
475 -- (the VoidRep takes up zero stack space). Also, spot
476 -- (# b #) and treat it as b.
478 -- 3. Application of a constructor, by defn saturated.
479 -- Split the args into ptrs and non-ptrs, and push the nonptrs,
480 -- then the ptrs, and then do PACK and RETURN.
482 -- 4. Otherwise, it must be a function call. Push the args
483 -- right to left, SLIDE and ENTER.
485 schemeT :: Int -- Stack depth
486 -> Sequel -- Sequel depth
487 -> BCEnv -- stack env
488 -> AnnExpr' Id VarSet
493 -- | trace ("schemeT: env in = \n" ++ showSDocDebug (ppBCEnv p)) False
494 -- = panic "schemeT ?!?!"
496 -- | trace ("\nschemeT\n" ++ showSDoc (pprCoreExpr (deAnnotate' app)) ++ "\n") False
500 | Just (arg, constr_names) <- maybe_is_tagToEnum_call
501 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
502 implement_tagToId constr_names `thenBc` \ tagToId_sequence ->
503 returnBc (push `appOL` tagToId_sequence
504 `appOL` mkSLIDE 1 (d+arg_words-s)
508 | Just (CCall ccall_spec) <- isFCallId_maybe fn
509 = generateCCall d s p ccall_spec fn args_r_to_l
511 -- Case 2: Constructor application
512 | Just con <- maybe_saturated_dcon,
513 isUnboxedTupleCon con
514 = case args_r_to_l of
515 [arg1,arg2] | isVoidRepAtom arg1 ->
516 unboxedTupleReturn d s p arg2
517 [arg1,arg2] | isVoidRepAtom arg2 ->
518 unboxedTupleReturn d s p arg1
519 _other -> unboxedTupleException
521 -- Case 3: Ordinary data constructor
522 | Just con <- maybe_saturated_dcon
523 = mkConAppCode d s p con args_r_to_l `thenBc` \ alloc_con ->
524 returnBc (alloc_con `appOL`
525 mkSLIDE 1 (d - s) `snocOL`
528 -- Case 4: Tail call of function
530 = doTailCall d s p fn args_r_to_l
533 -- Detect and extract relevant info for the tagToEnum kludge.
534 maybe_is_tagToEnum_call
535 = let extract_constr_Names ty
536 = case splitTyConApp_maybe (repType ty) of
537 (Just (tyc, [])) | isDataTyCon tyc
538 -> map getName (tyConDataCons tyc)
539 other -> panic "maybe_is_tagToEnum_call.extract_constr_Ids"
542 (AnnApp (_, AnnApp (_, AnnVar v) (_, AnnType t)) arg)
543 -> case isPrimOpId_maybe v of
544 Just TagToEnumOp -> Just (snd arg, extract_constr_Names t)
548 -- Extract the args (R->L) and fn
549 -- The function will necessarily be a variable,
550 -- because we are compiling a tail call
551 (AnnVar fn, args_r_to_l) = splitApp app
553 -- Only consider this to be a constructor application iff it is
554 -- saturated. Otherwise, we'll call the constructor wrapper.
555 n_args = length args_r_to_l
557 = case isDataConWorkId_maybe fn of
558 Just con | dataConRepArity con == n_args -> Just con
561 -- -----------------------------------------------------------------------------
562 -- Generate code to build a constructor application,
563 -- leaving it on top of the stack
565 mkConAppCode :: Int -> Sequel -> BCEnv
566 -> DataCon -- The data constructor
567 -> [AnnExpr' Id VarSet] -- Args, in *reverse* order
570 mkConAppCode orig_d s p con [] -- Nullary constructor
571 = ASSERT( isNullaryDataCon con )
572 returnBc (unitOL (PUSH_G (getName (dataConWorkId con))))
573 -- Instead of doing a PACK, which would allocate a fresh
574 -- copy of this constructor, use the single shared version.
576 mkConAppCode orig_d s p con args_r_to_l
577 = ASSERT( dataConRepArity con == length args_r_to_l )
578 do_pushery orig_d (non_ptr_args ++ ptr_args)
580 -- The args are already in reverse order, which is the way PACK
581 -- expects them to be. We must push the non-ptrs after the ptrs.
582 (ptr_args, non_ptr_args) = partition isPtrAtom args_r_to_l
584 do_pushery d (arg:args)
585 = pushAtom d p arg `thenBc` \ (push, arg_words) ->
586 do_pushery (d+arg_words) args `thenBc` \ more_push_code ->
587 returnBc (push `appOL` more_push_code)
589 = returnBc (unitOL (PACK con n_arg_words))
591 n_arg_words = d - orig_d
594 -- -----------------------------------------------------------------------------
595 -- Returning an unboxed tuple with one non-void component (the only
596 -- case we can handle).
598 -- Remember, we don't want to *evaluate* the component that is being
599 -- returned, even if it is a pointed type. We always just return.
602 :: Int -> Sequel -> BCEnv
603 -> AnnExpr' Id VarSet -> BcM BCInstrList
604 unboxedTupleReturn d s p arg = do
605 (push, sz) <- pushAtom d p arg
606 returnBc (push `appOL`
607 mkSLIDE sz (d-s) `snocOL`
608 RETURN_UBX (atomRep arg))
610 -- -----------------------------------------------------------------------------
611 -- Generate code for a tail-call
614 :: Int -> Sequel -> BCEnv
615 -> Id -> [AnnExpr' Id VarSet]
617 doTailCall init_d s p fn args
618 = do_pushes init_d args (map (primRepToArgRep.atomRep) args)
620 do_pushes d [] reps = do
622 (push_fn, sz) <- pushAtom d p (AnnVar fn)
624 returnBc (push_fn `appOL` (
625 mkSLIDE ((d-init_d) + 1) (init_d - s) `appOL`
627 do_pushes d args reps = do
628 let (push_apply, n, rest_of_reps) = findPushSeq reps
629 (these_args, rest_of_args) = splitAt n args
630 (next_d, push_code) <- push_seq d these_args
631 instrs <- do_pushes (next_d + 1) rest_of_args rest_of_reps
632 -- ^^^ for the PUSH_APPLY_ instruction
633 returnBc (push_code `appOL` (push_apply `consOL` instrs))
635 push_seq d [] = return (d, nilOL)
636 push_seq d (arg:args) = do
637 (push_code, sz) <- pushAtom d p arg
638 (final_d, more_push_code) <- push_seq (d+sz) args
639 return (final_d, push_code `appOL` more_push_code)
641 -- v. similar to CgStackery.findMatch, ToDo: merge
642 findPushSeq (RepP: RepP: RepP: RepP: RepP: RepP: RepP: rest)
643 = (PUSH_APPLY_PPPPPPP, 7, rest)
644 findPushSeq (RepP: RepP: RepP: RepP: RepP: RepP: rest)
645 = (PUSH_APPLY_PPPPPP, 6, rest)
646 findPushSeq (RepP: RepP: RepP: RepP: RepP: rest)
647 = (PUSH_APPLY_PPPPP, 5, rest)
648 findPushSeq (RepP: RepP: RepP: RepP: rest)
649 = (PUSH_APPLY_PPPP, 4, rest)
650 findPushSeq (RepP: RepP: RepP: rest)
651 = (PUSH_APPLY_PPP, 3, rest)
652 findPushSeq (RepP: RepP: rest)
653 = (PUSH_APPLY_PP, 2, rest)
654 findPushSeq (RepP: rest)
655 = (PUSH_APPLY_P, 1, rest)
656 findPushSeq (RepV: rest)
657 = (PUSH_APPLY_V, 1, rest)
658 findPushSeq (RepN: rest)
659 = (PUSH_APPLY_N, 1, rest)
660 findPushSeq (RepF: rest)
661 = (PUSH_APPLY_F, 1, rest)
662 findPushSeq (RepD: rest)
663 = (PUSH_APPLY_D, 1, rest)
664 findPushSeq (RepL: rest)
665 = (PUSH_APPLY_L, 1, rest)
667 = panic "ByteCodeGen.findPushSeq"
669 -- -----------------------------------------------------------------------------
672 doCase :: Int -> Sequel -> BCEnv
673 -> AnnExpr Id VarSet -> Id -> [AnnAlt Id VarSet]
674 -> Bool -- True <=> is an unboxed tuple case, don't enter the result
676 doCase d s p (_,scrut)
677 bndr alts is_unboxed_tuple
679 -- Top of stack is the return itbl, as usual.
680 -- underneath it is the pointer to the alt_code BCO.
681 -- When an alt is entered, it assumes the returned value is
682 -- on top of the itbl.
685 -- An unlifted value gets an extra info table pushed on top
686 -- when it is returned.
687 unlifted_itbl_sizeW | isAlgCase = 0
690 -- depth of stack after the return value has been pushed
691 d_bndr = d + ret_frame_sizeW + idSizeW bndr
693 -- depth of stack after the extra info table for an unboxed return
694 -- has been pushed, if any. This is the stack depth at the
696 d_alts = d_bndr + unlifted_itbl_sizeW
698 -- Env in which to compile the alts, not including
699 -- any vars bound by the alts themselves
700 p_alts = addToFM p bndr (d_bndr - 1)
702 bndr_ty = idType bndr
703 isAlgCase = not (isUnLiftedType bndr_ty) && not is_unboxed_tuple
705 -- given an alt, return a discr and code for it.
706 codeALt alt@(DEFAULT, _, (_,rhs))
707 = schemeE d_alts s p_alts rhs `thenBc` \ rhs_code ->
708 returnBc (NoDiscr, rhs_code)
709 codeAlt alt@(discr, bndrs, (_,rhs))
710 -- primitive or nullary constructor alt: no need to UNPACK
711 | null real_bndrs = do
712 rhs_code <- schemeE d_alts s p_alts rhs
713 returnBc (my_discr alt, rhs_code)
714 -- algebraic alt with some binders
715 | ASSERT(isAlgCase) otherwise =
717 (ptrs,nptrs) = partition (isFollowableRep.idPrimRep) real_bndrs
718 ptr_sizes = map idSizeW ptrs
719 nptrs_sizes = map idSizeW nptrs
720 bind_sizes = ptr_sizes ++ nptrs_sizes
721 size = sum ptr_sizes + sum nptrs_sizes
722 -- the UNPACK instruction unpacks in reverse order...
723 p' = addListToFM p_alts
724 (zip (reverse (ptrs ++ nptrs))
725 (mkStackOffsets d_alts (reverse bind_sizes)))
727 rhs_code <- schemeE (d_alts+size) s p' rhs
728 return (my_discr alt, unitOL (UNPACK size) `appOL` rhs_code)
730 real_bndrs = filter (not.isTyVar) bndrs
733 my_discr (DEFAULT, binds, rhs) = NoDiscr {-shouldn't really happen-}
734 my_discr (DataAlt dc, binds, rhs)
735 | isUnboxedTupleCon dc
736 = unboxedTupleException
738 = DiscrP (dataConTag dc - fIRST_TAG)
739 my_discr (LitAlt l, binds, rhs)
740 = case l of MachInt i -> DiscrI (fromInteger i)
741 MachFloat r -> DiscrF (fromRational r)
742 MachDouble r -> DiscrD (fromRational r)
743 MachChar i -> DiscrI i
744 _ -> pprPanic "schemeE(AnnCase).my_discr" (ppr l)
747 | not isAlgCase = Nothing
749 = case [dc | (DataAlt dc, _, _) <- alts] of
751 (dc:_) -> Just (tyConFamilySize (dataConTyCon dc))
753 -- the bitmap is relative to stack depth d, i.e. before the
754 -- BCO, info table and return value are pushed on.
755 -- This bit of code is v. similar to buildLivenessMask in CgBindery,
756 -- except that here we build the bitmap from the known bindings of
757 -- things that are pointers, whereas in CgBindery the code builds the
758 -- bitmap from the free slots and unboxed bindings.
760 bitmap = intsToBitmap d{-size-} (sortLt (<) rel_slots)
763 rel_slots = concat (map spread binds)
765 | isFollowableRep (idPrimRep id) = [ rel_offset ]
767 where rel_offset = d - offset - 1
770 alt_stuff <- mapM codeAlt alts
771 alt_final <- mkMultiBranch maybe_ncons alt_stuff
773 alt_bco_name = getName bndr
774 alt_bco = mkProtoBCO alt_bco_name alt_final (Left alts)
775 0{-no arity-} d{-bitmap size-} bitmap True{-is alts-}
777 -- trace ("case: bndr = " ++ showSDocDebug (ppr bndr) ++ "\ndepth = " ++ show d ++ "\nenv = \n" ++ showSDocDebug (ppBCEnv p) ++
778 -- "\n bitmap = " ++ show bitmap) $ do
779 scrut_code <- schemeE (d + ret_frame_sizeW) (d + ret_frame_sizeW) p scrut
780 alt_bco' <- emitBc alt_bco
782 | isAlgCase = PUSH_ALTS alt_bco'
783 | otherwise = PUSH_ALTS_UNLIFTED alt_bco' (typePrimRep bndr_ty)
784 returnBc (push_alts `consOL` scrut_code)
787 -- -----------------------------------------------------------------------------
788 -- Deal with a CCall.
790 -- Taggedly push the args onto the stack R->L,
791 -- deferencing ForeignObj#s and (ToDo: adjusting addrs to point to
792 -- payloads in Ptr/Byte arrays). Then, generate the marshalling
793 -- (machine) code for the ccall, and create bytecodes to call that and
794 -- then return in the right way.
796 generateCCall :: Int -> Sequel -- stack and sequel depths
798 -> CCallSpec -- where to call
799 -> Id -- of target, for type info
800 -> [AnnExpr' Id VarSet] -- args (atoms)
803 generateCCall d0 s p ccall_spec@(CCallSpec target cconv safety) fn args_r_to_l
806 addr_sizeW = getPrimRepSize AddrRep
808 -- Get the args on the stack, with tags and suitably
809 -- dereferenced for the CCall. For each arg, return the
810 -- depth to the first word of the bits for that arg, and the
811 -- PrimRep of what was actually pushed.
813 pargs d [] = returnBc []
815 = let arg_ty = repType (exprType (deAnnotate' a))
817 in case splitTyConApp_maybe arg_ty of
818 -- Don't push the FO; instead push the Addr# it
821 | t == arrayPrimTyCon || t == mutableArrayPrimTyCon
822 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
823 parg_ArrayishRep arrPtrsHdrSize d p a
825 returnBc ((code,AddrRep):rest)
827 | t == byteArrayPrimTyCon || t == mutableByteArrayPrimTyCon
828 -> pargs (d + addr_sizeW) az `thenBc` \ rest ->
829 parg_ArrayishRep arrWordsHdrSize d p a
831 returnBc ((code,AddrRep):rest)
833 -- Default case: push taggedly, but otherwise intact.
835 -> pushAtom d p a `thenBc` \ (code_a, sz_a) ->
836 pargs (d+sz_a) az `thenBc` \ rest ->
837 returnBc ((code_a, atomRep a) : rest)
839 -- Do magic for Ptr/Byte arrays. Push a ptr to the array on
840 -- the stack but then advance it over the headers, so as to
841 -- point to the payload.
842 parg_ArrayishRep hdrSizeW d p a
843 = pushAtom d p a `thenBc` \ (push_fo, _) ->
844 -- The ptr points at the header. Advance it over the
845 -- header and then pretend this is an Addr#.
846 returnBc (push_fo `snocOL`
847 SWIZZLE 0 (hdrSizeW * getPrimRepSize WordRep
851 pargs d0 args_r_to_l `thenBc` \ code_n_reps ->
853 (pushs_arg, a_reps_pushed_r_to_l) = unzip code_n_reps
855 push_args = concatOL pushs_arg
856 d_after_args = d0 + sum (map getPrimRepSize a_reps_pushed_r_to_l)
858 | null a_reps_pushed_r_to_l || head a_reps_pushed_r_to_l /= VoidRep
859 = panic "ByteCodeGen.generateCCall: missing or invalid World token?"
861 = reverse (tail a_reps_pushed_r_to_l)
863 -- Now: a_reps_pushed_RAW are the reps which are actually on the stack.
864 -- push_args is the code to do that.
865 -- d_after_args is the stack depth once the args are on.
867 -- Get the result rep.
868 (returns_void, r_rep)
869 = case maybe_getCCallReturnRep (idType fn) of
870 Nothing -> (True, VoidRep)
871 Just rr -> (False, rr)
873 Because the Haskell stack grows down, the a_reps refer to
874 lowest to highest addresses in that order. The args for the call
875 are on the stack. Now push an unboxed Addr# indicating
876 the C function to call. Then push a dummy placeholder for the
877 result. Finally, emit a CCALL insn with an offset pointing to the
878 Addr# just pushed, and a literal field holding the mallocville
879 address of the piece of marshalling code we generate.
880 So, just prior to the CCALL insn, the stack looks like this
881 (growing down, as usual):
886 Addr# address_of_C_fn
887 <placeholder-for-result#> (must be an unboxed type)
889 The interpreter then calls the marshall code mentioned
890 in the CCALL insn, passing it (& <placeholder-for-result#>),
891 that is, the addr of the topmost word in the stack.
892 When this returns, the placeholder will have been
893 filled in. The placeholder is slid down to the sequel
894 depth, and we RETURN.
896 This arrangement makes it simple to do f-i-dynamic since the Addr#
897 value is the first arg anyway.
899 The marshalling code is generated specifically for this
900 call site, and so knows exactly the (Haskell) stack
901 offsets of the args, fn address and placeholder. It
902 copies the args to the C stack, calls the stacked addr,
903 and parks the result back in the placeholder. The interpreter
904 calls it as a normal C call, assuming it has a signature
905 void marshall_code ( StgWord* ptr_to_top_of_stack )
907 -- resolve static address
911 -> returnBc (False, panic "ByteCodeGen.generateCCall(dyn)")
913 -> ioToBc (lookupStaticPtr target) `thenBc` \res ->
916 -> pprPanic "ByteCodeGen.generateCCall: casm" (ppr ccall_spec)
918 get_target_info `thenBc` \ (is_static, static_target_addr) ->
921 -- Get the arg reps, zapping the leading Addr# in the dynamic case
922 a_reps -- | trace (showSDoc (ppr a_reps_pushed_RAW)) False = error "???"
923 | is_static = a_reps_pushed_RAW
924 | otherwise = if null a_reps_pushed_RAW
925 then panic "ByteCodeGen.generateCCall: dyn with no args"
926 else tail a_reps_pushed_RAW
929 (push_Addr, d_after_Addr)
931 = (toOL [PUSH_UBX (Right static_target_addr) addr_sizeW],
932 d_after_args + addr_sizeW)
933 | otherwise -- is already on the stack
934 = (nilOL, d_after_args)
936 -- Push the return placeholder. For a call returning nothing,
937 -- this is a VoidRep (tag).
938 r_sizeW = getPrimRepSize r_rep
939 d_after_r = d_after_Addr + r_sizeW
940 r_lit = mkDummyLiteral r_rep
941 push_r = (if returns_void
943 else unitOL (PUSH_UBX (Left r_lit) r_sizeW))
945 -- generate the marshalling code we're going to call
948 arg1_offW = r_sizeW + addr_sizeW
949 args_offW = map (arg1_offW +)
950 (init (scanl (+) 0 (map getPrimRepSize a_reps)))
952 ioToBc (mkMarshalCode cconv
953 (r_offW, r_rep) addr_offW
954 (zip args_offW a_reps)) `thenBc` \ addr_of_marshaller ->
955 recordMallocBc addr_of_marshaller `thenBc_`
957 -- Offset of the next stack frame down the stack. The CCALL
958 -- instruction needs to describe the chunk of stack containing
959 -- the ccall args to the GC, so it needs to know how large it
960 -- is. See comment in Interpreter.c with the CCALL instruction.
961 stk_offset = d_after_r - s
964 do_call = unitOL (CCALL stk_offset (castPtr addr_of_marshaller))
966 wrapup = mkSLIDE r_sizeW (d_after_r - r_sizeW - s)
967 `snocOL` RETURN_UBX r_rep
969 --trace (show (arg1_offW, args_offW , (map getPrimRepSize a_reps) )) $
972 push_Addr `appOL` push_r `appOL` do_call `appOL` wrapup
976 -- Make a dummy literal, to be used as a placeholder for FFI return
977 -- values on the stack.
978 mkDummyLiteral :: PrimRep -> Literal
981 CharRep -> MachChar 0
983 WordRep -> MachWord 0
984 DoubleRep -> MachDouble 0
985 FloatRep -> MachFloat 0
986 AddrRep | getPrimRepSize AddrRep == getPrimRepSize WordRep -> MachWord 0
987 _ -> moan64 "mkDummyLiteral" (ppr pr)
991 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
992 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)
995 -- and check that an unboxed pair is returned wherein the first arg is VoidRep'd.
997 -- Alternatively, for call-targets returning nothing, convert
999 -- GHC.Prim.Char# -> GHC.Prim.State# GHC.Prim.RealWorld
1000 -- -> (# GHC.Prim.State# GHC.Prim.RealWorld #)
1004 maybe_getCCallReturnRep :: Type -> Maybe PrimRep
1005 maybe_getCCallReturnRep fn_ty
1006 = let (a_tys, r_ty) = splitFunTys (dropForAlls fn_ty)
1008 = if isSingleton r_reps then Nothing else Just (r_reps !! 1)
1010 = case splitTyConApp_maybe (repType r_ty) of
1011 (Just (tyc, tys)) -> (tyc, map typePrimRep tys)
1013 ok = ( ( r_reps `lengthIs` 2 && VoidRep == head r_reps)
1014 || r_reps == [VoidRep] )
1015 && isUnboxedTupleTyCon r_tycon
1016 && case maybe_r_rep_to_go of
1018 Just r_rep -> r_rep /= PtrRep
1019 -- if it was, it would be impossible
1020 -- to create a valid return value
1021 -- placeholder on the stack
1022 blargh = pprPanic "maybe_getCCallReturn: can't handle:"
1025 --trace (showSDoc (ppr (a_reps, r_reps))) $
1026 if ok then maybe_r_rep_to_go else blargh
1028 -- Compile code which expects an unboxed Int on the top of stack,
1029 -- (call it i), and pushes the i'th closure in the supplied list
1030 -- as a consequence.
1031 implement_tagToId :: [Name] -> BcM BCInstrList
1032 implement_tagToId names
1033 = ASSERT( notNull names )
1034 getLabelsBc (length names) `thenBc` \ labels ->
1035 getLabelBc `thenBc` \ label_fail ->
1036 getLabelBc `thenBc` \ label_exit ->
1037 zip4 labels (tail labels ++ [label_fail])
1038 [0 ..] names `bind` \ infos ->
1039 map (mkStep label_exit) infos `bind` \ steps ->
1040 returnBc (concatOL steps
1042 toOL [LABEL label_fail, CASEFAIL, LABEL label_exit])
1044 mkStep l_exit (my_label, next_label, n, name_for_n)
1045 = toOL [LABEL my_label,
1046 TESTEQ_I n next_label,
1051 -- -----------------------------------------------------------------------------
1054 -- Push an atom onto the stack, returning suitable code & number of
1055 -- stack words used.
1057 -- The env p must map each variable to the highest- numbered stack
1058 -- slot for it. For example, if the stack has depth 4 and we
1059 -- tagged-ly push (v :: Int#) on it, the value will be in stack[4],
1060 -- the tag in stack[5], the stack will have depth 6, and p must map v
1061 -- to 5 and not to 4. Stack locations are numbered from zero, so a
1062 -- depth 6 stack has valid words 0 .. 5.
1064 pushAtom :: Int -> BCEnv -> AnnExpr' Id VarSet -> BcM (BCInstrList, Int)
1066 pushAtom d p (AnnApp f (_, AnnType _))
1067 = pushAtom d p (snd f)
1069 pushAtom d p (AnnNote note e)
1070 = pushAtom d p (snd e)
1072 pushAtom d p (AnnLam x e)
1074 = pushAtom d p (snd e)
1076 pushAtom d p (AnnVar v)
1078 | idPrimRep v == VoidRep
1079 = returnBc (nilOL, 0)
1082 = pprPanic "pushAtom: shouldn't get an FCallId here" (ppr v)
1084 | Just primop <- isPrimOpId_maybe v
1085 = returnBc (unitOL (PUSH_PRIMOP primop), 1)
1087 | Just d_v <- lookupBCEnv_maybe p v -- v is a local variable
1088 = returnBc (toOL (nOfThem sz (PUSH_L (d-d_v+sz-2))), sz)
1089 -- d - d_v the number of words between the TOS
1090 -- and the 1st slot of the object
1092 -- d - d_v - 1 the offset from the TOS of the 1st slot
1094 -- d - d_v - 1 + sz - 1 the offset from the TOS of the last slot
1097 -- Having found the last slot, we proceed to copy the right number of
1098 -- slots on to the top of the stack.
1100 | otherwise -- v must be a global variable
1102 returnBc (unitOL (PUSH_G (getName v)), sz)
1108 pushAtom d p (AnnLit lit)
1110 MachLabel fs -> code CodePtrRep
1111 MachWord w -> code WordRep
1112 MachInt i -> code IntRep
1113 MachFloat r -> code FloatRep
1114 MachDouble r -> code DoubleRep
1115 MachChar c -> code CharRep
1116 MachStr s -> pushStr s
1119 = let size_host_words = getPrimRepSize rep
1120 in returnBc (unitOL (PUSH_UBX (Left lit) size_host_words),
1124 = let getMallocvilleAddr
1126 FastString _ l ba ->
1127 -- sigh, a string in the heap is no good to us.
1128 -- We need a static C pointer, since the type of
1129 -- a string literal is Addr#. So, copy the string
1130 -- into C land and remember the pointer so we can
1133 -- CAREFUL! Chars are 32 bits in ghc 4.09+
1134 in ioToBc (mallocBytes (n+1)) `thenBc` \ ptr ->
1135 recordMallocBc ptr `thenBc_`
1137 do memcpy ptr ba (fromIntegral n)
1138 pokeByteOff ptr n (fromIntegral (ord '\0') :: Word8)
1141 other -> panic "ByteCodeGen.pushAtom.pushStr"
1143 getMallocvilleAddr `thenBc` \ addr ->
1144 -- Get the addr on the stack, untaggedly
1145 returnBc (unitOL (PUSH_UBX (Right addr) 1), 1)
1148 = pprPanic "ByteCodeGen.pushAtom"
1149 (pprCoreExpr (deAnnotate (undefined, other)))
1151 foreign import ccall unsafe "memcpy"
1152 memcpy :: Ptr a -> ByteArray# -> CInt -> IO ()
1155 -- -----------------------------------------------------------------------------
1156 -- Given a bunch of alts code and their discrs, do the donkey work
1157 -- of making a multiway branch using a switch tree.
1158 -- What a load of hassle!
1160 mkMultiBranch :: Maybe Int -- # datacons in tycon, if alg alt
1161 -- a hint; generates better code
1162 -- Nothing is always safe
1163 -> [(Discr, BCInstrList)]
1165 mkMultiBranch maybe_ncons raw_ways
1166 = let d_way = filter (isNoDiscr.fst) raw_ways
1167 notd_ways = naturalMergeSortLe
1168 (\w1 w2 -> leAlt (fst w1) (fst w2))
1169 (filter (not.isNoDiscr.fst) raw_ways)
1171 mkTree :: [(Discr, BCInstrList)] -> Discr -> Discr -> BcM BCInstrList
1172 mkTree [] range_lo range_hi = returnBc the_default
1174 mkTree [val] range_lo range_hi
1175 | range_lo `eqAlt` range_hi
1176 = returnBc (snd val)
1178 = getLabelBc `thenBc` \ label_neq ->
1179 returnBc (mkTestEQ (fst val) label_neq
1181 `appOL` unitOL (LABEL label_neq)
1182 `appOL` the_default))
1184 mkTree vals range_lo range_hi
1185 = let n = length vals `div` 2
1186 vals_lo = take n vals
1187 vals_hi = drop n vals
1188 v_mid = fst (head vals_hi)
1190 getLabelBc `thenBc` \ label_geq ->
1191 mkTree vals_lo range_lo (dec v_mid) `thenBc` \ code_lo ->
1192 mkTree vals_hi v_mid range_hi `thenBc` \ code_hi ->
1193 returnBc (mkTestLT v_mid label_geq
1195 `appOL` unitOL (LABEL label_geq)
1199 = case d_way of [] -> unitOL CASEFAIL
1202 -- None of these will be needed if there are no non-default alts
1203 (mkTestLT, mkTestEQ, init_lo, init_hi)
1205 = panic "mkMultiBranch: awesome foursome"
1207 = case fst (head notd_ways) of {
1208 DiscrI _ -> ( \(DiscrI i) fail_label -> TESTLT_I i fail_label,
1209 \(DiscrI i) fail_label -> TESTEQ_I i fail_label,
1212 DiscrF _ -> ( \(DiscrF f) fail_label -> TESTLT_F f fail_label,
1213 \(DiscrF f) fail_label -> TESTEQ_F f fail_label,
1216 DiscrD _ -> ( \(DiscrD d) fail_label -> TESTLT_D d fail_label,
1217 \(DiscrD d) fail_label -> TESTEQ_D d fail_label,
1220 DiscrP _ -> ( \(DiscrP i) fail_label -> TESTLT_P i fail_label,
1221 \(DiscrP i) fail_label -> TESTEQ_P i fail_label,
1223 DiscrP algMaxBound )
1226 (algMinBound, algMaxBound)
1227 = case maybe_ncons of
1228 Just n -> (0, n - 1)
1229 Nothing -> (minBound, maxBound)
1231 (DiscrI i1) `eqAlt` (DiscrI i2) = i1 == i2
1232 (DiscrF f1) `eqAlt` (DiscrF f2) = f1 == f2
1233 (DiscrD d1) `eqAlt` (DiscrD d2) = d1 == d2
1234 (DiscrP i1) `eqAlt` (DiscrP i2) = i1 == i2
1235 NoDiscr `eqAlt` NoDiscr = True
1238 (DiscrI i1) `leAlt` (DiscrI i2) = i1 <= i2
1239 (DiscrF f1) `leAlt` (DiscrF f2) = f1 <= f2
1240 (DiscrD d1) `leAlt` (DiscrD d2) = d1 <= d2
1241 (DiscrP i1) `leAlt` (DiscrP i2) = i1 <= i2
1242 NoDiscr `leAlt` NoDiscr = True
1245 isNoDiscr NoDiscr = True
1248 dec (DiscrI i) = DiscrI (i-1)
1249 dec (DiscrP i) = DiscrP (i-1)
1250 dec other = other -- not really right, but if you
1251 -- do cases on floating values, you'll get what you deserve
1253 -- same snotty comment applies to the following
1255 minD, maxD :: Double
1261 mkTree notd_ways init_lo init_hi
1264 -- -----------------------------------------------------------------------------
1265 -- Supporting junk for the compilation schemes
1267 -- Describes case alts
1275 instance Outputable Discr where
1276 ppr (DiscrI i) = int i
1277 ppr (DiscrF f) = text (show f)
1278 ppr (DiscrD d) = text (show d)
1279 ppr (DiscrP i) = int i
1280 ppr NoDiscr = text "DEF"
1283 lookupBCEnv_maybe :: BCEnv -> Id -> Maybe Int
1284 lookupBCEnv_maybe = lookupFM
1286 idSizeW :: Id -> Int
1287 idSizeW id = getPrimRepSize (typePrimRep (idType id))
1289 unboxedTupleException :: a
1290 unboxedTupleException
1293 ("Bytecode generator can't handle unboxed tuples. Possibly due\n" ++
1294 "\tto foreign import/export decls in source. Workaround:\n" ++
1295 "\tcompile this module to a .o file, then restart session."))
1298 mkSLIDE n d = if d == 0 then nilOL else unitOL (SLIDE n d)
1301 splitApp :: AnnExpr' id ann -> (AnnExpr' id ann, [AnnExpr' id ann])
1302 -- The arguments are returned in *right-to-left* order
1303 splitApp (AnnApp (_,f) (_,a))
1304 | isTypeAtom a = splitApp f
1305 | otherwise = case splitApp f of
1306 (f', as) -> (f', a:as)
1307 splitApp (AnnNote n (_,e)) = splitApp e
1308 splitApp e = (e, [])
1311 isTypeAtom :: AnnExpr' id ann -> Bool
1312 isTypeAtom (AnnType _) = True
1313 isTypeAtom _ = False
1315 isVoidRepAtom :: AnnExpr' id ann -> Bool
1316 isVoidRepAtom (AnnVar v) = typePrimRep (idType v) == VoidRep
1317 isVoidRepAtom (AnnNote n (_,e)) = isVoidRepAtom e
1318 isVoidRepAtom _ = False
1320 atomRep :: AnnExpr' Id ann -> PrimRep
1321 atomRep (AnnVar v) = typePrimRep (idType v)
1322 atomRep (AnnLit l) = literalPrimRep l
1323 atomRep (AnnNote n b) = atomRep (snd b)
1324 atomRep (AnnApp f (_, AnnType _)) = atomRep (snd f)
1325 atomRep (AnnLam x e) | isTyVar x = atomRep (snd e)
1326 atomRep other = pprPanic "atomRep" (ppr (deAnnotate (undefined,other)))
1328 isPtrAtom :: AnnExpr' Id ann -> Bool
1329 isPtrAtom e = isFollowableRep (atomRep e)
1331 -- Let szsw be the sizes in words of some items pushed onto the stack,
1332 -- which has initial depth d'. Return the values which the stack environment
1333 -- should map these items to.
1334 mkStackOffsets :: Int -> [Int] -> [Int]
1335 mkStackOffsets original_depth szsw
1336 = map (subtract 1) (tail (scanl (+) original_depth szsw))
1338 -- -----------------------------------------------------------------------------
1339 -- The bytecode generator's monad
1343 nextlabel :: Int, -- for generating local labels
1344 malloced :: [Ptr ()] } -- ptrs malloced for current BCO
1345 -- Should be free()d when it is GCd
1347 newtype BcM r = BcM (BcM_State -> IO (BcM_State, r))
1349 ioToBc :: IO a -> BcM a
1350 ioToBc io = BcM $ \st -> do
1354 runBc :: BcM r -> IO (BcM_State, r)
1355 runBc (BcM m) = m (BcM_State 0 [])
1357 thenBc :: BcM a -> (a -> BcM b) -> BcM b
1358 thenBc (BcM expr) cont = BcM $ \st0 -> do
1359 (st1, q) <- expr st0
1364 thenBc_ :: BcM a -> BcM b -> BcM b
1365 thenBc_ (BcM expr) (BcM cont) = BcM $ \st0 -> do
1366 (st1, q) <- expr st0
1367 (st2, r) <- cont st1
1370 returnBc :: a -> BcM a
1371 returnBc result = BcM $ \st -> (return (st, result))
1373 instance Monad BcM where
1378 emitBc :: ([Ptr ()] -> ProtoBCO Name) -> BcM (ProtoBCO Name)
1380 = BcM $ \st -> return (st{malloced=[]}, bco (malloced st))
1382 recordMallocBc :: Ptr a -> BcM ()
1384 = BcM $ \st -> return (st{malloced = castPtr a : malloced st}, ())
1386 getLabelBc :: BcM Int
1388 = BcM $ \st -> return (st{nextlabel = 1 + nextlabel st}, nextlabel st)
1390 getLabelsBc :: Int -> BcM [Int]
1392 = BcM $ \st -> let ctr = nextlabel st
1393 in return (st{nextlabel = ctr+n}, [ctr .. ctr+n-1])