----------------------------------------------------------------------------- -- -- Code generation for foreign calls. -- -- (c) The University of Glasgow 2004-2006 -- ----------------------------------------------------------------------------- module GHC.StgToCmm.Foreign ( cgForeignCall, emitPrimCall, emitCCall, emitForeignCall, -- For CmmParse emitSaveThreadState, saveThreadState, emitLoadThreadState, loadThreadState, emitOpenNursery, emitCloseNursery, ) where import GhcPrelude hiding( succ, (<*>) ) import StgSyn import GHC.StgToCmm.Prof (storeCurCCS, ccsType) import GHC.StgToCmm.Env import GHC.StgToCmm.Monad import GHC.StgToCmm.Utils import GHC.StgToCmm.Closure import GHC.StgToCmm.Layout import BlockId (newBlockId) import Cmm import CmmUtils import MkGraph import Type import RepType import CLabel import SMRep import ForeignCall import DynFlags import Maybes import Outputable import UniqSupply import BasicTypes import TyCoRep import TysPrim import Util (zipEqual) import Control.Monad ----------------------------------------------------------------------------- -- Code generation for Foreign Calls ----------------------------------------------------------------------------- -- | Emit code for a foreign call, and return the results to the sequel. -- Precondition: the length of the arguments list is the same as the -- arity of the foreign function. cgForeignCall :: ForeignCall -- the op -> Type -- type of foreign function -> [StgArg] -- x,y arguments -> Type -- result type -> FCode ReturnKind cgForeignCall (CCall (CCallSpec target cconv safety)) typ stg_args res_ty = do { dflags <- getDynFlags ; let -- in the stdcall calling convention, the symbol needs @size appended -- to it, where size is the total number of bytes of arguments. We -- attach this info to the CLabel here, and the CLabel pretty printer -- will generate the suffix when the label is printed. call_size args | StdCallConv <- cconv = Just (sum (map arg_size args)) | otherwise = Nothing -- ToDo: this might not be correct for 64-bit API arg_size (arg, _) = max (widthInBytes $ typeWidth $ cmmExprType dflags arg) (wORD_SIZE dflags) ; cmm_args <- getFCallArgs stg_args typ ; (res_regs, res_hints) <- newUnboxedTupleRegs res_ty ; let ((call_args, arg_hints), cmm_target) = case target of StaticTarget _ _ _ False -> panic "cgForeignCall: unexpected FFI value import" StaticTarget _ lbl mPkgId True -> let labelSource = case mPkgId of Nothing -> ForeignLabelInThisPackage Just pkgId -> ForeignLabelInPackage pkgId size = call_size cmm_args in ( unzip cmm_args , CmmLit (CmmLabel (mkForeignLabel lbl size labelSource IsFunction))) DynamicTarget -> case cmm_args of (fn,_):rest -> (unzip rest, fn) [] -> panic "cgForeignCall []" fc = ForeignConvention cconv arg_hints res_hints CmmMayReturn call_target = ForeignTarget cmm_target fc -- we want to emit code for the call, and then emitReturn. -- However, if the sequel is AssignTo, we shortcut a little -- and generate a foreign call that assigns the results -- directly. Otherwise we end up generating a bunch of -- useless "r = r" assignments, which are not merely annoying: -- they prevent the common block elimination from working correctly -- in the case of a safe foreign call. -- See Note [safe foreign call convention] -- ; sequel <- getSequel ; case sequel of AssignTo assign_to_these _ -> emitForeignCall safety assign_to_these call_target call_args _something_else -> do { _ <- emitForeignCall safety res_regs call_target call_args ; emitReturn (map (CmmReg . CmmLocal) res_regs) } } {- Note [safe foreign call convention] The simple thing to do for a safe foreign call would be the same as an unsafe one: just emitForeignCall ... emitReturn ... but consider what happens in this case case foo x y z of (# s, r #) -> ... The sequel is AssignTo [r]. The call to newUnboxedTupleRegs picks [r] as the result reg, and we generate r = foo(x,y,z) returns to L1 -- emitForeignCall L1: r = r -- emitReturn goto L2 L2: ... Now L1 is a proc point (by definition, it is the continuation of the safe foreign call). If L2 does a heap check, then L2 will also be a proc point. Furthermore, the stack layout algorithm has to arrange to save r somewhere between the call and the jump to L1, which is annoying: we would have to treat r differently from the other live variables, which have to be saved *before* the call. So we adopt a special convention for safe foreign calls: the results are copied out according to the NativeReturn convention by the call, and the continuation of the call should copyIn the results. (The copyOut code is actually inserted when the safe foreign call is lowered later). The result regs attached to the safe foreign call are only used temporarily to hold the results before they are copied out. We will now generate this: r = foo(x,y,z) returns to L1 L1: r = R1 -- copyIn, inserted by mkSafeCall goto L2 L2: ... r ... And when the safe foreign call is lowered later (see Note [lower safe foreign calls]) we get this: suspendThread() r = foo(x,y,z) resumeThread() R1 = r -- copyOut, inserted by lowerSafeForeignCall jump L1 L1: r = R1 -- copyIn, inserted by mkSafeCall goto L2 L2: ... r ... Now consider what happens if L2 does a heap check: the Adams optimisation kicks in and commons up L1 with the heap-check continuation, resulting in just one proc point instead of two. Yay! -} emitCCall :: [(CmmFormal,ForeignHint)] -> CmmExpr -> [(CmmActual,ForeignHint)] -> FCode () emitCCall hinted_results fn hinted_args = void $ emitForeignCall PlayRisky results target args where (args, arg_hints) = unzip hinted_args (results, result_hints) = unzip hinted_results target = ForeignTarget fn fc fc = ForeignConvention CCallConv arg_hints result_hints CmmMayReturn emitPrimCall :: [CmmFormal] -> CallishMachOp -> [CmmActual] -> FCode () emitPrimCall res op args = void $ emitForeignCall PlayRisky res (PrimTarget op) args -- alternative entry point, used by CmmParse emitForeignCall :: Safety -> [CmmFormal] -- where to put the results -> ForeignTarget -- the op -> [CmmActual] -- arguments -> FCode ReturnKind emitForeignCall safety results target args | not (playSafe safety) = do dflags <- getDynFlags let (caller_save, caller_load) = callerSaveVolatileRegs dflags emit caller_save target' <- load_target_into_temp target args' <- mapM maybe_assign_temp args emit $ mkUnsafeCall target' results args' emit caller_load return AssignedDirectly | otherwise = do dflags <- getDynFlags updfr_off <- getUpdFrameOff target' <- load_target_into_temp target args' <- mapM maybe_assign_temp args k <- newBlockId let (off, _, copyout) = copyInOflow dflags NativeReturn (Young k) results [] -- see Note [safe foreign call convention] tscope <- getTickScope emit $ ( mkStore (CmmStackSlot (Young k) (widthInBytes (wordWidth dflags))) (CmmLit (CmmBlock k)) <*> mkLast (CmmForeignCall { tgt = target' , res = results , args = args' , succ = k , ret_args = off , ret_off = updfr_off , intrbl = playInterruptible safety }) <*> mkLabel k tscope <*> copyout ) return (ReturnedTo k off) load_target_into_temp :: ForeignTarget -> FCode ForeignTarget load_target_into_temp (ForeignTarget expr conv) = do tmp <- maybe_assign_temp expr return (ForeignTarget tmp conv) load_target_into_temp other_target@(PrimTarget _) = return other_target -- What we want to do here is create a new temporary for the foreign -- call argument if it is not safe to use the expression directly, -- because the expression mentions caller-saves GlobalRegs (see -- Note [Register Parameter Passing]). -- -- However, we can't pattern-match on the expression here, because -- this is used in a loop by CmmParse, and testing the expression -- results in a black hole. So we always create a temporary, and rely -- on CmmSink to clean it up later. (Yuck, ToDo). The generated code -- ends up being the same, at least for the RTS .cmm code. -- maybe_assign_temp :: CmmExpr -> FCode CmmExpr maybe_assign_temp e = do dflags <- getDynFlags reg <- newTemp (cmmExprType dflags e) emitAssign (CmmLocal reg) e return (CmmReg (CmmLocal reg)) -- ----------------------------------------------------------------------------- -- Save/restore the thread state in the TSO -- This stuff can't be done in suspendThread/resumeThread, because it -- refers to global registers which aren't available in the C world. emitSaveThreadState :: FCode () emitSaveThreadState = do dflags <- getDynFlags code <- saveThreadState dflags emit code -- | Produce code to save the current thread state to @CurrentTSO@ saveThreadState :: MonadUnique m => DynFlags -> m CmmAGraph saveThreadState dflags = do tso <- newTemp (gcWord dflags) close_nursery <- closeNursery dflags tso pure $ catAGraphs [ -- tso = CurrentTSO; mkAssign (CmmLocal tso) currentTSOExpr, -- tso->stackobj->sp = Sp; mkStore (cmmOffset dflags (CmmLoad (cmmOffset dflags (CmmReg (CmmLocal tso)) (tso_stackobj dflags)) (bWord dflags)) (stack_SP dflags)) spExpr, close_nursery, -- and save the current cost centre stack in the TSO when profiling: if gopt Opt_SccProfilingOn dflags then mkStore (cmmOffset dflags (CmmReg (CmmLocal tso)) (tso_CCCS dflags)) cccsExpr else mkNop ] emitCloseNursery :: FCode () emitCloseNursery = do dflags <- getDynFlags tso <- newTemp (bWord dflags) code <- closeNursery dflags tso emit $ mkAssign (CmmLocal tso) currentTSOExpr <*> code {- | @closeNursery dflags tso@ produces code to close the nursery. A local register holding the value of @CurrentTSO@ is expected for efficiency. Closing the nursery corresponds to the following code: @ tso = CurrentTSO; cn = CurrentNuresry; // Update the allocation limit for the current thread. We don't // check to see whether it has overflowed at this point, that check is // made when we run out of space in the current heap block (stg_gc_noregs) // and in the scheduler when context switching (schedulePostRunThread). tso->alloc_limit -= Hp + WDS(1) - cn->start; // Set cn->free to the next unoccupied word in the block cn->free = Hp + WDS(1); @ -} closeNursery :: MonadUnique m => DynFlags -> LocalReg -> m CmmAGraph closeNursery df tso = do let tsoreg = CmmLocal tso cnreg <- CmmLocal <$> newTemp (bWord df) pure $ catAGraphs [ mkAssign cnreg currentNurseryExpr, -- CurrentNursery->free = Hp+1; mkStore (nursery_bdescr_free df cnreg) (cmmOffsetW df hpExpr 1), let alloc = CmmMachOp (mo_wordSub df) [ cmmOffsetW df hpExpr 1 , CmmLoad (nursery_bdescr_start df cnreg) (bWord df) ] alloc_limit = cmmOffset df (CmmReg tsoreg) (tso_alloc_limit df) in -- tso->alloc_limit += alloc mkStore alloc_limit (CmmMachOp (MO_Sub W64) [ CmmLoad alloc_limit b64 , CmmMachOp (mo_WordTo64 df) [alloc] ]) ] emitLoadThreadState :: FCode () emitLoadThreadState = do dflags <- getDynFlags code <- loadThreadState dflags emit code -- | Produce code to load the current thread state from @CurrentTSO@ loadThreadState :: MonadUnique m => DynFlags -> m CmmAGraph loadThreadState dflags = do tso <- newTemp (gcWord dflags) stack <- newTemp (gcWord dflags) open_nursery <- openNursery dflags tso pure $ catAGraphs [ -- tso = CurrentTSO; mkAssign (CmmLocal tso) currentTSOExpr, -- stack = tso->stackobj; mkAssign (CmmLocal stack) (CmmLoad (cmmOffset dflags (CmmReg (CmmLocal tso)) (tso_stackobj dflags)) (bWord dflags)), -- Sp = stack->sp; mkAssign spReg (CmmLoad (cmmOffset dflags (CmmReg (CmmLocal stack)) (stack_SP dflags)) (bWord dflags)), -- SpLim = stack->stack + RESERVED_STACK_WORDS; mkAssign spLimReg (cmmOffsetW dflags (cmmOffset dflags (CmmReg (CmmLocal stack)) (stack_STACK dflags)) (rESERVED_STACK_WORDS dflags)), -- HpAlloc = 0; -- HpAlloc is assumed to be set to non-zero only by a failed -- a heap check, see HeapStackCheck.cmm:GC_GENERIC mkAssign hpAllocReg (zeroExpr dflags), open_nursery, -- and load the current cost centre stack from the TSO when profiling: if gopt Opt_SccProfilingOn dflags then storeCurCCS (CmmLoad (cmmOffset dflags (CmmReg (CmmLocal tso)) (tso_CCCS dflags)) (ccsType dflags)) else mkNop ] emitOpenNursery :: FCode () emitOpenNursery = do dflags <- getDynFlags tso <- newTemp (bWord dflags) code <- openNursery dflags tso emit $ mkAssign (CmmLocal tso) currentTSOExpr <*> code {- | @openNursery dflags tso@ produces code to open the nursery. A local register holding the value of @CurrentTSO@ is expected for efficiency. Opening the nursery corresponds to the following code: @ tso = CurrentTSO; cn = CurrentNursery; bdfree = CurrentNursery->free; bdstart = CurrentNursery->start; // We *add* the currently occupied portion of the nursery block to // the allocation limit, because we will subtract it again in // closeNursery. tso->alloc_limit += bdfree - bdstart; // Set Hp to the last occupied word of the heap block. Why not the // next unocupied word? Doing it this way means that we get to use // an offset of zero more often, which might lead to slightly smaller // code on some architectures. Hp = bdfree - WDS(1); // Set HpLim to the end of the current nursery block (note that this block // might be a block group, consisting of several adjacent blocks. HpLim = bdstart + CurrentNursery->blocks*BLOCK_SIZE_W - 1; @ -} openNursery :: MonadUnique m => DynFlags -> LocalReg -> m CmmAGraph openNursery df tso = do let tsoreg = CmmLocal tso cnreg <- CmmLocal <$> newTemp (bWord df) bdfreereg <- CmmLocal <$> newTemp (bWord df) bdstartreg <- CmmLocal <$> newTemp (bWord df) -- These assignments are carefully ordered to reduce register -- pressure and generate not completely awful code on x86. To see -- what code we generate, look at the assembly for -- stg_returnToStackTop in rts/StgStartup.cmm. pure $ catAGraphs [ mkAssign cnreg currentNurseryExpr, mkAssign bdfreereg (CmmLoad (nursery_bdescr_free df cnreg) (bWord df)), -- Hp = CurrentNursery->free - 1; mkAssign hpReg (cmmOffsetW df (CmmReg bdfreereg) (-1)), mkAssign bdstartreg (CmmLoad (nursery_bdescr_start df cnreg) (bWord df)), -- HpLim = CurrentNursery->start + -- CurrentNursery->blocks*BLOCK_SIZE_W - 1; mkAssign hpLimReg (cmmOffsetExpr df (CmmReg bdstartreg) (cmmOffset df (CmmMachOp (mo_wordMul df) [ CmmMachOp (MO_SS_Conv W32 (wordWidth df)) [CmmLoad (nursery_bdescr_blocks df cnreg) b32], mkIntExpr df (bLOCK_SIZE df) ]) (-1) ) ), -- alloc = bd->free - bd->start let alloc = CmmMachOp (mo_wordSub df) [CmmReg bdfreereg, CmmReg bdstartreg] alloc_limit = cmmOffset df (CmmReg tsoreg) (tso_alloc_limit df) in -- tso->alloc_limit += alloc mkStore alloc_limit (CmmMachOp (MO_Add W64) [ CmmLoad alloc_limit b64 , CmmMachOp (mo_WordTo64 df) [alloc] ]) ] nursery_bdescr_free, nursery_bdescr_start, nursery_bdescr_blocks :: DynFlags -> CmmReg -> CmmExpr nursery_bdescr_free dflags cn = cmmOffset dflags (CmmReg cn) (oFFSET_bdescr_free dflags) nursery_bdescr_start dflags cn = cmmOffset dflags (CmmReg cn) (oFFSET_bdescr_start dflags) nursery_bdescr_blocks dflags cn = cmmOffset dflags (CmmReg cn) (oFFSET_bdescr_blocks dflags) tso_stackobj, tso_CCCS, tso_alloc_limit, stack_STACK, stack_SP :: DynFlags -> ByteOff tso_stackobj dflags = closureField dflags (oFFSET_StgTSO_stackobj dflags) tso_alloc_limit dflags = closureField dflags (oFFSET_StgTSO_alloc_limit dflags) tso_CCCS dflags = closureField dflags (oFFSET_StgTSO_cccs dflags) stack_STACK dflags = closureField dflags (oFFSET_StgStack_stack dflags) stack_SP dflags = closureField dflags (oFFSET_StgStack_sp dflags) closureField :: DynFlags -> ByteOff -> ByteOff closureField dflags off = off + fixedHdrSize dflags -- Note [Unlifted boxed arguments to foreign calls] -- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -- -- For certain types passed to foreign calls, we adjust the actual -- value passed to the call. For ByteArray#, Array#, SmallArray#, -- and ArrayArray#, we pass the address of the array's payload, not -- the address of the heap object. For example, consider -- foreign import "c_foo" foo :: ByteArray# -> Int# -> IO () -- At a Haskell call like `foo x y`, we'll generate a C call that -- is more like -- c_foo( x+8, y ) -- where the "+8" takes the heap pointer (x :: ByteArray#) and moves -- it past the header words of the ByteArray object to point directly -- to the data inside the ByteArray#. (The exact offset depends -- on the target architecture and on profiling) By contrast, (y :: Int#) -- requires no such adjustment. -- -- This adjustment is performed by 'add_shim'. The size of the -- adjustment depends on the type of heap object. But -- how can we determine that type? There are two available options. -- We could use the types of the actual values that the foreign call -- has been applied to, or we could use the types present in the -- foreign function's type. Prior to GHC 8.10, we used the former -- strategy since it's a little more simple. However, in issue #16650 -- and more compellingly in the comments of -- https://gitlab.haskell.org/ghc/ghc/merge_requests/939, it was -- demonstrated that this leads to bad behavior in the presence -- of unsafeCoerce#. Returning to the above example, suppose the -- Haskell call looked like -- foo (unsafeCoerce# p) -- where the types of expressions comprising the arguments are -- p :: (Any :: TYPE 'UnliftedRep) -- i :: Int# -- so that the unsafe-coerce is between Any and ByteArray#. -- These two types have the same kind (they are both represented by -- a heap pointer) so no GC errors will occur if we do this unsafe coerce. -- By the time this gets to the code generator the cast has been -- discarded so we have -- foo p y -- But we *must* adjust the pointer to p by a ByteArray# shim, -- *not* by an Any shim (the Any shim involves no offset at all). -- -- To avoid this bad behavior, we adopt the second strategy: use -- the types present in the foreign function's type. -- In collectStgFArgTypes, we convert the foreign function's -- type to a list of StgFArgType. Then, in add_shim, we interpret -- these as numeric offsets. getFCallArgs :: [StgArg] -> Type -- the type of the foreign function -> FCode [(CmmExpr, ForeignHint)] -- (a) Drop void args -- (b) Add foreign-call shim code -- It's (b) that makes this differ from getNonVoidArgAmodes -- Precondition: args and typs have the same length -- See Note [Unlifted boxed arguments to foreign calls] getFCallArgs args typ = do { mb_cmms <- mapM get (zipEqual "getFCallArgs" args (collectStgFArgTypes typ)) ; return (catMaybes mb_cmms) } where get (arg,typ) | null arg_reps = return Nothing | otherwise = do { cmm <- getArgAmode (NonVoid arg) ; dflags <- getDynFlags ; return (Just (add_shim dflags typ cmm, hint)) } where arg_ty = stgArgType arg arg_reps = typePrimRep arg_ty hint = typeForeignHint arg_ty -- The minimum amount of information needed to determine -- the offset to apply to an argument to a foreign call. -- See Note [Unlifted boxed arguments to foreign calls] data StgFArgType = StgPlainType | StgArrayType | StgSmallArrayType | StgByteArrayType -- See Note [Unlifted boxed arguments to foreign calls] add_shim :: DynFlags -> StgFArgType -> CmmExpr -> CmmExpr add_shim dflags ty expr = case ty of StgPlainType -> expr StgArrayType -> cmmOffsetB dflags expr (arrPtrsHdrSize dflags) StgSmallArrayType -> cmmOffsetB dflags expr (smallArrPtrsHdrSize dflags) StgByteArrayType -> cmmOffsetB dflags expr (arrWordsHdrSize dflags) -- From a function, extract information needed to determine -- the offset of each argument when used as a C FFI argument. -- See Note [Unlifted boxed arguments to foreign calls] collectStgFArgTypes :: Type -> [StgFArgType] collectStgFArgTypes = go [] where -- Skip foralls go bs (ForAllTy _ res) = go bs res go bs (AppTy{}) = reverse bs go bs (TyConApp{}) = reverse bs go bs (LitTy{}) = reverse bs go bs (TyVarTy{}) = reverse bs go _ (CastTy{}) = panic "myCollectTypeArgs: CastTy" go _ (CoercionTy{}) = panic "myCollectTypeArgs: CoercionTy" go bs (FunTy {ft_arg = arg, ft_res=res}) = go (typeToStgFArgType arg:bs) res -- Choose the offset based on the type. For anything other -- than an unlifted boxed type, there is no offset. -- See Note [Unlifted boxed arguments to foreign calls] typeToStgFArgType :: Type -> StgFArgType typeToStgFArgType typ | tycon == arrayPrimTyCon = StgArrayType | tycon == mutableArrayPrimTyCon = StgArrayType | tycon == arrayArrayPrimTyCon = StgArrayType | tycon == mutableArrayArrayPrimTyCon = StgArrayType | tycon == smallArrayPrimTyCon = StgSmallArrayType | tycon == smallMutableArrayPrimTyCon = StgSmallArrayType | tycon == byteArrayPrimTyCon = StgByteArrayType | tycon == mutableByteArrayPrimTyCon = StgByteArrayType | otherwise = StgPlainType where -- Should be a tycon app, since this is a foreign call. We look -- through newtypes so the offset does not change if a user replaces -- a type in a foreign function signature with a representationally -- equivalent newtype. tycon = tyConAppTyCon (unwrapType typ)