X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FcodeGen%2FCgTailCall.lhs;h=dd7327b745f1d6ad41fb89414b049da1cb25c301;hb=04db0e9fa47ce4dfbcb73ec1752d94195f3b394e;hp=c2ece1ee2c6a2ae4fde799945ba9fc16426daa3b;hpb=10521d8418fd3a1cf32882718b5bd28992db36fd;p=ghc-hetmet.git diff --git a/ghc/compiler/codeGen/CgTailCall.lhs b/ghc/compiler/codeGen/CgTailCall.lhs index c2ece1e..dd7327b 100644 --- a/ghc/compiler/codeGen/CgTailCall.lhs +++ b/ghc/compiler/codeGen/CgTailCall.lhs @@ -1,5 +1,7 @@ % -% (c) The GRASP/AQUA Project, Glasgow University, 1992-1995 +% (c) The GRASP/AQUA Project, Glasgow University, 1992-1998 +% +% $Id: CgTailCall.lhs,v 1.43 2005/06/21 10:44:41 simonmar Exp $ % %******************************************************** %* * @@ -8,538 +10,446 @@ %******************************************************** \begin{code} -#include "HsVersions.h" - module CgTailCall ( - cgTailCall, - performReturn, - mkStaticAlgReturnCode, mkDynamicAlgReturnCode, - mkPrimReturnCode, - - tailCallBusiness, - - -- and to make the interface self-sufficient... - StgAtom, Id, CgState, CAddrMode, TyCon, - CgInfoDownwards, HeapOffset, Maybe + cgTailCall, performTailCall, + performReturn, performPrimReturn, + emitKnownConReturnCode, emitAlgReturnCode, + returnUnboxedTuple, ccallReturnUnboxedTuple, + pushUnboxedTuple, + tailCallPrimOp, + + pushReturnAddress ) where -IMPORT_Trace -import Pretty -- Pretty/Outputable: rm (debugging only) ToDo -import Outputable +#include "HsVersions.h" -import StgSyn import CgMonad -import AbsCSyn - -import AbsUniType ( isPrimType, UniType ) -import CgBindery ( getAtomAmodes, getCAddrMode, getCAddrModeAndInfo ) -import CgCompInfo ( oTHER_TAG, iND_TAG ) -import CgRetConv ( dataReturnConvPrim, ctrlReturnConvAlg, dataReturnConvAlg, - mkLiveRegsBitMask, - CtrlReturnConvention(..), DataReturnConvention(..) - ) -import CgStackery ( adjustRealSps, mkStkAmodes ) -import CgUsages ( getSpARelOffset, getSpBRelOffset ) -import CLabelInfo ( CLabel, mkStdUpdCodePtrVecLabel, mkConUpdCodePtrVecLabel ) -import ClosureInfo ( nodeMustPointToIt, getEntryConvention, EntryConvention(..) ) -import CmdLineOpts ( GlobalSwitch(..) ) -import Id ( getDataConTyCon, getDataConTag, - getIdUniType, getIdKind, fIRST_TAG, Id, - ConTag(..) - ) -import Maybes ( assocMaybe, maybeToBool, Maybe(..) ) -import PrimKind ( retKindSize ) -import Util -\end{code} - -%************************************************************************ -%* * -\subsection[tailcall-doc]{Documentation} -%* * -%************************************************************************ - -\begin{code} -cgTailCall :: PlainStgAtom -> [PlainStgAtom] -> PlainStgLiveVars -> Code -\end{code} - -Here's the code we generate for a tail call. (NB there may be no -arguments, in which case this boils down to just entering a variable.) - -\begin{itemize} -\item Adjust the stack ptr to \tr{tailSp + #args}. -\item Put args in the top locations of the resulting stack. -\item Make Node point to the function closure. -\item Enter the function closure. -\end{itemize} - -Things to be careful about: -\begin{itemize} -\item Don't overwrite stack locations before you have finished with - them (remember you need the function and the as-yet-unmoved - arguments). -\item Preferably, generate no code to replace x by x on the stack (a - common situation in tail-recursion). -\item Adjust the stack high water mark appropriately. -\end{itemize} - -Literals are similar to constructors; they return by putting -themselves in an appropriate register and returning to the address on -top of the B stack. - -\begin{code} -cgTailCall (StgLitAtom lit) [] live_vars - = performPrimReturn (CLit lit) live_vars -\end{code} - -Treat unboxed locals exactly like literals (above) except use the addr -mode for the local instead of (CLit lit) in the assignment. - -Case for unboxed @Ids@ first: -\begin{code} -cgTailCall atom@(StgVarAtom fun) [] live_vars - | isPrimType (getIdUniType fun) - = getCAddrMode fun `thenFC` \ amode -> - performPrimReturn amode live_vars -\end{code} - -The general case (@fun@ is boxed): -\begin{code} -cgTailCall (StgVarAtom fun) args live_vars = performTailCall fun args live_vars -\end{code} - -%************************************************************************ -%* * -\subsection[return-and-tail-call]{Return and tail call} -%* * -%************************************************************************ - -ADR-HACK - - A quick bit of hacking to try to solve my void#-leaking blues... - - I think I'm getting bitten by this stuff because code like - - \begin{pseudocode} - case ds.s12 :: IoWorld of { - -- lvs: [ds.s12]; rhs lvs: []; uniq: c0 - IoWorld ds.s13# -> ds.s13#; - } :: Universe# - \end{pseudocode} - - causes me to try to allocate a register to return the result in. The - hope is that the following will avoid such problems (and that Will - will do this in a cleaner way when he hits the same problem). - -KCAH-RDA - -\begin{code} -performPrimReturn :: CAddrMode -- The thing to return - -> PlainStgLiveVars - -> Code +import CgBindery ( getArgAmodes, getCgIdInfo, CgIdInfo, maybeLetNoEscape, + idInfoToAmode, cgIdInfoId, cgIdInfoLF, + cgIdInfoArgRep ) +import CgInfoTbls ( entryCode, emitDirectReturnInstr, dataConTagZ, + emitVectoredReturnInstr, closureInfoPtr ) +import CgCallConv +import CgStackery ( setRealSp, mkStkAmodes, adjustStackHW, + getSpRelOffset ) +import CgHeapery ( setRealHp, getHpRelOffset ) +import CgUtils ( emitSimultaneously ) +import CgTicky +import ClosureInfo +import SMRep ( CgRep, isVoidArg, separateByPtrFollowness ) +import Cmm +import CmmUtils +import CLabel ( CLabel, mkRtsPrimOpLabel, mkSeqInfoLabel ) +import Type ( isUnLiftedType ) +import Id ( Id, idName, idUnique, idType ) +import DataCon ( DataCon, dataConTyCon ) +import StgSyn ( StgArg ) +import TyCon ( TyCon ) +import PrimOp ( PrimOp ) +import Outputable -performPrimReturn amode live_vars - = let - kind = getAmodeKind amode - ret_reg = dataReturnConvPrim kind - - assign_possibly = case kind of - VoidKind -> AbsCNop - kind -> (CAssign (CReg ret_reg) amode) - in - performReturn assign_possibly mkPrimReturnCode live_vars - -mkPrimReturnCode :: Sequel -> Code ---UNUSED:mkPrimReturnCode RestoreCostCentre = panic "mkPrimReturnCode: RCC" -mkPrimReturnCode (UpdateCode _) = panic "mkPrimReturnCode: Upd" -mkPrimReturnCode sequel = sequelToAmode sequel `thenFC` \ dest_amode -> - absC (CReturn dest_amode DirectReturn) - -- Direct, no vectoring - --- All constructor arguments in registers; Node and InfoPtr are set. --- All that remains is --- (a) to set TagReg, if necessary --- (b) to set InfoPtr to the info ptr, if necessary --- (c) to do the right sort of jump. - -mkStaticAlgReturnCode :: Id -- The constructor - -> Maybe CLabel -- The info ptr, if it isn't already set - -> Sequel -- where to return to - -> Code - -mkStaticAlgReturnCode con maybe_info_lbl sequel - = -- Generate profiling code if necessary - (case return_convention of - VectoredReturn sz -> profCtrC SLIT("VEC_RETURN") [mkIntCLit sz] - other -> nopC - ) `thenC` - - -- Set tag if necessary - -- This is done by a macro, because if we are short of registers - -- we don't set TagReg; instead the continuation gets the tag - -- by indexing off the info ptr - (case return_convention of - - UnvectoredReturn no_of_constrs - | no_of_constrs > 1 - -> absC (CMacroStmt SET_TAG [mkIntCLit zero_indexed_tag]) - - other -> nopC - ) `thenC` - - -- Generate the right jump or return - (case sequel of - UpdateCode _ -> -- Ha! We know the constructor, - -- so we can go direct to the correct - -- update code for that constructor - - -- Set the info pointer, and jump - set_info_ptr `thenC` - getIntSwitchChkrC `thenFC` \ isw_chkr -> - absC (CJump (CLbl (update_label isw_chkr) CodePtrKind)) - - CaseAlts _ (Just (alts, _)) -> -- Ho! We know the constructor so - -- we can go right to the alternative - - -- No need to set info ptr when returning to a - -- known join point. After all, the code at - -- the destination knows what constructor it - -- is going to handle. - - case assocMaybe alts tag of - Just (alt_absC, join_lbl) -> absC (CJump (CLbl join_lbl CodePtrKind)) - Nothing -> panic "mkStaticAlgReturnCode: default" - -- The Nothing case should never happen; it's the subject - -- of a wad of special-case code in cgReturnCon - - other -> -- OnStack, or (CaseAlts) ret_amode Nothing) - -- Set the info pointer, and jump - set_info_ptr `thenC` - sequelToAmode sequel `thenFC` \ ret_amode -> - absC (CReturn ret_amode return_info) - ) +import Monad ( when ) + +----------------------------------------------------------------------------- +-- Tail Calls + +cgTailCall :: Id -> [StgArg] -> Code + +-- Here's the code we generate for a tail call. (NB there may be no +-- arguments, in which case this boils down to just entering a variable.) +-- +-- * Put args in the top locations of the stack. +-- * Adjust the stack ptr +-- * Make R1 point to the function closure if necessary. +-- * Perform the call. +-- +-- Things to be careful about: +-- +-- * Don't overwrite stack locations before you have finished with +-- them (remember you need the function and the as-yet-unmoved +-- arguments). +-- * Preferably, generate no code to replace x by x on the stack (a +-- common situation in tail-recursion). +-- * Adjust the stack high water mark appropriately. +-- +-- Treat unboxed locals exactly like literals (above) except use the addr +-- mode for the local instead of (CLit lit) in the assignment. + +cgTailCall fun args + = do { fun_info <- getCgIdInfo fun + + ; if isUnLiftedType (idType fun) + then -- Primitive return + ASSERT( null args ) + do { fun_amode <- idInfoToAmode fun_info + ; performPrimReturn (cgIdInfoArgRep fun_info) fun_amode } + + else -- Normal case, fun is boxed + do { arg_amodes <- getArgAmodes args + ; performTailCall fun_info arg_amodes noStmts } + } + +-- ----------------------------------------------------------------------------- +-- The guts of a tail-call + +performTailCall + :: CgIdInfo -- The function + -> [(CgRep,CmmExpr)] -- Args + -> CmmStmts -- Pending simultaneous assignments + -- *** GUARANTEED to contain only stack assignments. + -> Code + +performTailCall fun_info arg_amodes pending_assts + | Just join_sp <- maybeLetNoEscape fun_info + = -- A let-no-escape is slightly different, because we + -- arrange the stack arguments into pointers and non-pointers + -- to make the heap check easier. The tail-call sequence + -- is very similar to returning an unboxed tuple, so we + -- share some code. + do { (final_sp, arg_assts) <- pushUnboxedTuple join_sp arg_amodes + ; emitSimultaneously (pending_assts `plusStmts` arg_assts) + ; let lbl = enterReturnPtLabel (idUnique (cgIdInfoId fun_info)) + ; doFinalJump final_sp True {- Is LNE -} (jumpToLbl lbl) } + + | otherwise + = do { fun_amode <- idInfoToAmode fun_info + ; let node_asst = oneStmt (CmmAssign nodeReg fun_amode) + opt_node_asst | nodeMustPointToIt lf_info = node_asst + | otherwise = noStmts + ; EndOfBlockInfo sp _ <- getEndOfBlockInfo + ; hmods <- getHomeModules + + ; case (getCallMethod hmods fun_name lf_info (length arg_amodes)) of + + -- Node must always point to things we enter + EnterIt -> do + { emitSimultaneously (node_asst `plusStmts` pending_assts) + ; let target = entryCode (closureInfoPtr (CmmReg nodeReg)) + ; doFinalJump sp False (stmtC (CmmJump target [])) } + + -- A function, but we have zero arguments. It is already in WHNF, + -- so we can just return it. + -- As with any return, Node must point to it. + ReturnIt -> do + { emitSimultaneously (node_asst `plusStmts` pending_assts) + ; doFinalJump sp False emitDirectReturnInstr } + + -- A real constructor. Don't bother entering it, + -- just do the right sort of return instead. + -- As with any return, Node must point to it. + ReturnCon con -> do + { emitSimultaneously (node_asst `plusStmts` pending_assts) + ; doFinalJump sp False (emitKnownConReturnCode con) } + + JumpToIt lbl -> do + { emitSimultaneously (opt_node_asst `plusStmts` pending_assts) + ; doFinalJump sp False (jumpToLbl lbl) } + + -- A slow function call via the RTS apply routines + -- Node must definitely point to the thing + SlowCall -> do + { when (not (null arg_amodes)) $ do + { if (isKnownFun lf_info) + then tickyKnownCallTooFewArgs + else tickyUnknownCall + ; tickySlowCallPat (map fst arg_amodes) + } + + ; let (apply_lbl, args, extra_args) + = constructSlowCall arg_amodes + + ; directCall sp apply_lbl args extra_args + (node_asst `plusStmts` pending_assts) + } + + -- A direct function call (possibly with some left-over arguments) + DirectEntry lbl arity -> do + { if arity == length arg_amodes + then tickyKnownCallExact + else do tickyKnownCallExtraArgs + tickySlowCallPat (map fst (drop arity arg_amodes)) + + ; let + -- The args beyond the arity go straight on the stack + (arity_args, extra_args) = splitAt arity arg_amodes + + ; directCall sp lbl arity_args extra_args + (opt_node_asst `plusStmts` pending_assts) + } + } where - tag = getDataConTag con - tycon = getDataConTyCon con - return_convention = ctrlReturnConvAlg tycon - zero_indexed_tag = tag - fIRST_TAG -- Adjust tag to be zero-indexed - -- cf AbsCFuns.mkAlgAltsCSwitch + fun_name = idName (cgIdInfoId fun_info) + lf_info = cgIdInfoLF fun_info - update_label isw_chkr - = case (dataReturnConvAlg isw_chkr con) of - ReturnInHeap -> mkStdUpdCodePtrVecLabel tycon tag - ReturnInRegs _ -> mkConUpdCodePtrVecLabel tycon tag - return_info = case return_convention of - UnvectoredReturn _ -> DirectReturn - VectoredReturn _ -> StaticVectoredReturn zero_indexed_tag - set_info_ptr = case maybe_info_lbl of - Nothing -> nopC - Just info_lbl -> absC (CAssign (CReg infoptr) (CLbl info_lbl DataPtrKind)) +directCall sp lbl args extra_args assts = do + let + -- First chunk of args go in registers + (reg_arg_amodes, stk_args) = assignCallRegs args + + -- Any "extra" arguments are placed in frames on the + -- stack after the other arguments. + slow_stk_args = slowArgs extra_args + reg_assts = assignToRegs reg_arg_amodes + -- + (final_sp, stk_assts) <- mkStkAmodes sp (stk_args ++ slow_stk_args) -mkDynamicAlgReturnCode :: TyCon -> CAddrMode -> Sequel -> Code + emitSimultaneously (reg_assts `plusStmts` + stk_assts `plusStmts` + assts) -mkDynamicAlgReturnCode tycon dyn_tag sequel - = case ctrlReturnConvAlg tycon of - VectoredReturn sz -> + doFinalJump final_sp False (jumpToLbl lbl) - profCtrC SLIT("VEC_RETURN") [mkIntCLit sz] `thenC` - sequelToAmode sequel `thenFC` \ ret_addr -> - absC (CReturn ret_addr (DynamicVectoredReturn dyn_tag)) +-- ----------------------------------------------------------------------------- +-- The final clean-up before we do a jump at the end of a basic block. +-- This code is shared by tail-calls and returns. - UnvectoredReturn no_of_constrs -> - - -- Set tag if necessary - -- This is done by a macro, because if we are short of registers - -- we don't set TagReg; instead the continuation gets the tag - -- by indexing off the info ptr - (if no_of_constrs > 1 then - absC (CMacroStmt SET_TAG [dyn_tag]) - else - nopC - ) `thenC` +doFinalJump :: VirtualSpOffset -> Bool -> Code -> Code +doFinalJump final_sp is_let_no_escape jump_code + = do { -- Adjust the high-water mark if necessary + adjustStackHW final_sp + -- Push a return address if necessary (after the assignments + -- above, in case we clobber a live stack location) + -- + -- DONT push the return address when we're about to jump to a + -- let-no-escape: the final tail call in the let-no-escape + -- will do this. + ; eob <- getEndOfBlockInfo + ; whenC (not is_let_no_escape) (pushReturnAddress eob) - sequelToAmode sequel `thenFC` \ ret_addr -> - -- Generate the right jump or return - absC (CReturn ret_addr DirectReturn) -\end{code} + -- Final adjustment of Sp/Hp + ; adjustSpAndHp final_sp -\begin{code} -performReturn :: AbstractC -- Simultaneous assignments to perform - -> (Sequel -> Code) -- The code to execute to actually do - -- the return, given an addressing mode - -- for the return address - -> PlainStgLiveVars - -> Code + -- and do the jump + ; jump_code } -performReturn sim_assts finish_code live_vars - = getEndOfBlockInfo `thenFC` \ (EndOfBlockInfo args_spa args_spb sequel) -> +-- ----------------------------------------------------------------------------- +-- A general return (just a special case of doFinalJump, above) - -- Do the simultaneous assignments, - doSimAssts args_spa live_vars {-UNUSED:live_regs-} sim_assts `thenC` +performReturn :: Code -- The code to execute to actually do the return + -> Code - -- Adjust stack pointers - adjustRealSps args_spa args_spb `thenC` +performReturn finish_code + = do { EndOfBlockInfo args_sp sequel <- getEndOfBlockInfo + ; doFinalJump args_sp False{-not a LNE-} finish_code } - -- Do the return - finish_code sequel -- "sequel" is `robust' in that it doesn't - -- depend on stk-ptr values --- where ---UNUSED: live_regs = getDestinationRegs sim_assts - -- ToDo: this is a *really* boring way to compute the - -- live-reg set! -\end{code} +-- ----------------------------------------------------------------------------- +-- Primitive Returns +-- Just load the return value into the right register, and return. -\begin{code} -performTailCall :: Id -- Function - -> [PlainStgAtom] -- Args - -> PlainStgLiveVars - -> Code - -performTailCall fun args live_vars - = -- Get all the info we have about the function and args and go on to - -- the business end - getCAddrModeAndInfo fun `thenFC` \ (fun_amode, lf_info) -> - getAtomAmodes args `thenFC` \ arg_amodes -> - - tailCallBusiness - fun fun_amode lf_info arg_amodes - live_vars AbsCNop {- No pending assignments -} - - -tailCallBusiness :: Id -> CAddrMode -- Function and its amode - -> LambdaFormInfo -- Info about the function - -> [CAddrMode] -- Arguments - -> PlainStgLiveVars -- Live in continuation - - -> AbstractC -- Pending simultaneous assignments - -- *** GUARANTEED to contain only stack assignments. - -- In ptic, we don't need to look in here to - -- discover all live regs - - -> Code - -tailCallBusiness fun fun_amode lf_info arg_amodes live_vars pending_assts - = isSwitchSetC EmitArityChecks `thenFC` \ do_arity_chks -> - - nodeMustPointToIt lf_info `thenFC` \ node_points -> - getEntryConvention fun lf_info - (map getAmodeKind arg_amodes) `thenFC` \ entry_conv -> - - getEndOfBlockInfo `thenFC` \ (EndOfBlockInfo args_spa args_spb sequel) -> - - let - node_asst - = if node_points then - CAssign (CReg node) fun_amode - else - AbsCNop - - (arg_regs, finish_code) - = case entry_conv of - ViaNode -> - ([], - mkAbstractCs [ - CCallProfCtrMacro SLIT("ENT_VIA_NODE") [], - CAssign (CReg infoptr) - - (CMacroExpr DataPtrKind INFO_PTR [CReg node]), - CJump (CMacroExpr CodePtrKind ENTRY_CODE [CReg infoptr]) - ]) - StdEntry lbl Nothing -> ([], CJump (CLbl lbl CodePtrKind)) - StdEntry lbl (Just itbl) -> ([], CAssign (CReg infoptr) (CLbl itbl DataPtrKind) - `mkAbsCStmts` - CJump (CLbl lbl CodePtrKind)) - DirectEntry lbl arity regs -> - (regs, (if do_arity_chks - then CMacroStmt SET_ARITY [mkIntCLit arity] - else AbsCNop) - `mkAbsCStmts` CJump (CLbl lbl CodePtrKind)) - - no_of_args = length arg_amodes - -{- UNUSED: live_regs = if node_points then - node : arg_regs - else - arg_regs --} - (reg_arg_assts, stk_arg_amodes) - = (mkAbstractCs (zipWith assign_to_reg arg_regs arg_amodes), - drop (length arg_regs) arg_amodes) -- No regs, or - -- args beyond arity - - assign_to_reg reg_id amode = CAssign (CReg reg_id) amode - - in - case fun_amode of - CJoinPoint join_spa join_spb -> -- Ha! A let-no-escape thingy - - ASSERT(not (args_spa > join_spa) || (args_spb > join_spb)) - -- If ASSERTion fails: Oops: the join point has *lower* - -- stack ptrs than the continuation Note that we take - -- the SpB point without the return address here. The - -- return address is put on by the let-no-escapey thing - -- when it finishes. - - mkStkAmodes join_spa join_spb stk_arg_amodes - `thenFC` \ (final_spa, final_spb, stk_arg_assts) -> - - -- Do the simultaneous assignments, - doSimAssts join_spa live_vars {-UNUSED: live_regs-} - (mkAbstractCs [pending_assts, reg_arg_assts, stk_arg_assts]) - `thenC` - - -- Adjust stack ptrs - adjustRealSps final_spa final_spb `thenC` - - -- Jump to join point - absC finish_code - - _ -> -- else: not a let-no-escape (the common case) - - -- Make instruction to save return address - loadRetAddrIntoRetReg sequel `thenFC` \ ret_asst -> +performPrimReturn :: CgRep -> CmmExpr -- The thing to return + -> Code +performPrimReturn rep amode + = do { whenC (not (isVoidArg rep)) + (stmtC (CmmAssign ret_reg amode)) + ; performReturn emitDirectReturnInstr } + where + ret_reg = dataReturnConvPrim rep + +-- ----------------------------------------------------------------------------- +-- Algebraic constructor returns + +-- Constructor is built on the heap; Node is set. +-- All that remains is to do the right sort of jump. + +emitKnownConReturnCode :: DataCon -> Code +emitKnownConReturnCode con + = emitAlgReturnCode (dataConTyCon con) + (CmmLit (mkIntCLit (dataConTagZ con))) + -- emitAlgReturnCode requires zero-indexed tag + +emitAlgReturnCode :: TyCon -> CmmExpr -> Code +-- emitAlgReturnCode is used both by emitKnownConReturnCode, +-- and by by PrimOps that return enumerated types (i.e. +-- all the comparison operators). +emitAlgReturnCode tycon tag + = do { case ctrlReturnConvAlg tycon of + VectoredReturn fam_sz -> do { tickyVectoredReturn fam_sz + ; emitVectoredReturnInstr tag } + UnvectoredReturn _ -> emitDirectReturnInstr + } + + +-- --------------------------------------------------------------------------- +-- Unboxed tuple returns + +-- These are a bit like a normal tail call, except that: +-- +-- - The tail-call target is an info table on the stack +-- +-- - We separate stack arguments into pointers and non-pointers, +-- to make it easier to leave things in a sane state for a heap check. +-- This is OK because we can never partially-apply an unboxed tuple, +-- unlike a function. The same technique is used when calling +-- let-no-escape functions, because they also can't be partially +-- applied. + +returnUnboxedTuple :: [(CgRep, CmmExpr)] -> Code +returnUnboxedTuple amodes + = do { eob@(EndOfBlockInfo args_sp sequel) <- getEndOfBlockInfo + ; tickyUnboxedTupleReturn (length amodes) + ; (final_sp, assts) <- pushUnboxedTuple args_sp amodes + ; emitSimultaneously assts + ; doFinalJump final_sp False{-not a LNE-} emitDirectReturnInstr } + +pushUnboxedTuple :: VirtualSpOffset -- Sp at which to start pushing + -> [(CgRep, CmmExpr)] -- amodes of the components + -> FCode (VirtualSpOffset, -- final Sp + CmmStmts) -- assignments (regs+stack) + +pushUnboxedTuple sp [] + = return (sp, noStmts) +pushUnboxedTuple sp amodes + = do { let (reg_arg_amodes, stk_arg_amodes) = assignReturnRegs amodes + + -- separate the rest of the args into pointers and non-pointers + (ptr_args, nptr_args) = separateByPtrFollowness stk_arg_amodes + reg_arg_assts = assignToRegs reg_arg_amodes - mkStkAmodes args_spa args_spb stk_arg_amodes - `thenFC` - \ (final_spa, final_spb, stk_arg_assts) -> - - -- The B-stack space for the pushed return addess, with any args pushed - -- on top, is recorded in final_spb. - - -- Do the simultaneous assignments, - doSimAssts args_spa live_vars {-UNUSED: live_regs-} - (mkAbstractCs [pending_assts, node_asst, ret_asst, - reg_arg_assts, stk_arg_assts]) - `thenC` - - -- Final adjustment of stack pointers - adjustRealSps final_spa final_spb `thenC` - - -- Now decide about semi-tagging - isSwitchSetC DoSemiTagging `thenFC` \ semi_tagging_on -> - case (semi_tagging_on, arg_amodes, node_points, sequel) of - - -- - -- *************** The semi-tagging case *************** - -- - ( True, [], True, CaseAlts _ (Just (st_alts, maybe_deflt_join_details))) -> - - -- Whoppee! Semi-tagging rules OK! - -- (a) semi-tagging is switched on - -- (b) there are no arguments, - -- (c) Node points to the closure - -- (d) we have a case-alternative sequel with - -- some visible alternatives - - -- Why is test (c) necessary? - -- Usually Node will point to it at this point, because we're - -- scrutinsing something which is either a thunk or a - -- constructor. - -- But not always! The example I came across is when we have - -- a top-level Double: - -- lit.3 = D# 3.000 - -- ... (case lit.3 of ...) ... - -- Here, lit.3 is built as a re-entrant thing, which you must enter. - -- (OK, the simplifier should have eliminated this, but it's - -- easy to deal with the case anyway.) - let - join_details_to_code (load_regs_and_profiling_code, join_lbl) - = load_regs_and_profiling_code `mkAbsCStmts` - CJump (CLbl join_lbl CodePtrKind) - - semi_tagged_alts = [ (mkMachInt (toInteger (tag - fIRST_TAG)), - join_details_to_code join_details) - | (tag, join_details) <- st_alts - ] - - enter_jump - -- Enter Node (we know infoptr will have the info ptr in it)! - = mkAbstractCs [ - CCallProfCtrMacro SLIT("RET_SEMI_FAILED") - [CMacroExpr IntKind INFO_TAG [CReg infoptr]], - CJump (CMacroExpr CodePtrKind ENTRY_CODE [CReg infoptr]) ] - in - -- Final switch - absC (mkAbstractCs [ - CAssign (CReg infoptr) - (CVal (NodeRel zeroOff) DataPtrKind), - - case maybe_deflt_join_details of - Nothing -> - CSwitch (CMacroExpr IntKind INFO_TAG [CReg infoptr]) - (semi_tagged_alts) - (enter_jump) - Just (_, details) -> - CSwitch (CMacroExpr IntKind EVAL_TAG [CReg infoptr]) - [(mkMachInt 0, enter_jump)] - (CSwitch - (CMacroExpr IntKind INFO_TAG [CReg infoptr]) - (semi_tagged_alts) - (join_details_to_code details)) - ]) - - -- - -- *************** The non-semi-tagging case *************** - -- - other -> absC finish_code + -- push ptrs, then nonptrs, on the stack + ; (ptr_sp, ptr_assts) <- mkStkAmodes sp ptr_args + ; (final_sp, nptr_assts) <- mkStkAmodes ptr_sp nptr_args + + ; returnFC (final_sp, + reg_arg_assts `plusStmts` + ptr_assts `plusStmts` nptr_assts) } + + +-- ----------------------------------------------------------------------------- +-- Returning unboxed tuples. This is mainly to support _ccall_GC_, where +-- we want to do things in a slightly different order to normal: +-- +-- - push return address +-- - adjust stack pointer +-- - r = call(args...) +-- - assign regs for unboxed tuple (usually just R1 = r) +-- - return to continuation +-- +-- The return address (i.e. stack frame) must be on the stack before +-- doing the call in case the call ends up in the garbage collector. +-- +-- Sadly, the information about the continuation is lost after we push it +-- (in order to avoid pushing it again), so we end up doing a needless +-- indirect jump (ToDo). + +ccallReturnUnboxedTuple :: [(CgRep, CmmExpr)] -> Code -> Code +ccallReturnUnboxedTuple amodes before_jump + = do { eob@(EndOfBlockInfo args_sp _) <- getEndOfBlockInfo + + -- Push a return address if necessary + ; pushReturnAddress eob + ; setEndOfBlockInfo (EndOfBlockInfo args_sp OnStack) + (do { adjustSpAndHp args_sp + ; before_jump + ; returnUnboxedTuple amodes }) + } + +-- ----------------------------------------------------------------------------- +-- Calling an out-of-line primop + +tailCallPrimOp :: PrimOp -> [StgArg] -> Code +tailCallPrimOp op args + = do { -- We're going to perform a normal-looking tail call, + -- except that *all* the arguments will be in registers. + -- Hence the ASSERT( null leftovers ) + arg_amodes <- getArgAmodes args + ; let (arg_regs, leftovers) = assignPrimOpCallRegs arg_amodes + jump_to_primop = jumpToLbl (mkRtsPrimOpLabel op) + + ; ASSERT(null leftovers) -- no stack-resident args + emitSimultaneously (assignToRegs arg_regs) + + ; EndOfBlockInfo args_sp _ <- getEndOfBlockInfo + ; doFinalJump args_sp False{-not a LNE-} jump_to_primop } + +-- ----------------------------------------------------------------------------- +-- Return Addresses + +-- We always push the return address just before performing a tail call +-- or return. The reason we leave it until then is because the stack +-- slot that the return address is to go into might contain something +-- useful. +-- +-- If the end of block info is 'CaseAlts', then we're in the scrutinee of a +-- case expression and the return address is still to be pushed. +-- +-- There are cases where it doesn't look necessary to push the return +-- address: for example, just before doing a return to a known +-- continuation. However, the continuation will expect to find the +-- return address on the stack in case it needs to do a heap check. + +pushReturnAddress :: EndOfBlockInfo -> Code + +pushReturnAddress (EndOfBlockInfo args_sp sequel@(CaseAlts lbl _ _ False)) + = do { sp_rel <- getSpRelOffset args_sp + ; stmtC (CmmStore sp_rel (mkLblExpr lbl)) } + +-- For a polymorphic case, we have two return addresses to push: the case +-- return, and stg_seq_frame_info which turns a possible vectored return +-- into a direct one. +pushReturnAddress (EndOfBlockInfo args_sp sequel@(CaseAlts lbl _ _ True)) + = do { sp_rel <- getSpRelOffset (args_sp-1) + ; stmtC (CmmStore sp_rel (mkLblExpr lbl)) + ; sp_rel <- getSpRelOffset args_sp + ; stmtC (CmmStore sp_rel (CmmLit (CmmLabel mkSeqInfoLabel))) } + +pushReturnAddress _ = nopC + +-- ----------------------------------------------------------------------------- +-- Misc. + +jumpToLbl :: CLabel -> Code +-- Passes no argument to the destination procedure +jumpToLbl lbl = stmtC (CmmJump (CmmLit (CmmLabel lbl)) [{- No args -}]) + +assignToRegs :: [(CmmExpr, GlobalReg)] -> CmmStmts +assignToRegs reg_args + = mkStmts [ CmmAssign (CmmGlobal reg_id) expr + | (expr, reg_id) <- reg_args ] \end{code} -\begin{code} -loadRetAddrIntoRetReg :: Sequel -> FCode AbstractC - -loadRetAddrIntoRetReg InRetReg - = returnFC AbsCNop -- Return address already there - -loadRetAddrIntoRetReg sequel - = sequelToAmode sequel `thenFC` \ amode -> - returnFC (CAssign (CReg RetReg) amode) - -\end{code} %************************************************************************ %* * -\subsection[doSimAssts]{@doSimAssts@} +\subsection[CgStackery-adjust]{Adjusting the stack pointers} %* * %************************************************************************ -@doSimAssts@ happens at the end of every block of code. -They are separate because we sometimes do some jiggery-pokery in between. +This function adjusts the stack and heap pointers just before a tail +call or return. The stack pointer is adjusted to its final position +(i.e. to point to the last argument for a tail call, or the activation +record for a return). The heap pointer may be moved backwards, in +cases where we overallocated at the beginning of the basic block (see +CgCase.lhs for discussion). + +These functions {\em do not} deal with high-water-mark adjustment. +That's done by functions which allocate stack space. \begin{code} -doSimAssts :: VirtualSpAOffset -- tail_spa: SpA as seen by continuation - -> PlainStgLiveVars -- Live in continuation ---UNUSED: -> [MagicId] -- Live regs (ptrs and non-ptrs) - -> AbstractC - -> Code - -doSimAssts tail_spa live_vars {-UNUSED: live_regs-} sim_assts - = -- Do the simultaneous assignments - absC (CSimultaneous sim_assts) `thenC` - - -- Stub any unstubbed slots; the only live variables are indicated in - -- the end-of-block info in the monad - nukeDeadBindings live_vars `thenC` - getUnstubbedAStackSlots tail_spa `thenFC` \ a_slots -> - -- Passing in tail_spa here should actually be redundant, because - -- the stack should be trimmed (by nukeDeadBindings) to - -- exactly the tail_spa position anyhow. - - -- Emit code to stub dead regs; this only generates actual - -- machine instructions in in the DEBUG version - -- *** NOT DONE YET *** - - (if (null a_slots) - then nopC - else profCtrC SLIT("A_STK_STUB") [mkIntCLit (length a_slots)] `thenC` - mapCs stub_A_slot a_slots - ) - where - stub_A_slot :: VirtualSpAOffset -> Code - stub_A_slot offset = getSpARelOffset offset `thenFC` \ spa_rel -> - absC (CAssign (CVal spa_rel PtrKind) - (CReg StkStubReg)) +adjustSpAndHp :: VirtualSpOffset -- New offset for Arg stack ptr + -> Code +adjustSpAndHp newRealSp + = do { -- Adjust stack, if necessary. + -- NB: the conditional on the monad-carried realSp + -- is out of line (via codeOnly), to avoid a black hole + ; new_sp <- getSpRelOffset newRealSp + ; checkedAbsC (CmmAssign spReg new_sp) -- Will generate no code in the case + ; setRealSp newRealSp -- where realSp==newRealSp + + -- Adjust heap. The virtual heap pointer may be less than the real Hp + -- because the latter was advanced to deal with the worst-case branch + -- of the code, and we may be in a better-case branch. In that case, + -- move the real Hp *back* and retract some ticky allocation count. + ; hp_usg <- getHpUsage + ; let rHp = realHp hp_usg + vHp = virtHp hp_usg + ; new_hp <- getHpRelOffset vHp + ; checkedAbsC (CmmAssign hpReg new_hp) -- Generates nothing when vHp==rHp + ; tickyAllocHeap (vHp - rHp) -- ...ditto + ; setRealHp vHp + } \end{code}