X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FcodeGen%2FCgCase.lhs;h=af053f6a9b512de5e35b5e4d2cf90f14d40b9cc6;hb=40df40956025fdb4d7d81e89f6c27755a9df8f46;hp=17d61261c1c1529dbb5c6d850361bca6ec3d28ff;hpb=e7498a3ee1d0484d02a9e86633cc179c76ebf36e;p=ghc-hetmet.git diff --git a/ghc/compiler/codeGen/CgCase.lhs b/ghc/compiler/codeGen/CgCase.lhs index 17d6126..af053f6 100644 --- a/ghc/compiler/codeGen/CgCase.lhs +++ b/ghc/compiler/codeGen/CgCase.lhs @@ -1,5 +1,7 @@ % -% (c) The GRASP/AQUA Project, Glasgow University, 1992-1996 +% (c) The GRASP/AQUA Project, Glasgow University, 1992-1998 +% +% $Id: CgCase.lhs,v 1.66 2003/07/22 16:11:26 simonmar Exp $ % %******************************************************** %* * @@ -8,67 +10,52 @@ %******************************************************** \begin{code} -#include "HsVersions.h" +module CgCase ( cgCase, saveVolatileVarsAndRegs, + mkRetDirectTarget, restoreCurrentCostCentre + ) where -module CgCase ( cgCase, saveVolatileVarsAndRegs ) where +#include "HsVersions.h" -IMP_Ubiq(){-uitous-} -IMPORT_DELOOPER(CgLoop2) ( cgExpr, getPrimOpArgAmodes ) +import {-# SOURCE #-} CgExpr ( cgExpr ) import CgMonad import StgSyn import AbsCSyn -import AbsCUtils ( mkAbstractCs, mkAbsCStmts, mkAlgAltsCSwitch, - magicIdPrimRep, getAmodeRep - ) -import CgBindery ( getVolatileRegs, getArgAmode, getArgAmodes, +import AbsCUtils ( mkAbstractCs, mkAbsCStmts, mkAlgAltsCSwitch, getAmodeRep ) +import CgBindery ( getVolatileRegs, getArgAmodes, bindNewToReg, bindNewToTemp, - bindNewPrimToAmode, - rebindToAStack, rebindToBStack, - getCAddrModeAndInfo, getCAddrModeIfVolatile, - idInfoToAmode - ) -import CgCon ( buildDynCon, bindConArgs ) -import CgHeapery ( heapCheck, yield ) -import CgRetConv ( dataReturnConvAlg, dataReturnConvPrim, - ctrlReturnConvAlg, - DataReturnConvention(..), CtrlReturnConvention(..), - assignPrimOpResultRegs, - makePrimOpArgsRobust - ) -import CgStackery ( allocAStack, allocBStack, allocAStackTop, allocBStackTop ) -import CgTailCall ( tailCallBusiness, performReturn ) -import CgUsages ( getSpARelOffset, getSpBRelOffset, freeBStkSlot ) -import CLabel ( mkVecTblLabel, mkReturnPtLabel, mkDefaultLabel, - mkAltLabel + getCAddrModeAndInfo, + rebindToStack, getCAddrMode, getCAddrModeIfVolatile, + buildContLivenessMask, nukeDeadBindings, ) -import ClosureInfo ( mkConLFInfo, mkLFArgument, layOutDynCon ) -import CmdLineOpts ( opt_SccProfilingOn, opt_GranMacros ) -import CostCentre ( useCurrentCostCentre ) -import HeapOffs ( VirtualSpBOffset(..), VirtualHeapOffset(..) ) -import Id ( idPrimRep, toplevelishId, - dataConTag, fIRST_TAG, ConTag(..), - isDataCon, DataCon(..), - idSetToList, GenId{-instance Uniquable,Eq-} +import CgCon ( bindConArgs, bindUnboxedTupleComponents ) +import CgHeapery ( altHeapCheck, unbxTupleHeapCheck ) +import CgRetConv ( dataReturnConvPrim, ctrlReturnConvAlg, + CtrlReturnConvention(..) ) -import Maybes ( catMaybes ) -import PprStyle ( PprStyle(..) ) -import PprType ( GenType{-instance Outputable-} ) -import PrimOp ( primOpCanTriggerGC, PrimOp(..), - primOpStackRequired, StackRequirement(..) +import CgStackery ( allocPrimStack, allocStackTop, + deAllocStackTop, freeStackSlots, dataStackSlots ) -import PrimRep ( getPrimRepSize, isFollowableRep, retPrimRepSize, - PrimRep(..) +import CgTailCall ( performTailCall ) +import CgUsages ( getSpRelOffset ) +import CLabel ( mkVecTblLabel, mkClosureTblLabel, + mkDefaultLabel, mkAltLabel, mkReturnInfoLabel ) -import TyCon ( isEnumerationTyCon ) -import Type ( typePrimRep, - getAppSpecDataTyConExpandingDicts, maybeAppSpecDataTyConExpandingDicts, - isEnumerationTyCon - ) -import Util ( sortLt, isIn, isn'tIn, zipEqual, - pprError, panic, assertPanic +import ClosureInfo ( mkLFArgument ) +import CmdLineOpts ( opt_SccProfilingOn ) +import Id ( Id, idName, isDeadBinder ) +import DataCon ( dataConTag, fIRST_TAG, ConTag ) +import VarSet ( varSetElems ) +import CoreSyn ( AltCon(..) ) +import PrimOp ( primOpOutOfLine, PrimOp(..) ) +import PrimRep ( getPrimRepSize, retPrimRepSize, PrimRep(..) ) +import TyCon ( TyCon, isEnumerationTyCon, tyConPrimRep ) +import Unique ( Unique, Uniquable(..), newTagUnique ) +import Util ( only ) +import List ( sortBy ) +import Outputable \end{code} \begin{code} @@ -88,13 +75,13 @@ op which can trigger GC. A more interesting situation is this: -\begin{verbatim} + \begin{verbatim} !A!; ...A... case x# of 0# -> !B!; ...B... default -> !C!; ...C... -\end{verbatim} + \end{verbatim} where \tr{!x!} indicates a possible heap-check point. The heap checks in the alternatives {\em can} be omitted, in which case the topmost @@ -102,205 +89,44 @@ heapcheck will take their worst case into account. In favour of omitting \tr{!B!}, \tr{!C!}: -\begin{itemize} -\item -{\em May} save a heap overflow test, + - {\em May} save a heap overflow test, if ...A... allocates anything. The other advantage of this is that we can use relative addressing from a single Hp to get at all the closures so allocated. -\item - No need to save volatile vars etc across the case -\end{itemize} + + - No need to save volatile vars etc across the case Against: -\begin{itemize} -\item - May do more allocation than reqd. This sometimes bites us + - May do more allocation than reqd. This sometimes bites us badly. For example, nfib (ha!) allocates about 30\% more space if the worst-casing is done, because many many calls to nfib are leaf calls which don't need to allocate anything. This never hurts us if there is only one alternative. -\end{itemize} - - -*** NOT YET DONE *** The difficulty is that \tr{!B!}, \tr{!C!} need -to take account of what is live, and that includes all live volatile -variables, even if they also have stable analogues. Furthermore, the -stack pointers must be lined up properly so that GC sees tidy stacks. -If these things are done, then the heap checks can be done at \tr{!B!} and -\tr{!C!} without a full save-volatile-vars sequence. \begin{code} cgCase :: StgExpr -> StgLiveVars -> StgLiveVars - -> Unique - -> StgCaseAlts + -> Id + -> SRT + -> AltType + -> [StgAlt] -> Code \end{code} -Several special cases for primitive operations. - -******* TO DO TO DO: fix what follows - -Special case for - - case (op x1 ... xn) of - y -> e - -where the type of the case scrutinee is a multi-constuctor algebraic type. -Then we simply compile code for - - let y = op x1 ... xn - in - e - -In this case: - - case (op x1 ... xn) of - C a b -> ... - y -> e - -where the type of the case scrutinee is a multi-constuctor algebraic type. -we just bomb out at the moment. It never happens in practice. - -**** END OF TO DO TO DO - -\begin{code} -cgCase scrut@(StgPrim op args _) live_in_whole_case live_in_alts uniq - (StgAlgAlts _ alts (StgBindDefault id _ deflt_rhs)) - = if not (null alts) then - panic "cgCase: case on PrimOp with default *and* alts\n" - -- For now, die if alts are non-empty - else - cgExpr (StgLet (StgNonRec id scrut_rhs) deflt_rhs) - where - scrut_rhs = StgRhsClosure useCurrentCostCentre stgArgOcc{-safe-} scrut_free_vars - Updatable [] scrut - scrut_free_vars = [ fv | StgVarArg fv <- args, not (toplevelishId fv) ] - -- Hack, hack -\end{code} - +Special case #1: case of literal. \begin{code} -cgCase (StgPrim op args _) live_in_whole_case live_in_alts uniq alts - | not (primOpCanTriggerGC op) - = - -- Get amodes for the arguments and results - getPrimOpArgAmodes op args `thenFC` \ arg_amodes -> - let - result_amodes = getPrimAppResultAmodes uniq alts - liveness_mask = panic "cgCase: liveness of non-GC-ing primop touched\n" - in - -- Perform the operation - getVolatileRegs live_in_alts `thenFC` \ vol_regs -> - - -- seq cannot happen here => no additional B Stack alloc - - absC (COpStmt result_amodes op - arg_amodes -- note: no liveness arg - liveness_mask vol_regs) `thenC` - - -- Scrutinise the result - cgInlineAlts NoGC uniq alts - - | otherwise -- *Can* trigger GC - = getPrimOpArgAmodes op args `thenFC` \ arg_amodes -> - - -- Get amodes for the arguments and results, and assign to regs - -- (Can-trigger-gc primops guarantee to have their (nonRobust) - -- args in regs) - let - op_result_regs = assignPrimOpResultRegs op - - op_result_amodes = map CReg op_result_regs - - (op_arg_amodes, liveness_mask, arg_assts) - = makePrimOpArgsRobust op arg_amodes - - liveness_arg = mkIntCLit liveness_mask - in - -- Tidy up in case GC happens... - - -- Nota Bene the use of live_in_whole_case in nukeDeadBindings. - -- Reason: the arg_assts computed above may refer to some stack slots - -- which are not live in the alts. So we mustn't use those slots - -- to save volatile vars in! - nukeDeadBindings live_in_whole_case `thenC` - saveVolatileVars live_in_alts `thenFC` \ volatile_var_save_assts -> - - -- Allocate stack words for the prim-op itself, - -- these are guaranteed to be ON TOP OF the stack. - -- Currently this is used *only* by the seq# primitive op. - let - (a_req,b_req) = case (primOpStackRequired op) of - NoStackRequired -> (0, 0) - FixedStackRequired a b -> (a, b) - VariableStackRequired -> (0, 0) -- i.e. don't care - in - allocAStackTop a_req `thenFC` \ a_slot -> - allocBStackTop b_req `thenFC` \ b_slot -> - - getEndOfBlockInfo `thenFC` \ eob_info@(EndOfBlockInfo args_spa args_spb sequel) -> - -- a_req and b_req allocate stack space that is taken care of by the - -- macros generated for the primops; thus, we there is no need to adjust - -- this part of the stacks later on (=> +a_req in EndOfBlockInfo) - -- currently all this is only used for SeqOp - forkEval (if True {- a_req==0 && b_req==0 -} - then eob_info - else (EndOfBlockInfo (args_spa+a_req) - (args_spb+b_req) sequel)) nopC - ( - getAbsC (cgInlineAlts GCMayHappen uniq alts) `thenFC` \ abs_c -> - absC (CRetUnVector vtbl_label (CLabelledCode return_label abs_c)) - `thenC` - returnFC (CaseAlts (CUnVecLbl return_label vtbl_label) - Nothing{-no semi-tagging-})) - `thenFC` \ new_eob_info -> - - -- Record the continuation info - setEndOfBlockInfo new_eob_info ( - - -- Now "return" to the inline alternatives; this will get - -- compiled to a fall-through. - let - simultaneous_assts = arg_assts `mkAbsCStmts` volatile_var_save_assts - - -- do_op_and_continue will be passed an amode for the continuation - do_op_and_continue sequel - = absC (COpStmt op_result_amodes - op - (pin_liveness op liveness_arg op_arg_amodes) - liveness_mask - [{-no vol_regs-}]) - `thenC` - - sequelToAmode sequel `thenFC` \ dest_amode -> - absC (CReturn dest_amode DirectReturn) - - -- Note: we CJump even for algebraic data types, - -- because cgInlineAlts always generates code, never a - -- vector. - in - performReturn simultaneous_assts do_op_and_continue live_in_alts - ) - where - -- for all PrimOps except ccalls, we pin the liveness info - -- on as the first "argument" - -- ToDo: un-duplicate? - - pin_liveness (CCallOp _ _ _ _ _) _ args = args - pin_liveness other_op liveness_arg args - = liveness_arg :args - - vtbl_label = mkVecTblLabel uniq - return_label = mkReturnPtLabel uniq - +cgCase (StgLit lit) live_in_whole_case live_in_alts bndr srt + alt_type@(PrimAlt tycon) alts + = bindNewToTemp bndr `thenFC` \ tmp_amode -> + absC (CAssign tmp_amode (CLit lit)) `thenC` + cgPrimAlts NoGC tmp_amode alts alt_type \end{code} -Another special case: scrutinising a primitive-typed variable. No +Special case #2: scrutinising a primitive-typed variable. No evaluation required. We don't save volatile variables, nor do we do a heap-check in the alternatives. Instead, the heap usage of the alternatives is worst-cased and passed upstream. This can result in @@ -308,134 +134,202 @@ allocating more heap than strictly necessary, but it will sometimes eliminate a heap check altogether. \begin{code} -cgCase (StgApp v [] _) live_in_whole_case live_in_alts uniq (StgPrimAlts ty alts deflt) - = getArgAmode v `thenFC` \ amode -> - cgPrimAltsGivenScrutinee NoGC amode alts deflt +cgCase (StgApp v []) live_in_whole_case live_in_alts bndr srt + alt_type@(PrimAlt tycon) alts + + = -- Careful! we can't just bind the default binder to the same thing + -- as the scrutinee, since it might be a stack location, and having + -- two bindings pointing at the same stack locn doesn't work (it + -- confuses nukeDeadBindings). Hence, use a new temp. + getCAddrMode v `thenFC` \ amode -> + bindNewToTemp bndr `thenFC` \ tmp_amode -> + absC (CAssign tmp_amode amode) `thenC` + cgPrimAlts NoGC tmp_amode alts alt_type +\end{code} + +Special case #3: inline PrimOps. + +\begin{code} +cgCase (StgOpApp op@(StgPrimOp primop) args _) + live_in_whole_case live_in_alts bndr srt alt_type alts + | not (primOpOutOfLine primop) + = -- Get amodes for the arguments and results + getArgAmodes args `thenFC` \ arg_amodes -> + getVolatileRegs live_in_alts `thenFC` \ vol_regs -> + + case alt_type of + PrimAlt tycon -- PRIMITIVE ALTS + -> bindNewToTemp bndr `thenFC` \ tmp_amode -> + absC (COpStmt [tmp_amode] op arg_amodes vol_regs) `thenC` + -- Note: no liveness arg + cgPrimAlts NoGC tmp_amode alts alt_type + + UbxTupAlt tycon -- UNBOXED TUPLE ALTS + -> -- No heap check, no yield, just get in there and do it. + -- NB: the case binder isn't bound to anything; + -- it has a unboxed tuple type + mapFCs bindNewToTemp res_ids `thenFC` \ res_tmps -> + absC (COpStmt res_tmps op arg_amodes vol_regs) `thenC` + cgExpr rhs + where + [(_, res_ids, _, rhs)] = alts + + AlgAlt tycon -- ENUMERATION TYPE RETURN + -> ASSERT( isEnumerationTyCon tycon ) + do_enum_primop primop `thenFC` \ tag_amode -> + + -- Bind the default binder if necessary + -- (avoiding it avoids the assignment) + -- The deadness info is set by StgVarInfo + (if (isDeadBinder bndr) + then nopC + else bindNewToTemp bndr `thenFC` \ tmp_amode -> + absC (CAssign tmp_amode (tagToClosure tycon tag_amode)) + ) `thenC` + + -- Compile the alts + cgAlgAlts NoGC (getUnique bndr) + Nothing{-cc_slot-} False{-no semi-tagging-} + (AlgAlt tycon) alts `thenFC` \ tagged_alts -> + + -- Do the switch + absC (mkAlgAltsCSwitch tag_amode tagged_alts) + where + do_enum_primop :: PrimOp -> FCode CAddrMode -- Returns amode for result + do_enum_primop TagToEnumOp -- No code! + = returnFC (only arg_amodes) + + do_enum_primop primop + = absC (COpStmt [tag_amode] op arg_amodes vol_regs) `thenC` + returnFC tag_amode + where + tag_amode = CTemp (newTagUnique (getUnique bndr) 'C') IntRep + -- Being a bit short of uniques for temporary variables here, + -- we use newTagUnique to generate a new unique from the case + -- binder. The case binder's unique will presumably have + -- the 'c' tag (generated by CoreToStg), so we just change + -- its tag to 'C' (for 'case') to ensure it doesn't clash with + -- anything else. + -- We can't use the unique from the case binder, becaus e + -- this is used to hold the actual result closure + -- (via the call to bindNewToTemp) + + other -> pprPanic "cgCase: case of primop has strange alt type" (ppr alt_type) \end{code} +TODO: Case-of-case of primop can probably be done inline too (but +maybe better to translate it out beforehand). See +ghc/lib/misc/PackedString.lhs for examples where this crops up (with +4.02). + Special case: scrutinising a non-primitive variable. This can be done a little better than the general case, because we can reuse/trim the stack slot holding the variable (if it is in one). \begin{code} -cgCase (StgApp (StgVarArg fun) args _ {-lvs must be same as live_in_alts-}) - live_in_whole_case live_in_alts uniq alts@(StgAlgAlts _ _ _) - = - getCAddrModeAndInfo fun `thenFC` \ (fun_amode, lf_info) -> +cgCase (StgApp fun args) + live_in_whole_case live_in_alts bndr srt alt_type alts + = getCAddrModeAndInfo fun `thenFC` \ (fun', fun_amode, lf_info) -> getArgAmodes args `thenFC` \ arg_amodes -> - -- Squish the environment + -- Nuking dead bindings *before* calculating the saves is the + -- value-add here. We might end up freeing up some slots currently + -- occupied by variables only required for the call. + -- NOTE: we need to look up the variables used in the call before + -- doing this, because some of them may not be in the environment + -- afterward. nukeDeadBindings live_in_alts `thenC` saveVolatileVarsAndRegs live_in_alts `thenFC` \ (save_assts, alts_eob_info, maybe_cc_slot) -> - forkEval alts_eob_info - nopC (cgEvalAlts maybe_cc_slot uniq alts) `thenFC` \ scrut_eob_info -> - setEndOfBlockInfo scrut_eob_info ( - tailCallBusiness fun fun_amode lf_info arg_amodes live_in_alts save_assts - ) + forkEval alts_eob_info + ( allocStackTop retPrimRepSize + `thenFC` \_ -> nopC ) + ( deAllocStackTop retPrimRepSize `thenFC` \_ -> + cgEvalAlts maybe_cc_slot bndr srt alt_type alts ) + `thenFC` \ scrut_eob_info -> + setEndOfBlockInfo (maybeReserveSeqFrame alt_type scrut_eob_info) $ + performTailCall fun' fun_amode lf_info arg_amodes save_assts \end{code} +Note about return addresses: we *always* push a return address, even +if because of an optimisation we end up jumping direct to the return +code (not through the address itself). The alternatives always assume +that the return address is on the stack. The return address is +required in case the alternative performs a heap check, since it +encodes the liveness of the slots in the activation record. + +On entry to the case alternative, we can re-use the slot containing +the return address immediately after the heap check. That's what the +deAllocStackTop call is doing above. + Finally, here is the general case. \begin{code} -cgCase expr live_in_whole_case live_in_alts uniq alts +cgCase expr live_in_whole_case live_in_alts bndr srt alt_type alts = -- Figure out what volatile variables to save nukeDeadBindings live_in_whole_case `thenC` + saveVolatileVarsAndRegs live_in_alts `thenFC` \ (save_assts, alts_eob_info, maybe_cc_slot) -> - -- Save those variables right now! + -- Save those variables right now! absC save_assts `thenC` + -- generate code for the alts forkEval alts_eob_info - (nukeDeadBindings live_in_alts) - (cgEvalAlts maybe_cc_slot uniq alts) `thenFC` \ scrut_eob_info -> - - setEndOfBlockInfo scrut_eob_info (cgExpr expr) + (nukeDeadBindings live_in_alts `thenC` + allocStackTop retPrimRepSize -- space for retn address + `thenFC` \_ -> nopC + ) + (deAllocStackTop retPrimRepSize `thenFC` \_ -> + cgEvalAlts maybe_cc_slot bndr srt alt_type alts) `thenFC` \ scrut_eob_info -> + + setEndOfBlockInfo (maybeReserveSeqFrame alt_type scrut_eob_info) $ + cgExpr expr \end{code} -%************************************************************************ -%* * -\subsection[CgCase-primops]{Primitive applications} -%* * -%************************************************************************ - -Get result amodes for a primitive operation, in the case wher GC can't happen. -The amodes are returned in canonical order, ready for the prim-op! - - Alg case: temporaries named as in the alternatives, - plus (CTemp u) for the tag (if needed) - Prim case: (CTemp u) - -This is all disgusting, because these amodes must be consistent with those -invented by CgAlgAlts. +There's a lot of machinery going on behind the scenes to manage the +stack pointer here. forkEval takes the virtual Sp and free list from +the first argument, and turns that into the *real* Sp for the second +argument. It also uses this virtual Sp as the args-Sp in the EOB info +returned, so that the scrutinee will trim the real Sp back to the +right place before doing whatever it does. + --SDM (who just spent an hour figuring this out, and didn't want to + forget it). + +Why don't we push the return address just before evaluating the +scrutinee? Because the slot reserved for the return address might +contain something useful, so we wait until performing a tail call or +return before pushing the return address (see +CgTailCall.pushReturnAddress). + +This also means that the environment doesn't need to know about the +free stack slot for the return address (for generating bitmaps), +because we don't reserve it until just before the eval. + +TODO!! Problem: however, we have to save the current cost centre +stack somewhere, because at the eval point the current CCS might be +different. So we pick a free stack slot and save CCCS in it. The +problem with this is that this slot isn't recorded as free/unboxed in +the environment, so a case expression in the scrutinee will have the +wrong bitmap attached. Fortunately we don't ever seem to see +case-of-case at the back end. One solution might be to shift the +saved CCS to the correct place in the activation record just before +the jump. + --SDM + +(one consequence of the above is that activation records on the stack +don't follow the layout of closures when we're profiling. The CCS +could be anywhere within the record). \begin{code} -getPrimAppResultAmodes - :: Unique - -> StgCaseAlts - -> [CAddrMode] +maybeReserveSeqFrame PolyAlt (EndOfBlockInfo args_sp (CaseAlts amode stuff _)) + = EndOfBlockInfo (args_sp + retPrimRepSize) (CaseAlts amode stuff True) +maybeReserveSeqFrame other scrut_eob_info = scrut_eob_info \end{code} -\begin{code} --- If there's an StgBindDefault which does use the bound --- variable, then we can only handle it if the type involved is --- an enumeration type. That's important in the case --- of comparisions: --- --- case x ># y of --- r -> f r --- --- The only reason for the restriction to *enumeration* types is our --- inability to invent suitable temporaries to hold the results; --- Elaborating the CTemp addr mode to have a second uniq field --- (which would simply count from 1) would solve the problem. --- Anyway, cgInlineAlts is now capable of handling all cases; --- it's only this function which is being wimpish. - -getPrimAppResultAmodes uniq (StgAlgAlts ty alts (StgBindDefault _ True {- used -} _)) - | isEnumerationTyCon spec_tycon = [tag_amode] - | otherwise = panic "getPrimAppResultAmodes: non-enumeration algebraic alternatives with default" - where - -- A temporary variable to hold the tag; this is unaffected by GC because - -- the heap-checks in the branches occur after the switch - tag_amode = CTemp uniq IntRep - (spec_tycon, _, _) = getAppSpecDataTyConExpandingDicts ty - -getPrimAppResultAmodes uniq (StgAlgAlts ty alts other_default) - -- Default is either StgNoDefault or StgBindDefault with unused binder - = case alts of - [_] -> arg_amodes -- No need for a tag - other -> tag_amode : arg_amodes - where - -- A temporary variable to hold the tag; this is unaffected by GC because - -- the heap-checks in the branches occur after the switch - tag_amode = CTemp uniq IntRep - - -- Sort alternatives into canonical order; there must be a complete - -- set because there's no default case. - sorted_alts = sortLt lt alts - (con1,_,_,_) `lt` (con2,_,_,_) = dataConTag con1 < dataConTag con2 - - arg_amodes :: [CAddrMode] - - -- Turn them into amodes - arg_amodes = concat (map mk_amodes sorted_alts) - mk_amodes (con, args, use_mask, rhs) - = [ CTemp (uniqueOf arg) (idPrimRep arg) | arg <- args ] -\end{code} - -The situation is simpler for primitive -results, because there is only one! - -\begin{code} -getPrimAppResultAmodes uniq (StgPrimAlts ty _ _) - = [CTemp uniq (typePrimRep ty)] -\end{code} - - %************************************************************************ %* * \subsection[CgCase-alts]{Alternatives} @@ -447,16 +341,55 @@ alternatives of a @case@, used in a context when there is some evaluation to be done. \begin{code} -cgEvalAlts :: Maybe VirtualSpBOffset -- Offset of cost-centre to be restored, if any - -> Unique - -> StgCaseAlts - -> FCode Sequel -- Any addr modes inside are guaranteed to be a label - -- so that we can duplicate it without risk of - -- duplicating code +cgEvalAlts :: Maybe VirtualSpOffset -- Offset of cost-centre to be restored, if any + -> Id + -> SRT -- SRT for the continuation + -> AltType + -> [StgAlt] + -> FCode Sequel -- Any addr modes inside are guaranteed + -- to be a label so that we can duplicate it + -- without risk of duplicating code + +cgEvalAlts cc_slot bndr srt (UbxTupAlt _) [(con,args,_,rhs)] + = -- Unboxed tuple case + -- By now, the simplifier should have have turned it + -- into case e of (# a,b #) -> e + -- There shouldn't be a + -- case e of DEFAULT -> e + ASSERT2( case con of { DataAlt _ -> True; other -> False }, + text "cgEvalAlts: dodgy case of unboxed tuple type" ) + + forkAbsC ( -- forkAbsC for the RHS, so that the envt is + -- not changed for the mkRetDirect call + bindUnboxedTupleComponents args `thenFC` \ (live_regs, ptrs, nptrs, _) -> + -- restore the CC *after* binding the tuple components, so that we + -- get the stack offset of the saved CC right. + restoreCurrentCostCentre cc_slot True `thenC` + -- Generate a heap check if necessary + unbxTupleHeapCheck live_regs ptrs nptrs AbsCNop ( + -- And finally the code for the alternative + cgExpr rhs + )) `thenFC` \ abs_c -> + mkRetDirectTarget bndr abs_c srt `thenFC` \ lbl -> + returnFC (CaseAlts lbl Nothing False) + +cgEvalAlts cc_slot bndr srt alt_type@(PrimAlt tycon) alts + = forkAbsC ( -- forkAbsC for the RHS, so that the envt is + -- not changed for the mkRetDirect call + restoreCurrentCostCentre cc_slot True `thenC` + bindNewToReg bndr reg (mkLFArgument bndr) `thenC` + cgPrimAlts GCMayHappen (CReg reg) alts alt_type + ) `thenFC` \ abs_c -> + mkRetDirectTarget bndr abs_c srt `thenFC` \ lbl -> + returnFC (CaseAlts lbl Nothing False) + where + reg = dataReturnConvPrim kind + kind = tyConPrimRep tycon -cgEvalAlts cc_slot uniq (StgAlgAlts ty alts deflt) - = -- Generate the instruction to restore cost centre, if any - restoreCurrentCostCentre cc_slot `thenFC` \ cc_restore -> +cgEvalAlts cc_slot bndr srt alt_type alts + = -- Algebraic and polymorphic case + -- Bind the default binder + bindNewToReg bndr node (mkLFArgument bndr) `thenC` -- Generate sequel info for use downstream -- At the moment, we only do it if the type is vector-returnable. @@ -467,93 +400,34 @@ cgEvalAlts cc_slot uniq (StgAlgAlts ty alts deflt) -- -- which is worse than having the alt code in the switch statement - let - (spec_tycon, _, _) = getAppSpecDataTyConExpandingDicts ty - - use_labelled_alts - = case ctrlReturnConvAlg spec_tycon of - VectoredReturn _ -> True - _ -> False - - semi_tagged_stuff - = if not use_labelled_alts then - Nothing -- no semi-tagging info - else - cgSemiTaggedAlts uniq alts deflt -- Just - in - cgAlgAlts GCMayHappen uniq cc_restore use_labelled_alts ty alts deflt True - `thenFC` \ (tagged_alt_absCs, deflt_absC) -> + let ret_conv = case alt_type of + AlgAlt tc -> ctrlReturnConvAlg tc + PolyAlt -> UnvectoredReturn 0 - mkReturnVector uniq ty tagged_alt_absCs deflt_absC `thenFC` \ return_vec -> + use_labelled_alts = case ret_conv of + VectoredReturn _ -> True + _ -> False - returnFC (CaseAlts return_vec semi_tagged_stuff) + semi_tagged_stuff = cgSemiTaggedAlts use_labelled_alts bndr alts -cgEvalAlts cc_slot uniq (StgPrimAlts ty alts deflt) - = -- Generate the instruction to restore cost centre, if any - restoreCurrentCostCentre cc_slot `thenFC` \ cc_restore -> + in + cgAlgAlts GCMayHappen (getUnique bndr) + cc_slot use_labelled_alts + alt_type alts `thenFC` \ tagged_alt_absCs -> - -- Generate the switch - getAbsC (cgPrimAlts GCMayHappen uniq ty alts deflt) `thenFC` \ abs_c -> + mkRetVecTarget bndr tagged_alt_absCs + srt ret_conv `thenFC` \ return_vec -> - -- Generate the labelled block, starting with restore-cost-centre - absC (CRetUnVector vtbl_label - (CLabelledCode return_label (cc_restore `mkAbsCStmts` abs_c))) - `thenC` - -- Return an amode for the block - returnFC (CaseAlts (CUnVecLbl return_label vtbl_label) Nothing{-no semi-tagging-}) - where - vtbl_label = mkVecTblLabel uniq - return_label = mkReturnPtLabel uniq + returnFC (CaseAlts return_vec semi_tagged_stuff False) \end{code} -\begin{code} -cgInlineAlts :: GCFlag -> Unique - -> StgCaseAlts - -> Code -\end{code} - HWL comment on {\em GrAnSim\/} (adding GRAN_YIELDs for context switch): If we do an inlining of the case no separate functions for returning are created, so we don't have to generate a GRAN_YIELD in that case. This info must be propagated to cgAlgAltRhs (where the GRAN_YIELD macro might be emitted). Hence, the new Bool arg to cgAlgAltRhs. -First case: algebraic case, exactly one alternative, no default. -In this case the primitive op will not have set a temporary to the -tag, so we shouldn't generate a switch statment. Instead we just -do the right thing. - -\begin{code} -cgInlineAlts gc_flag uniq (StgAlgAlts ty [alt@(con,args,use_mask,rhs)] StgNoDefault) - = cgAlgAltRhs gc_flag con args use_mask rhs False{-no yield macro if alt gets inlined-} -\end{code} - -Second case: algebraic case, several alternatives. -Tag is held in a temporary. - -\begin{code} -cgInlineAlts gc_flag uniq (StgAlgAlts ty alts deflt) - = cgAlgAlts gc_flag uniq AbsCNop{-restore_cc-} False{-no semi-tagging-} - ty alts deflt - False{-don't emit yield-} `thenFC` \ (tagged_alts, deflt_c) -> - - -- Do the switch - absC (mkAlgAltsCSwitch tag_amode tagged_alts deflt_c) - where - -- A temporary variable to hold the tag; this is unaffected by GC because - -- the heap-checks in the branches occur after the switch - tag_amode = CTemp uniq IntRep -\end{code} - -Third (real) case: primitive result type. - -\begin{code} -cgInlineAlts gc_flag uniq (StgPrimAlts ty alts deflt) - = cgPrimAlts gc_flag uniq ty alts deflt -\end{code} - - %************************************************************************ %* * \subsection[CgCase-alg-alts]{Algebraic alternatives} @@ -570,231 +444,43 @@ are inlined alternatives. \begin{code} cgAlgAlts :: GCFlag - -> Unique - -> AbstractC -- Restore-cost-centre instruction - -> Bool -- True <=> branches must be labelled - -> Type -- From the case statement - -> [(Id, [Id], [Bool], StgExpr)] -- The alternatives - -> StgCaseDefault -- The default - -> Bool -- Context switch at alts? - -> FCode ([(ConTag, AbstractC)], -- The branches - AbstractC -- The default case - ) -\end{code} - -The case with a default which has a binder is different. We need to -pick all the constructors which aren't handled explicitly by an -alternative, and which return their results in registers, allocate -them explicitly in the heap, and jump to a join point for the default -case. - -OLD: All of this only works if a heap-check is required anyway, because -otherwise it isn't safe to allocate. - -NEW (July 94): now false! It should work regardless of gc_flag, -because of the extra_branches argument now added to forkAlts. - -We put a heap-check at the join point, for the benefit of constructors -which don't need to do allocation. This means that ones which do need -to allocate may end up doing two heap-checks; but that's just too bad. -(We'd need two join labels otherwise. ToDo.) - -It's all pretty turgid anyway. - -\begin{code} -cgAlgAlts gc_flag uniq restore_cc semi_tagging - ty alts deflt@(StgBindDefault binder True{-used-} _) - emit_yield{-should a yield macro be emitted?-} - = let - extra_branches :: [FCode (ConTag, AbstractC)] - extra_branches = catMaybes (map mk_extra_branch default_cons) - - must_label_default = semi_tagging || not (null extra_branches) - in - forkAlts (map (cgAlgAlt gc_flag uniq restore_cc semi_tagging emit_yield) alts) - extra_branches - (cgAlgDefault gc_flag uniq restore_cc must_label_default deflt emit_yield) - where - - default_join_lbl = mkDefaultLabel uniq - jump_instruction = CJump (CLbl default_join_lbl CodePtrRep) - - (spec_tycon, _, spec_cons) = getAppSpecDataTyConExpandingDicts ty - - alt_cons = [ con | (con,_,_,_) <- alts ] - - default_cons = [ spec_con | spec_con <- spec_cons, -- In this type - spec_con `not_elem` alt_cons ] -- Not handled explicitly - where - not_elem = isn'tIn "cgAlgAlts" - - -- (mk_extra_branch con) returns the a maybe for the extra branch for con. - -- The "maybe" is because con may return in heap, in which case there is - -- nothing to do. Otherwise, we have a special case for a nullary constructor, - -- but in the general case we do an allocation and heap-check. - - mk_extra_branch :: DataCon -> (Maybe (FCode (ConTag, AbstractC))) - - mk_extra_branch con - = ASSERT(isDataCon con) - case dataReturnConvAlg con of - ReturnInHeap -> Nothing - ReturnInRegs rs -> Just (getAbsC (alloc_code rs) `thenFC` \ abs_c -> - returnFC (tag, abs_c) - ) - where - lf_info = mkConLFInfo con - tag = dataConTag con - - -- alloc_code generates code to allocate constructor con, whose args are - -- in the arguments to alloc_code, assigning the result to Node. - alloc_code :: [MagicId] -> Code - - alloc_code regs - = possibleHeapCheck gc_flag regs False ( - buildDynCon binder useCurrentCostCentre con - (map CReg regs) (all zero_size regs) - `thenFC` \ idinfo -> - idInfoToAmode PtrRep idinfo `thenFC` \ amode -> - - absC (CAssign (CReg node) amode) `thenC` - absC jump_instruction - ) - where - zero_size reg = getPrimRepSize (magicIdPrimRep reg) == 0 -\end{code} - -Now comes the general case - -\begin{code} -cgAlgAlts gc_flag uniq restore_cc must_label_branches ty alts deflt - {- The deflt is either StgNoDefault or a BindDefault which doesn't use the binder -} - emit_yield{-should a yield macro be emitted?-} - - = forkAlts (map (cgAlgAlt gc_flag uniq restore_cc must_label_branches emit_yield) alts) - [{- No "extra branches" -}] - (cgAlgDefault gc_flag uniq restore_cc must_label_branches deflt emit_yield) -\end{code} - -\begin{code} -cgAlgDefault :: GCFlag - -> Unique -> AbstractC -> Bool -- turgid state... - -> StgCaseDefault -- input - -> Bool - -> FCode AbstractC -- output - -cgAlgDefault gc_flag uniq restore_cc must_label_branch - StgNoDefault _ - = returnFC AbsCNop - -cgAlgDefault gc_flag uniq restore_cc must_label_branch - (StgBindDefault _ False{-binder not used-} rhs) - emit_yield{-should a yield macro be emitted?-} - - = getAbsC (absC restore_cc `thenC` - let - emit_gran_macros = opt_GranMacros - in - (if emit_gran_macros && emit_yield - then yield [] False - else absC AbsCNop) `thenC` - -- liveness same as in possibleHeapCheck below - possibleHeapCheck gc_flag [] False (cgExpr rhs)) `thenFC` \ abs_c -> - let - final_abs_c | must_label_branch = CJump (CLabelledCode lbl abs_c) - | otherwise = abs_c - in - returnFC final_abs_c - where - lbl = mkDefaultLabel uniq - - -cgAlgDefault gc_flag uniq restore_cc must_label_branch - (StgBindDefault binder True{-binder used-} rhs) - emit_yield{-should a yield macro be emitted?-} - - = -- We have arranged that Node points to the thing, even - -- even if we return in registers - bindNewToReg binder node mkLFArgument `thenC` - getAbsC (absC restore_cc `thenC` - let - emit_gran_macros = opt_GranMacros - in - (if emit_gran_macros && emit_yield - then yield [node] False - else absC AbsCNop) `thenC` - -- liveness same as in possibleHeapCheck below - possibleHeapCheck gc_flag [node] False (cgExpr rhs) - -- Node is live, but doesn't need to point at the thing itself; - -- it's ok for Node to point to an indirection or FETCH_ME - -- Hence no need to re-enter Node. - ) `thenFC` \ abs_c -> - - let - final_abs_c | must_label_branch = CJump (CLabelledCode lbl abs_c) - | otherwise = abs_c - in - returnFC final_abs_c - where - lbl = mkDefaultLabel uniq - --- HWL comment on GrAnSim: GRAN_YIELDs needed; emitted in cgAlgAltRhs + -> Unique + -> Maybe VirtualSpOffset + -> Bool -- True <=> branches must be labelled + -- (used for semi-tagging) + -> AltType -- ** AlgAlt or PolyAlt only ** + -> [StgAlt] -- The alternatives + -> FCode [(AltCon, AbstractC)] -- The branches + +cgAlgAlts gc_flag uniq restore_cc must_label_branches alt_type alts + = forkAlts [ cgAlgAlt gc_flag uniq restore_cc must_label_branches alt_type alt + | alt <- alts] cgAlgAlt :: GCFlag - -> Unique -> AbstractC -> Bool -- turgid state - -> Bool -- Context switch at alts? - -> (Id, [Id], [Bool], StgExpr) - -> FCode (ConTag, AbstractC) - -cgAlgAlt gc_flag uniq restore_cc must_label_branch - emit_yield{-should a yield macro be emitted?-} - (con, args, use_mask, rhs) - = getAbsC (absC restore_cc `thenC` - cgAlgAltRhs gc_flag con args use_mask rhs - emit_yield - ) `thenFC` \ abs_c -> + -> Unique -> Maybe VirtualSpOffset -> Bool -- turgid state + -> AltType -- ** AlgAlt or PolyAlt only ** + -> StgAlt + -> FCode (AltCon, AbstractC) + +cgAlgAlt gc_flag uniq cc_slot must_label_branch + alt_type (con, args, use_mask, rhs) + = getAbsC (bind_con_args con args `thenFC` \ _ -> + restoreCurrentCostCentre cc_slot True `thenC` + maybeAltHeapCheck gc_flag alt_type (cgExpr rhs) + ) `thenFC` \ abs_c -> let - final_abs_c | must_label_branch = CJump (CLabelledCode lbl abs_c) + final_abs_c | must_label_branch = CCodeBlock lbl abs_c | otherwise = abs_c in - returnFC (tag, final_abs_c) + returnFC (con, final_abs_c) where - tag = dataConTag con - lbl = mkAltLabel uniq tag - -cgAlgAltRhs :: GCFlag - -> Id - -> [Id] - -> [Bool] - -> StgExpr - -> Bool -- context switch? - -> Code -cgAlgAltRhs gc_flag con args use_mask rhs emit_yield - = let - (live_regs, node_reqd) - = case (dataReturnConvAlg con) of - ReturnInHeap -> ([], True) - ReturnInRegs regs -> ([reg | (reg,True) <- zipEqual "cgAlgAltRhs" regs use_mask], False) - -- Pick the live registers using the use_mask - -- Doing so is IMPORTANT, because with semi-tagging - -- enabled only the live registers will have valid - -- pointers in them. - in - let - emit_gran_macros = opt_GranMacros - in - (if emit_gran_macros && emit_yield - then yield live_regs node_reqd - else absC AbsCNop) `thenC` - -- liveness same as in possibleHeapCheck below - possibleHeapCheck gc_flag live_regs node_reqd ( - (case gc_flag of - NoGC -> mapFCs bindNewToTemp args `thenFC` \ _ -> - nopC - GCMayHappen -> bindConArgs con args - ) `thenC` - cgExpr rhs - ) + lbl = case con of + DataAlt dc -> mkAltLabel uniq (dataConTag dc) + DEFAULT -> mkDefaultLabel uniq + other -> pprPanic "cgAlgAlt" (ppr con) + + bind_con_args DEFAULT args = nopC + bind_con_args (DataAlt dc) args = bindConArgs dc args \end{code} %************************************************************************ @@ -807,60 +493,44 @@ Turgid-but-non-monadic code to conjure up the required info from algebraic case alternatives for semi-tagging. \begin{code} -cgSemiTaggedAlts :: Unique - -> [(Id, [Id], [Bool], StgExpr)] - -> GenStgCaseDefault Id Id +cgSemiTaggedAlts :: Bool -- True <=> use semitagging: each alt will be labelled + -> Id + -> [StgAlt] -> SemiTaggingStuff -cgSemiTaggedAlts uniq alts deflt - = Just (map st_alt alts, st_deflt deflt) +cgSemiTaggedAlts False binder alts + = Nothing +cgSemiTaggedAlts True binder alts + = Just ([st_alt con args | (DataAlt con, args, _, _) <- alts], + case head alts of + (DEFAULT, _, _, _) -> Just st_deflt + other -> Nothing) where - st_deflt StgNoDefault = Nothing - - st_deflt (StgBindDefault binder binder_used _) - = Just (if binder_used then Just binder else Nothing, - (CCallProfCtrMacro SLIT("RET_SEMI_BY_DEFAULT") [], -- ToDo: monadise? - mkDefaultLabel uniq) - ) - - st_alt (con, args, use_mask, _) - = case (dataReturnConvAlg con) of - - ReturnInHeap -> - -- Ha! Nothing to do; Node already points to the thing - (con_tag, - (CCallProfCtrMacro SLIT("RET_SEMI_IN_HEAP") -- ToDo: monadise? - [mkIntCLit (length args)], -- how big the thing in the heap is + uniq = getUnique binder + + st_deflt = (binder, + (CCallProfCtrMacro FSLIT("RET_SEMI_BY_DEFAULT") [], -- ToDo: monadise? + mkDefaultLabel uniq)) + + st_alt con args -- Ha! Nothing to do; Node already points to the thing + = (con_tag, + (CCallProfCtrMacro FSLIT("RET_SEMI_IN_HEAP") -- ToDo: monadise? + [mkIntCLit (length args)], -- how big the thing in the heap is join_label) ) - - ReturnInRegs regs -> - -- We have to load the live registers from the constructor - -- pointed to by Node. - let - (_, regs_w_offsets) = layOutDynCon con magicIdPrimRep regs - - used_regs = selectByMask use_mask regs - - used_regs_w_offsets = [ ro | ro@(reg,offset) <- regs_w_offsets, - reg `is_elem` used_regs] - - is_elem = isIn "cgSemiTaggedAlts" - in - (con_tag, - (mkAbstractCs [ - CCallProfCtrMacro SLIT("RET_SEMI_IN_REGS") -- ToDo: macroise? - [mkIntCLit (length regs_w_offsets), - mkIntCLit (length used_regs_w_offsets)], - CSimultaneous (mkAbstractCs (map move_to_reg used_regs_w_offsets))], - join_label)) where - con_tag = dataConTag con - join_label = mkAltLabel uniq con_tag + con_tag = dataConTag con + join_label = mkAltLabel uniq con_tag + - move_to_reg :: (MagicId, VirtualHeapOffset {-from Node-}) -> AbstractC - move_to_reg (reg, offset) - = CAssign (CReg reg) (CVal (NodeRel offset) (magicIdPrimRep reg)) +tagToClosure :: TyCon -> CAddrMode -> CAddrMode +-- Primops returning an enumeration type (notably Bool) +-- actually return an index into +-- the table of closures for the enumeration type +tagToClosure tycon tag_amode + = CVal (CIndex closure_tbl tag_amode PtrRep) PtrRep + where + closure_tbl = CLbl (mkClosureTblLabel tycon) PtrRep \end{code} %************************************************************************ @@ -869,69 +539,38 @@ cgSemiTaggedAlts uniq alts deflt %* * %************************************************************************ -@cgPrimAlts@ generates a suitable @CSwitch@ for dealing with the -alternatives of a primitive @case@, given an addressing mode for the -thing to scrutinise. It also keeps track of the maximum stack depth -encountered down any branch. +@cgPrimAlts@ generates suitable a @CSwitch@ +for dealing with the alternatives of a primitive @case@, given an +addressing mode for the thing to scrutinise. It also keeps track of +the maximum stack depth encountered down any branch. As usual, no binders in the alternatives are yet bound. \begin{code} cgPrimAlts :: GCFlag - -> Unique - -> Type - -> [(Literal, StgExpr)] -- Alternatives - -> StgCaseDefault -- Default + -> CAddrMode -- Scrutinee + -> [StgAlt] -- Alternatives + -> AltType -> Code - -cgPrimAlts gc_flag uniq ty alts deflt - = cgPrimAltsGivenScrutinee gc_flag scrutinee alts deflt - where - -- A temporary variable, or standard register, to hold the result - scrutinee = case gc_flag of - NoGC -> CTemp uniq kind - GCMayHappen -> CReg (dataReturnConvPrim kind) - - kind = typePrimRep ty - - -cgPrimAltsGivenScrutinee gc_flag scrutinee alts deflt - = forkAlts (map (cgPrimAlt gc_flag) alts) - [{- No "extra branches" -}] - (cgPrimDefault gc_flag scrutinee deflt) `thenFC` \ (alt_absCs, deflt_absC) -> +-- INVARIANT: the default binder is already bound +cgPrimAlts gc_flag scrutinee alts alt_type + = forkAlts (map (cgPrimAlt gc_flag alt_type) alts) `thenFC` \ tagged_absCs -> + let + ((DEFAULT, deflt_absC) : others) = tagged_absCs -- There is always a default + alt_absCs = [(lit,rhs) | (LitAlt lit, rhs) <- others] + in absC (CSwitch scrutinee alt_absCs deflt_absC) - -- CSwitch does sensible things with one or zero alternatives - + -- CSwitch does sensible things with one or zero alternatives cgPrimAlt :: GCFlag - -> (Literal, StgExpr) -- The alternative - -> FCode (Literal, AbstractC) -- Its compiled form - -cgPrimAlt gc_flag (lit, rhs) - = getAbsC rhs_code `thenFC` \ absC -> - returnFC (lit,absC) - where - rhs_code = possibleHeapCheck gc_flag [] False (cgExpr rhs ) - -cgPrimDefault :: GCFlag - -> CAddrMode -- Scrutinee - -> StgCaseDefault - -> FCode AbstractC - -cgPrimDefault gc_flag scrutinee StgNoDefault - = panic "cgPrimDefault: No default in prim case" - -cgPrimDefault gc_flag scrutinee (StgBindDefault _ False{-binder not used-} rhs) - = getAbsC (possibleHeapCheck gc_flag [] False (cgExpr rhs )) - -cgPrimDefault gc_flag scrutinee (StgBindDefault binder True{-used-} rhs) - = getAbsC (possibleHeapCheck gc_flag regs False rhs_code) - where - regs = if isFollowableRep (getAmodeRep scrutinee) then - [node] else [] - - rhs_code = bindNewPrimToAmode binder scrutinee `thenC` - cgExpr rhs + -> AltType + -> StgAlt -- The alternative + -> FCode (AltCon, AbstractC) -- Its compiled form + +cgPrimAlt gc_flag alt_type (con, [], [], rhs) + = ASSERT( case con of { DEFAULT -> True; LitAlt _ -> True; other -> False } ) + getAbsC (maybeAltHeapCheck gc_flag alt_type (cgExpr rhs)) `thenFC` \ abs_c -> + returnFC (con, abs_c) \end{code} @@ -942,28 +581,37 @@ cgPrimDefault gc_flag scrutinee (StgBindDefault binder True{-used-} rhs) %************************************************************************ \begin{code} +maybeAltHeapCheck + :: GCFlag + -> AltType -- PolyAlt, PrimAlt, AlgAlt, but *not* UbxTupAlt + -> Code -- Continuation + -> Code +maybeAltHeapCheck NoGC _ code = code +maybeAltHeapCheck GCMayHappen alt_type code + = -- HWL: maybe need yield here + -- yield [node] True -- XXX live regs wrong + altHeapCheck alt_type code + saveVolatileVarsAndRegs - :: StgLiveVars -- Vars which should be made safe + :: StgLiveVars -- Vars which should be made safe -> FCode (AbstractC, -- Assignments to do the saves - EndOfBlockInfo, -- New sequel, recording where the return - -- address now is - Maybe VirtualSpBOffset) -- Slot for current cost centre - + EndOfBlockInfo, -- sequel for the alts + Maybe VirtualSpOffset) -- Slot for current cost centre saveVolatileVarsAndRegs vars - = saveVolatileVars vars `thenFC` \ var_saves -> - saveCurrentCostCentre `thenFC` \ (maybe_cc_slot, cc_save) -> - saveReturnAddress `thenFC` \ (new_eob_info, ret_save) -> - returnFC (mkAbstractCs [var_saves, cc_save, ret_save], - new_eob_info, + = saveVolatileVars vars `thenFC` \ var_saves -> + saveCurrentCostCentre `thenFC` \ (maybe_cc_slot, cc_save) -> + getEndOfBlockInfo `thenFC` \ eob_info -> + returnFC (mkAbstractCs [var_saves, cc_save], + eob_info, maybe_cc_slot) -saveVolatileVars :: StgLiveVars -- Vars which should be made safe +saveVolatileVars :: StgLiveVars -- Vars which should be made safe -> FCode AbstractC -- Assignments to to the saves saveVolatileVars vars - = save_em (idSetToList vars) + = save_em (varSetElems vars) where save_em [] = returnFC AbsCNop @@ -979,101 +627,49 @@ saveVolatileVars vars returnFC (abs_c `mkAbsCStmts` abs_cs) save_var var vol_amode - | isFollowableRep kind - = allocAStack `thenFC` \ a_slot -> - rebindToAStack var a_slot `thenC` - getSpARelOffset a_slot `thenFC` \ spa_rel -> - returnFC (CAssign (CVal spa_rel kind) vol_amode) - | otherwise - = allocBStack (getPrimRepSize kind) `thenFC` \ b_slot -> - rebindToBStack var b_slot `thenC` - getSpBRelOffset b_slot `thenFC` \ spb_rel -> - returnFC (CAssign (CVal spb_rel kind) vol_amode) + = allocPrimStack (getPrimRepSize kind) `thenFC` \ slot -> + rebindToStack var slot `thenC` + getSpRelOffset slot `thenFC` \ sp_rel -> + returnFC (CAssign (CVal sp_rel kind) vol_amode) where kind = getAmodeRep vol_amode - -saveReturnAddress :: FCode (EndOfBlockInfo, AbstractC) -saveReturnAddress - = getEndOfBlockInfo `thenFC` \ eob_info@(EndOfBlockInfo vA vB sequel) -> - - -- See if it is volatile - case sequel of - InRetReg -> -- Yes, it's volatile - allocBStack retPrimRepSize `thenFC` \ b_slot -> - getSpBRelOffset b_slot `thenFC` \ spb_rel -> - - returnFC (EndOfBlockInfo vA vB (OnStack b_slot), - CAssign (CVal spb_rel RetRep) (CReg RetReg)) - - UpdateCode _ -> -- It's non-volatile all right, but we still need - -- to allocate a B-stack slot for it, *solely* to make - -- sure that update frames for different values do not - -- appear adjacent on the B stack. This makes sure - -- that B-stack squeezing works ok. - -- See note below - allocBStack retPrimRepSize `thenFC` \ b_slot -> - returnFC (eob_info, AbsCNop) - - other -> -- No, it's non-volatile, so do nothing - returnFC (eob_info, AbsCNop) \end{code} -Note about B-stack squeezing. Consider the following:` - - y = [...] \u [] -> ... - x = [y] \u [] -> case y of (a,b) -> a - -The code for x will push an update frame, and then enter y. The code -for y will push another update frame. If the B-stack-squeezer then -wakes up, it will see two update frames right on top of each other, -and will combine them. This is WRONG, of course, because x's value is -not the same as y's. - -The fix implemented above makes sure that we allocate an (unused) -B-stack slot before entering y. You can think of this as holding the -saved value of RetAddr, which (after pushing x's update frame will be -some update code ptr). The compiler is clever enough to load the -static update code ptr into RetAddr before entering ~a~, but the slot -is still there to separate the update frames. +--------------------------------------------------------------------------- When we save the current cost centre (which is done for lexical -scoping), we allocate a free B-stack location, and return (a)~the +scoping), we allocate a free stack location, and return (a)~the virtual offset of the location, to pass on to the alternatives, and (b)~the assignment to do the save (just as for @saveVolatileVars@). \begin{code} saveCurrentCostCentre :: - FCode (Maybe VirtualSpBOffset, -- Where we decide to store it - -- Nothing if not lexical CCs + FCode (Maybe VirtualSpOffset, -- Where we decide to store it AbstractC) -- Assignment to save it - -- AbsCNop if not lexical CCs saveCurrentCostCentre - = let - doing_profiling = opt_SccProfilingOn - in - if not doing_profiling then + = if not opt_SccProfilingOn then returnFC (Nothing, AbsCNop) else - allocBStack (getPrimRepSize CostCentreRep) `thenFC` \ b_slot -> - getSpBRelOffset b_slot `thenFC` \ spb_rel -> - returnFC (Just b_slot, - CAssign (CVal spb_rel CostCentreRep) (CReg CurCostCentre)) - -restoreCurrentCostCentre :: Maybe VirtualSpBOffset -> FCode AbstractC - -restoreCurrentCostCentre Nothing - = returnFC AbsCNop -restoreCurrentCostCentre (Just b_slot) - = getSpBRelOffset b_slot `thenFC` \ spb_rel -> - freeBStkSlot b_slot `thenC` - returnFC (CCallProfCCMacro SLIT("RESTORE_CCC") [CVal spb_rel CostCentreRep]) - -- we use the RESTORE_CCC macro, rather than just - -- assigning into CurCostCentre, in case RESTORE_CCC + allocPrimStack (getPrimRepSize CostCentreRep) `thenFC` \ slot -> + dataStackSlots [slot] `thenC` + getSpRelOffset slot `thenFC` \ sp_rel -> + returnFC (Just slot, + CAssign (CVal sp_rel CostCentreRep) (CReg CurCostCentre)) + +-- Sometimes we don't free the slot containing the cost centre after restoring it +-- (see CgLetNoEscape.cgLetNoEscapeBody). +restoreCurrentCostCentre :: Maybe VirtualSpOffset -> Bool -> Code +restoreCurrentCostCentre Nothing _freeit = nopC +restoreCurrentCostCentre (Just slot) freeit + = getSpRelOffset slot `thenFC` \ sp_rel -> + (if freeit then freeStackSlots [slot] else nopC) `thenC` + absC (CCallProfCCMacro FSLIT("RESTORE_CCCS") [CVal sp_rel CostCentreRep]) + -- we use the RESTORE_CCCS macro, rather than just + -- assigning into CurCostCentre, in case RESTORE_CCCS -- has some sanity-checking in it. \end{code} - %************************************************************************ %* * \subsection[CgCase-return-vec]{Building a return vector} @@ -1084,77 +680,88 @@ Build a return vector, and return a suitable label addressing mode for it. \begin{code} -mkReturnVector :: Unique - -> Type - -> [(ConTag, AbstractC)] -- Branch codes - -> AbstractC -- Default case - -> FCode CAddrMode - -mkReturnVector uniq ty tagged_alt_absCs deflt_absC - = let - (return_vec_amode, vtbl_body) = case (ctrlReturnConvAlg spec_tycon) of { - - UnvectoredReturn _ -> - (CUnVecLbl ret_label vtbl_label, - absC (CRetUnVector vtbl_label - (CLabelledCode ret_label - (mkAlgAltsCSwitch (CReg TagReg) - tagged_alt_absCs - deflt_absC)))); - VectoredReturn table_size -> - (CLbl vtbl_label DataPtrRep, - absC (CRetVector vtbl_label - -- must restore cc before each alt, if required - (map mk_vector_entry [fIRST_TAG .. (table_size+fIRST_TAG-1)]) - deflt_absC)) - --- Leave nops and comments in for now; they are eliminated --- lazily as it's printed. --- (case (nonemptyAbsC deflt_absC) of --- Nothing -> AbsCNop --- Just def -> def) - - } in - vtbl_body `thenC` - returnFC return_vec_amode - -- ) +mkRetDirectTarget :: Id -- Used for labelling only + -> AbstractC -- Return code + -> SRT -- Live CAFs in return code + -> FCode CAddrMode -- Emit the labelled return block, + -- and return its label +mkRetDirectTarget bndr abs_c srt + = buildContLivenessMask bndr `thenFC` \ liveness -> + getSRTInfo name srt `thenFC` \ srt_info -> + absC (CRetDirect uniq abs_c srt_info liveness) `thenC` + return lbl where - - (spec_tycon,_,_) = case (maybeAppSpecDataTyConExpandingDicts ty) of -- *must* be a real "data" type constructor - Just xx -> xx - Nothing -> pprError "ERROR: can't generate code for polymorphic case;\nprobably a mis-use of `seq' or `par';\nthe User's Guide has more details.\nOffending type: " (ppr PprDebug ty) - - vtbl_label = mkVecTblLabel uniq - ret_label = mkReturnPtLabel uniq - - mk_vector_entry :: ConTag -> Maybe CAddrMode - mk_vector_entry tag - = case [ absC | (t, absC) <- tagged_alt_absCs, t == tag ] of - [] -> Nothing - [absC] -> Just (CCode absC) - _ -> panic "mkReturnVector: too many" + name = idName bndr + uniq = getUnique name + lbl = CLbl (mkReturnInfoLabel uniq) RetRep \end{code} -%************************************************************************ -%* * -\subsection[CgCase-utils]{Utilities for handling case expressions} -%* * -%************************************************************************ - -@possibleHeapCheck@ tests a flag passed in to decide whether to -do a heap check or not. - \begin{code} -possibleHeapCheck :: GCFlag -> [MagicId] -> Bool -> Code -> Code +mkRetVecTarget :: Id -- Just for its unique + -> [(AltCon, AbstractC)] -- Branch codes + -> SRT -- Continuation's SRT + -> CtrlReturnConvention + -> FCode CAddrMode -possibleHeapCheck GCMayHappen regs node_reqd code = heapCheck regs node_reqd code -possibleHeapCheck NoGC _ _ code = code -\end{code} +mkRetVecTarget bndr tagged_alt_absCs srt (UnvectoredReturn 0) + = ASSERT( null other_alts ) + mkRetDirectTarget bndr deflt_absC srt + where + ((DEFAULT, deflt_absC) : other_alts) = tagged_alt_absCs -Select a restricted set of registers based on a usage mask. +mkRetVecTarget bndr tagged_alt_absCs srt (UnvectoredReturn n) + = mkRetDirectTarget bndr switch_absC srt + where + -- Find the tag explicitly rather than using tag_reg for now. + -- on architectures with lots of regs the tag will be loaded + -- into tag_reg by the code doing the returning. + tag = CMacroExpr WordRep GET_TAG [CVal (nodeRel 0) DataPtrRep] + switch_absC = mkAlgAltsCSwitch tag tagged_alt_absCs + + +mkRetVecTarget bndr tagged_alt_absCs srt (VectoredReturn table_size) + = buildContLivenessMask bndr `thenFC` \ liveness -> + getSRTInfo name srt `thenFC` \ srt_info -> + let + ret_vector = CRetVector vtbl_lbl vector_table srt_info liveness + in + absC (mkAbstractCs alts_absCs `mkAbsCStmts` ret_vector) `thenC` + -- Alts come first, because we don't want to declare all the symbols -\begin{code} -selectByMask [] [] = [] -selectByMask (True:ms) (x:xs) = x : selectByMask ms xs -selectByMask (False:ms) (x:xs) = selectByMask ms xs + return (CLbl vtbl_lbl DataPtrRep) + where + tags = [fIRST_TAG .. (table_size+fIRST_TAG-1)] + vector_table = map mk_vector_entry tags + alts_absCs = map snd (sortBy cmp tagged_alt_absCs) + -- The sort is unnecessary; just there for now + -- to make the new order the same as the old + (DEFAULT,_) `cmp` (DEFAULT,_) = EQ + (DEFAULT,_) `cmp` _ = GT + (DataAlt d1,_) `cmp` (DataAlt d2,_) = dataConTag d1 `compare` dataConTag d2 + (DataAlt d1,_) `cmp` (DEFAULT, _) = LT + -- Others impossible + + name = idName bndr + uniq = getUnique name + vtbl_lbl = mkVecTblLabel uniq + + deflt_lbl :: CAddrMode + deflt_lbl = case tagged_alt_absCs of + (DEFAULT, abs_c) : _ -> get_block_label abs_c + other -> mkIntCLit 0 + -- 'other' case: the simplifier might have eliminated a case + -- so we may have e.g. case xs of + -- [] -> e + -- In that situation the default should never be taken, + -- so we just use '0' (=> seg fault if used) + + mk_vector_entry :: ConTag -> CAddrMode + mk_vector_entry tag + = case [ absC | (DataAlt d, absC) <- tagged_alt_absCs, dataConTag d == tag ] of + -- The comprehension neatly, and correctly, ignores the DEFAULT + [] -> deflt_lbl + [abs_c] -> get_block_label abs_c + _ -> panic "mkReturnVector: too many" + + get_block_label (CCodeBlock lbl _) = CLbl lbl CodePtrRep \end{code}