-cmmCPS :: DynFlags -- ^ Dynamic flags: -dcmm-lint -ddump-cps-cmm
- -> [GenCmm CmmStatic CmmInfo CmmStmt] -- ^ Input C-- with Proceedures
- -> IO [GenCmm CmmStatic [CmmStatic] CmmStmt] -- ^ Output CPS transformed C--
-cmmCPS dflags abstractC = do
- when (dopt Opt_DoCmmLinting dflags) $
- do showPass dflags "CmmLint"
- case firstJust $ map cmmLint abstractC of
- Just err -> do printDump err
- ghcExit dflags 1
- Nothing -> return ()
- showPass dflags "CPS"
-
- -- TODO: more lint checking
- -- check for use of branches to non-existant blocks
- -- check for use of Sp, SpLim, R1, R2, etc.
-
- uniqSupply <- mkSplitUniqSupply 'p'
- let supplies = listSplitUniqSupply uniqSupply
- let doCpsProc s (Cmm c) =
- Cmm $ concat $ zipWith cpsProc (listSplitUniqSupply s) c
- let continuationC = zipWith doCpsProc supplies abstractC
-
- dumpIfSet_dyn dflags Opt_D_dump_cps_cmm "CPS Cmm" (pprCmms continuationC)
-
- -- TODO: add option to dump Cmm to file
-
- return continuationC
-
-stg_gc_gen = mkRtsApFastLabel SLIT("gen_cg_TODO") --panic "Need the label for gc"
-make_gc_block block_id fun_label formals safety = BasicBlock block_id stmts
- where
- stmts = [CmmCall stg_gc_gen_target [] [] safety,
- CmmJump fun_expr actuals]
- stg_gc_gen_target =
- CmmForeignCall (CmmLit (CmmLabel stg_gc_gen)) CmmCallConv
- actuals = map (\x -> (CmmReg (CmmLocal x), NoHint)) formals
- fun_expr = CmmLit (CmmLabel fun_label)
-
-make_gc_check stack_use gc_block =
- [CmmCondBranch
- (CmmMachOp (MO_U_Lt $ cmmRegRep spReg)
- [CmmReg stack_use, CmmReg spLimReg])
- gc_block]
-
-force_gc_block old_info stack_use block_id fun_label formals =
- case old_info of
- CmmNonInfo (Just existing) -> (old_info, [], make_gc_check stack_use existing)
- CmmInfo _ (Just existing) _ _ -> (old_info, [], make_gc_check stack_use existing)
- CmmNonInfo Nothing
- -> (CmmNonInfo (Just block_id),
- [make_gc_block block_id fun_label formals (CmmSafe NoC_SRT)],
- make_gc_check stack_use block_id)
- CmmInfo prof Nothing type_tag type_info
- -> (CmmInfo prof (Just block_id) type_tag type_info,
- [make_gc_block block_id fun_label formals (CmmSafe srt)],
- make_gc_check stack_use block_id)
- where
- srt = case type_info of
- ConstrInfo _ _ _ -> NoC_SRT
- FunInfo _ srt' _ _ _ _ -> srt'
- ThunkInfo _ srt' -> srt'
- ThunkSelectorInfo _ srt' -> srt'
- ContInfo _ srt' -> srt'
-
------------------------------------------------------------------------------
--- |CPS a single CmmTop (proceedure)
--- Only 'CmmProc' are transformed 'CmmData' will be left alone.
------------------------------------------------------------------------------
-
-cpsProc :: UniqSupply
- -> GenCmmTop CmmStatic CmmInfo CmmStmt -- ^Input proceedure
- -> [GenCmmTop CmmStatic [CmmStatic] CmmStmt] -- ^Output proceedure and continuations
-cpsProc uniqSupply (CmmData sec dat) = [CmmData sec dat]
-cpsProc uniqSupply (CmmProc info ident params blocks) = info_procs
- where
- (uniqSupply1, uniqSupply2) = splitUniqSupply uniqSupply
- uniques :: [[Unique]]
- uniques = map uniqsFromSupply $ listSplitUniqSupply uniqSupply1
- (gc_unique:stack_use_unique:info_uniques):adaptor_uniques:block_uniques = uniques
- proc_uniques = map (map uniqsFromSupply . listSplitUniqSupply) $ listSplitUniqSupply uniqSupply2
-
- stack_use = CmmLocal (LocalReg stack_use_unique (cmmRegRep spReg) KindPtr)
-
- -- TODO: doc
- forced_gc :: (CmmInfo, [CmmBasicBlock], [CmmStmt])
- forced_gc = force_gc_block info stack_use (BlockId gc_unique) ident params
- (forced_info, gc_blocks, check_stmts) = forced_gc
-
- forced_blocks =
- case blocks of
- (BasicBlock id stmts) : bs ->
- (BasicBlock id (check_stmts ++ stmts)) : (bs ++ gc_blocks)
- [] -> [] -- If there is no code then we don't need a stack check
-
- forced_gc_id = case forced_info of
- CmmNonInfo (Just x) -> x
- CmmInfo _ (Just x) _ _ -> x
-
- -- Break the block at each function call.
- -- The part after the function call will have to become a continuation.
- broken_blocks :: ([(BlockId, ContFormat)], [BrokenBlock])
- broken_blocks =
- (\x -> (concatMap fst x, concatMap snd x)) $
- zipWith3 (breakBlock [forced_gc_id])
- block_uniques
- forced_blocks
- (FunctionEntry forced_info ident params :
- repeat ControlEntry)
-
- f' = selectContinuations (fst broken_blocks)
- broken_blocks' = map (makeContinuationEntries f') $
- concat $
- zipWith (adaptBlockToFormat f')
- adaptor_uniques
- (snd broken_blocks)
-
- -- Calculate live variables for each broken block.
- --
- -- Nothing can be live on entry to the first block
- -- so we could take the tail, but for now we wont
- -- to help future proof the code.
- live :: BlockEntryLiveness
- live = cmmLiveness $ map cmmBlockFromBrokenBlock broken_blocks'
-
- -- Calculate which blocks must be made into full fledged procedures.
- proc_points :: UniqSet BlockId
- proc_points = calculateProcPoints broken_blocks'
-
- -- Construct a map so we can lookup a broken block by its 'BlockId'.
- block_env :: BlockEnv BrokenBlock
- block_env = blocksToBlockEnv broken_blocks'
-
- -- Group the blocks into continuations based on the set of proc-points.
- continuations :: [Continuation (Either C_SRT CmmInfo)]
- continuations = zipWith
- (gatherBlocksIntoContinuation live proc_points block_env)
- (uniqSetToList proc_points)
- (Just forced_gc_id : repeat Nothing) {-dead-}
-
- -- Select the stack format on entry to each continuation.
- -- Return the max stack offset and an association list
- --
- -- This is an association list instead of a UniqFM because
- -- CLabel's don't have a 'Uniqueable' instance.
- formats :: [(CLabel, -- key
- (CmmFormals, -- arguments
- Maybe CLabel, -- label in top slot
- [Maybe LocalReg]))] -- slots
- formats = selectContinuationFormat live continuations
-
- -- Do a little meta-processing on the stack formats such as
- -- getting the individual frame sizes and the maximum frame size
- formats' :: (WordOff, [(CLabel, ContinuationFormat)])
- formats' = processFormats formats continuations
-
- -- Update the info table data on the continuations with
- -- the selected stack formats.
- continuations' :: [Continuation CmmInfo]
- continuations' = map (applyContinuationFormat (snd formats')) continuations
-
- -- Do the actual CPS transform.
- cps_procs :: [CmmTop]
- cps_procs = zipWith (continuationToProc formats' stack_use) proc_uniques continuations'
-
- -- Convert the info tables from CmmInfo to [CmmStatic]
- -- We might want to put this in another pass eventually
- info_procs :: [RawCmmTop]
- info_procs = concat (zipWith mkInfoTable info_uniques cps_procs)
-
---------------------------------------------------------------------------------
-
--- The format for the call to a continuation
--- The fst is the arguments that must be passed to the continuation
--- by the continuation's caller.
--- The snd is the live values that must be saved on stack.
--- A Nothing indicates an ignored slot.
--- The head of each list is the stack top or the first parameter.
-
--- The format for live values for a particular continuation
--- All on stack for now.
--- Head element is the top of the stack (or just under the header).
--- Nothing means an empty slot.
--- Future possibilities include callee save registers (i.e. passing slots in register)
--- and heap memory (not sure if that's usefull at all though, but it may
--- be worth exploring the design space).
-
-continuationLabel (Continuation _ l _ _ _) = l
-data Continuation info =
- Continuation
- info -- Left <=> Continuation created by the CPS
- -- Right <=> Function or Proc point
- CLabel -- Used to generate both info & entry labels
- CmmFormals -- Argument locals live on entry (C-- procedure params)
- Bool -- ^ True <=> GC block so ignore stack size
- [BrokenBlock] -- Code, may be empty. The first block is
- -- the entry point. The order is otherwise initially
- -- unimportant, but at some point the code gen will
- -- fix the order.
-
- -- the BlockId of the first block does not give rise
- -- to a label. To jump to the first block in a Proc,
- -- use the appropriate CLabel.
-
-data ContinuationFormat
- = ContinuationFormat {
- continuation_formals :: CmmFormals,
- continuation_label :: Maybe CLabel, -- The label occupying the top slot
- continuation_frame_size :: WordOff, -- Total frame size in words (not including arguments)
- continuation_stack :: [Maybe LocalReg] -- local reg offsets from stack top
- }
-
--- A block can be a continuation of a call
--- A block can be a continuation of another block (w/ or w/o joins)
--- A block can be an entry to a function
-
------------------------------------------------------------------------------
-
-collectNonProcPointTargets ::
- UniqSet BlockId -> BlockEnv BrokenBlock
- -> UniqSet BlockId -> [BlockId] -> UniqSet BlockId
-collectNonProcPointTargets proc_points blocks current_targets new_blocks =
- if sizeUniqSet current_targets == sizeUniqSet new_targets
- then current_targets
- else foldl
- (collectNonProcPointTargets proc_points blocks)
- new_targets
- (map (:[]) targets)
- where
- blocks' = map (lookupWithDefaultUFM blocks (panic "TODO")) new_blocks
- targets =
- -- Note the subtlety that since the extra branch after a call
- -- will always be to a block that is a proc-point,
- -- this subtraction will always remove that case
- uniqSetToList $ (unionManyUniqSets $ map (mkUniqSet . brokenBlockTargets) blocks')
- `minusUniqSet` proc_points
- -- TODO: remove redundant uniqSetToList
- new_targets = current_targets `unionUniqSets` (mkUniqSet targets)
-
--- TODO: insert proc point code here
--- * Branches and switches to proc points may cause new blocks to be created
--- (or proc points could leave behind phantom blocks that just jump to them)
--- * Proc points might get some live variables passed as arguments
-
-gatherBlocksIntoContinuation ::
- BlockEntryLiveness -> UniqSet BlockId -> BlockEnv BrokenBlock
- -> BlockId -> Maybe BlockId -> Continuation (Either C_SRT CmmInfo)
-gatherBlocksIntoContinuation live proc_points blocks start gc =
- Continuation info_table clabel params is_gc_cont body
- where
- --start_and_gc = [start] -- : maybeToList gc
- --children = (collectNonProcPointTargets proc_points blocks (mkUniqSet start_and_gc) start_and_gc) `minusUniqSet` (mkUniqSet start_and_gc)
- children = (collectNonProcPointTargets proc_points blocks (unitUniqSet start) [start]) `minusUniqSet` (unitUniqSet start)
- start_block = lookupWithDefaultUFM blocks (panic "TODO") start
- unknown_block = panic "unknown block in gatherBlocksIntoContinuation"
- --gc_block = map (lookupWithDefaultUFM blocks (panic "TODO)"))
- -- (maybeToList gc)
- children_blocks = map (lookupWithDefaultUFM blocks (panic "TODO")) (uniqSetToList children)
- body = start_block : {-gc_block ++ -} children_blocks
-
- -- We can't properly annotate the continuation's stack parameters
- -- at this point because this is before stack selection
- -- but we want to keep the C_SRT around so we use 'Either'.
- info_table = case start_block_entry of
- FunctionEntry info _ _ -> Right info
- ContinuationEntry _ srt _ -> Left srt
- ControlEntry -> Right (CmmNonInfo Nothing)
-
- is_gc_cont = case start_block_entry of
- FunctionEntry _ _ _ -> False
- ContinuationEntry _ _ gc_cont -> gc_cont
- ControlEntry -> False
-
- start_block_entry = brokenBlockEntry start_block
- clabel = case start_block_entry of
- FunctionEntry _ label _ -> label
- _ -> mkReturnPtLabel $ getUnique start
- params = case start_block_entry of
- FunctionEntry _ _ args -> args
- ContinuationEntry args _ _ -> args
- ControlEntry ->
- uniqSetToList $
- lookupWithDefaultUFM live unknown_block start
- -- it's a proc-point, pass lives in parameter registers
-
---------------------------------------------------------------------------------
--- For now just select the continuation orders in the order they are in the set with no gaps
-
-selectContinuationFormat :: BlockEnv CmmLive
- -> [Continuation (Either C_SRT CmmInfo)]
- -> [(CLabel, (CmmFormals, Maybe CLabel, [Maybe LocalReg]))]
-selectContinuationFormat live continuations =
- map (\c -> (continuationLabel c, selectContinuationFormat' c)) continuations
- where
- -- User written continuations
- selectContinuationFormat' (Continuation
- (Right (CmmInfo _ _ _ (ContInfo format srt)))
- label formals _ _) =
- (formals, Just label, format)
- -- Either user written non-continuation code
- -- or CPS generated proc-points
- selectContinuationFormat' (Continuation (Right _) _ formals _ _) =
- (formals, Nothing, [])
- -- CPS generated continuations
- selectContinuationFormat' (Continuation (Left srt) label formals _ blocks) =
- -- TODO: assumes the first block is the entry block
- let ident = brokenBlockId $ head blocks -- TODO: CLabel isn't a uniquable, but we need a better way than this
- in (formals,
- Just label,
- map Just $ uniqSetToList $
- lookupWithDefaultUFM live unknown_block ident)
-
- unknown_block = panic "unknown BlockId in selectContinuationFormat"
-
-processFormats :: [(CLabel, (CmmFormals, Maybe CLabel, [Maybe LocalReg]))]
- -> [Continuation (Either C_SRT CmmInfo)]
- -> (WordOff, [(CLabel, ContinuationFormat)])
-processFormats formats continuations = (max_size, formats')
- where
- max_size = maximum $
- 0 : map (continuationMaxStack formats') continuations
- formats' = map make_format formats
- make_format (label, (formals, top, stack)) =
- (label,
- ContinuationFormat {
- continuation_formals = formals,
- continuation_label = top,
- continuation_frame_size = stack_size stack +
- if isJust top
- then label_size
- else 0,
- continuation_stack = stack })
-
- -- TODO: get rid of "+ 1" etc.
- label_size = 1 :: WordOff
-
- stack_size [] = 0
- stack_size (Nothing:formats) = 1 + stack_size formats -- one dead word
- stack_size (Just reg:formats) = width + stack_size formats
- where
- width = machRepByteWidth (localRegRep reg) `quot` wORD_SIZE
- -- TODO: it would be better if we had a machRepWordWidth
-
-continuationMaxStack :: [(CLabel, ContinuationFormat)]
- -> Continuation a
- -> WordOff
-continuationMaxStack _ (Continuation _ _ _ True _) = 0
-continuationMaxStack formats (Continuation _ label _ False blocks) =
- max_arg_size + continuation_frame_size stack_format
- where
- stack_format = maybe unknown_format id $ lookup label formats
- unknown_format = panic "Unknown format in continuationMaxStack"
-
- max_arg_size = maximum $ 0 : map block_max_arg_size blocks
-
- block_max_arg_size block =
- maximum (final_arg_size (brokenBlockExit block) :
- map stmt_arg_size (brokenBlockStmts block))
-
- final_arg_size (FinalReturn args) =
- argumentsSize (cmmExprRep . fst) args
- final_arg_size (FinalJump _ args) =
- argumentsSize (cmmExprRep . fst) args
- final_arg_size (FinalCall next _ _ args _ True) = 0
- final_arg_size (FinalCall next _ _ args _ False) =
- -- We have to account for the stack used when we build a frame
- -- for the *next* continuation from *this* continuation
- argumentsSize (cmmExprRep . fst) args +
- continuation_frame_size next_format
- where
- next_format = maybe unknown_format id $ lookup next' formats
- next' = mkReturnPtLabel $ getUnique next
-
- final_arg_size _ = 0
-
- stmt_arg_size (CmmJump _ args) =
- argumentsSize (cmmExprRep . fst) args
- stmt_arg_size (CmmCall _ _ _ (CmmSafe _)) =
- panic "Safe call in processFormats"
- stmt_arg_size (CmmReturn _) =
- panic "CmmReturn in processFormats"
- stmt_arg_size _ = 0
-
------------------------------------------------------------------------------
-applyContinuationFormat :: [(CLabel, ContinuationFormat)]
- -> Continuation (Either C_SRT CmmInfo)
- -> Continuation CmmInfo
-
--- User written continuations
-applyContinuationFormat formats (Continuation
- (Right (CmmInfo prof gc tag (ContInfo _ srt)))
- label formals is_gc blocks) =
- Continuation (CmmInfo prof gc tag (ContInfo format srt))
- label formals is_gc blocks
- where
- format = continuation_stack $ maybe unknown_block id $ lookup label formats
- unknown_block = panic "unknown BlockId in applyContinuationFormat"
-
--- Either user written non-continuation code or CPS generated proc-point
-applyContinuationFormat formats (Continuation
- (Right info) label formals is_gc blocks) =
- Continuation info label formals is_gc blocks
-
--- CPS generated continuations
-applyContinuationFormat formats (Continuation
- (Left srt) label formals is_gc blocks) =
- Continuation (CmmInfo prof gc tag (ContInfo (continuation_stack $ format) srt))
- label formals is_gc blocks
- where
- gc = Nothing -- Generated continuations never need a stack check
- -- TODO prof: this is the same as the current implementation
- -- but I think it could be improved
- prof = ProfilingInfo zeroCLit zeroCLit
- tag = rET_SMALL -- cmmToRawCmm may convert it to rET_BIG
- format = maybe unknown_block id $ lookup label formats
- unknown_block = panic "unknown BlockId in applyContinuationFormat"
-
------------------------------------------------------------------------------
-continuationToProc :: (WordOff, [(CLabel, ContinuationFormat)])
- -> CmmReg
- -> [[Unique]]
- -> Continuation CmmInfo
- -> CmmTop
-continuationToProc (max_stack, formats) stack_use uniques
- (Continuation info label formals _ blocks) =
- CmmProc info label formals (concat $ zipWith3 continuationToProc' uniques blocks (True : repeat False))
- where
- curr_format = maybe unknown_block id $ lookup label formats
- unknown_block = panic "unknown BlockId in continuationToProc"
- curr_stack = continuation_frame_size curr_format
- arg_stack = argumentsSize localRegRep formals
-
- param_stmts :: [CmmStmt]
- param_stmts = function_entry curr_format
-
- gc_stmts :: [CmmStmt]
- gc_stmts =
- case info of
- CmmInfo _ (Just gc_block) _ _ ->
- gc_stack_check' stack_use arg_stack (max_stack - curr_stack)
- CmmInfo _ Nothing _ _ ->
- panic "continuationToProc: missing GC block"
- CmmNonInfo (Just gc_block) ->
- gc_stack_check' stack_use arg_stack (max_stack - curr_stack)
- CmmNonInfo Nothing ->
- panic "continuationToProc: missing non-info GC block"
-
--- At present neither the Cmm parser nor the code generator
--- produce code that will allow the target of a CmmCondBranch
--- or a CmmSwitch to become a continuation or a proc-point.
--- If future revisions, might allow these to happen
--- then special care will have to be take to allow for that case.
- continuationToProc' :: [Unique]
- -> BrokenBlock
- -> Bool
- -> [CmmBasicBlock]
- continuationToProc' uniques (BrokenBlock ident entry stmts _ exit) is_entry =
- prefix_blocks ++ [main_block]
- where
- prefix_blocks =
- case gc_prefix ++ param_prefix of
- [] -> []
- entry_stmts -> [BasicBlock prefix_id
- (entry_stmts ++ [CmmBranch ident])]
-
- prefix_unique : call_uniques = uniques
- toCLabel = mkReturnPtLabel . getUnique
-
- block_for_branch unique next
- | (Just cont_format) <- lookup (toCLabel next) formats
- = let
- new_next = BlockId unique
- cont_stack = continuation_frame_size cont_format
- arguments = map formal_to_actual (continuation_formals cont_format)
- in (new_next,
- [BasicBlock new_next $
- pack_continuation False curr_format cont_format ++
- tail_call (curr_stack - cont_stack)
- (CmmLit $ CmmLabel $ toCLabel next)
- arguments])
- | otherwise
- = (next, [])
-
- block_for_branch' :: Unique -> Maybe BlockId -> (Maybe BlockId, [CmmBasicBlock])
- block_for_branch' _ Nothing = (Nothing, [])
- block_for_branch' unique (Just next) = (Just new_next, new_blocks)
- where (new_next, new_blocks) = block_for_branch unique next
-
- main_block = BasicBlock ident (stmts ++ postfix_stmts)
- prefix_id = BlockId prefix_unique
- gc_prefix = case entry of
- FunctionEntry _ _ _ -> gc_stmts
- ControlEntry -> []
- ContinuationEntry _ _ _ -> []
- param_prefix = if is_entry
- then param_stmts
- else []
- postfix_stmts = case exit of
- FinalBranch next ->
- if (mkReturnPtLabel $ getUnique next) == label
- then [CmmBranch next]
- else case lookup (mkReturnPtLabel $ getUnique next) formats of
- Nothing -> [CmmBranch next]
- Just cont_format ->
- pack_continuation False curr_format cont_format ++
- tail_call (curr_stack - cont_stack)
- (CmmLit $ CmmLabel $ mkReturnPtLabel $ getUnique next)
- arguments
- where
- cont_stack = continuation_frame_size cont_format
- arguments = map formal_to_actual (continuation_formals cont_format)
- FinalSwitch expr targets -> [CmmSwitch expr targets]
- FinalReturn arguments ->
- tail_call curr_stack
- (CmmLoad (CmmReg spReg) wordRep)
- arguments
- FinalJump target arguments ->
- tail_call curr_stack target arguments
-
- -- A regular Cmm function call
- FinalCall next (CmmForeignCall target CmmCallConv)
- results arguments _ _ ->
- pack_continuation True curr_format cont_format ++
- tail_call (curr_stack - cont_stack)
- target arguments
- where
- cont_format = maybe unknown_block id $
- lookup (mkReturnPtLabel $ getUnique next) formats
- cont_stack = continuation_frame_size cont_format
-
- -- A safe foreign call
- FinalCall next (CmmForeignCall target conv)
- results arguments _ _ ->
- target_stmts ++
- foreignCall call_uniques' (CmmForeignCall new_target conv)
- results arguments
- where
- (call_uniques', target_stmts, new_target) =
- maybeAssignTemp call_uniques target
-
- -- A safe prim call
- FinalCall next (CmmPrim target)
- results arguments _ _ ->
- foreignCall call_uniques (CmmPrim target)
- results arguments
-
-formal_to_actual reg = (CmmReg (CmmLocal reg), NoHint)
-
-foreignCall :: [Unique] -> CmmCallTarget -> CmmHintFormals -> CmmActuals -> [CmmStmt]
-foreignCall uniques call results arguments =
- arg_stmts ++
- saveThreadState ++
- caller_save ++
- [CmmCall (CmmForeignCall suspendThread CCallConv)
- [ (id,PtrHint) ]
- [ (CmmReg (CmmGlobal BaseReg), PtrHint) ]
- CmmUnsafe,
- CmmCall call results new_args CmmUnsafe,
- CmmCall (CmmForeignCall resumeThread CCallConv)
- [ (new_base, PtrHint) ]
- [ (CmmReg (CmmLocal id), PtrHint) ]
- CmmUnsafe,
- -- Assign the result to BaseReg: we
- -- might now have a different Capability!
- CmmAssign (CmmGlobal BaseReg) (CmmReg (CmmLocal new_base))] ++
- caller_load ++
- loadThreadState tso_unique ++
- [CmmJump (CmmReg spReg) (map (formal_to_actual . fst) results)]
- where
- (_, arg_stmts, new_args) =
- loadArgsIntoTemps argument_uniques arguments
- (caller_save, caller_load) =
- callerSaveVolatileRegs (Just [{-only system regs-}])
- new_base = LocalReg base_unique (cmmRegRep (CmmGlobal BaseReg)) KindNonPtr
- id = LocalReg id_unique wordRep KindNonPtr
- tso_unique : base_unique : id_unique : argument_uniques = uniques
-
--- -----------------------------------------------------------------------------
--- Save/restore the thread state in the TSO
-
-suspendThread = CmmLit (CmmLabel (mkRtsCodeLabel SLIT("suspendThread")))
-resumeThread = CmmLit (CmmLabel (mkRtsCodeLabel SLIT("resumeThread")))
-
--- This stuff can't be done in suspendThread/resumeThread, because it
--- refers to global registers which aren't available in the C world.
-
-saveThreadState =
- -- CurrentTSO->sp = Sp;
- [CmmStore (cmmOffset stgCurrentTSO tso_SP) stgSp,
- closeNursery] ++
- -- and save the current cost centre stack in the TSO when profiling:
- if opt_SccProfilingOn
- then [CmmStore (cmmOffset stgCurrentTSO tso_CCCS) curCCS]
- else []
-
- -- CurrentNursery->free = Hp+1;
-closeNursery = CmmStore nursery_bdescr_free (cmmOffsetW stgHp 1)
-
-loadThreadState tso_unique =
- [
- -- tso = CurrentTSO;
- CmmAssign (CmmLocal tso) stgCurrentTSO,
- -- Sp = tso->sp;
- CmmAssign sp (CmmLoad (cmmOffset (CmmReg (CmmLocal tso)) tso_SP)
- wordRep),
- -- SpLim = tso->stack + RESERVED_STACK_WORDS;
- CmmAssign spLim (cmmOffsetW (cmmOffset (CmmReg (CmmLocal tso)) tso_STACK)
- rESERVED_STACK_WORDS)
- ] ++
- openNursery ++
- -- and load the current cost centre stack from the TSO when profiling:
- if opt_SccProfilingOn
- then [CmmStore curCCSAddr
- (CmmLoad (cmmOffset (CmmReg (CmmLocal tso)) tso_CCCS) wordRep)]
- else []
- where tso = LocalReg tso_unique wordRep KindNonPtr -- TODO FIXME NOW
-
-
-openNursery = [
- -- Hp = CurrentNursery->free - 1;
- CmmAssign hp (cmmOffsetW (CmmLoad nursery_bdescr_free wordRep) (-1)),
-
- -- HpLim = CurrentNursery->start +
- -- CurrentNursery->blocks*BLOCK_SIZE_W - 1;
- CmmAssign hpLim
- (cmmOffsetExpr
- (CmmLoad nursery_bdescr_start wordRep)
- (cmmOffset
- (CmmMachOp mo_wordMul [
- CmmMachOp (MO_S_Conv I32 wordRep)
- [CmmLoad nursery_bdescr_blocks I32],
- CmmLit (mkIntCLit bLOCK_SIZE)
- ])
- (-1)
- )
- )
- ]
-
-
-nursery_bdescr_free = cmmOffset stgCurrentNursery oFFSET_bdescr_free
-nursery_bdescr_start = cmmOffset stgCurrentNursery oFFSET_bdescr_start
-nursery_bdescr_blocks = cmmOffset stgCurrentNursery oFFSET_bdescr_blocks
-
-tso_SP = tsoFieldB oFFSET_StgTSO_sp
-tso_STACK = tsoFieldB oFFSET_StgTSO_stack
-tso_CCCS = tsoProfFieldB oFFSET_StgTSO_CCCS
-
--- The TSO struct has a variable header, and an optional StgTSOProfInfo in
--- the middle. The fields we're interested in are after the StgTSOProfInfo.
-tsoFieldB :: ByteOff -> ByteOff
-tsoFieldB off
- | opt_SccProfilingOn = off + sIZEOF_StgTSOProfInfo + fixedHdrSize * wORD_SIZE
- | otherwise = off + fixedHdrSize * wORD_SIZE
-
-tsoProfFieldB :: ByteOff -> ByteOff
-tsoProfFieldB off = off + fixedHdrSize * wORD_SIZE
-
-stgSp = CmmReg sp
-stgHp = CmmReg hp
-stgCurrentTSO = CmmReg currentTSO
-stgCurrentNursery = CmmReg currentNursery
-
-sp = CmmGlobal Sp
-spLim = CmmGlobal SpLim
-hp = CmmGlobal Hp
-hpLim = CmmGlobal HpLim
-currentTSO = CmmGlobal CurrentTSO
-currentNursery = CmmGlobal CurrentNursery
-
------------------------------------------------------------------------------
--- Functions that generate CmmStmt sequences
--- for packing/unpacking continuations
--- and entering/exiting functions
-
-tail_call :: WordOff -> CmmExpr -> CmmActuals -> [CmmStmt]
-tail_call spRel target arguments
- = store_arguments ++ adjust_spReg ++ jump where
- store_arguments =
- [stack_put spRel expr offset
- | ((expr, _), StackParam offset) <- argument_formats] ++
- [global_put expr global
- | ((expr, _), RegisterParam global) <- argument_formats]
- adjust_spReg =
- if spRel == 0
- then []
- else [CmmAssign spReg (CmmRegOff spReg (spRel*wORD_SIZE))]
- jump = [CmmJump target arguments]
-
- argument_formats = assignArguments (cmmExprRep . fst) arguments
-
-gc_stack_check' stack_use arg_stack max_frame_size =
- if max_frame_size > arg_stack
- then [CmmAssign stack_use (CmmRegOff spReg (-max_frame_size*wORD_SIZE))]
- else [CmmAssign stack_use (CmmReg spLimReg)]
- -- Trick the optimizer into eliminating the branch for us
-
-gc_stack_check :: BlockId -> WordOff -> [CmmStmt]
-gc_stack_check gc_block max_frame_size
- = check_stack_limit where
- check_stack_limit = [
- CmmCondBranch
- (CmmMachOp (MO_U_Lt $ cmmRegRep spReg)
- [CmmRegOff spReg (-max_frame_size*wORD_SIZE),
- CmmReg spLimReg])
- gc_block]
-
-
--- TODO: fix branches to proc point
--- (we have to insert a new block to marshel the continuation)
-
-
-pack_continuation :: Bool -- ^ Whether to set the top/header
- -- of the stack. We only need to
- -- set it if we are calling down
- -- as opposed to continuation
- -- adaptors.
- -> ContinuationFormat -- ^ The current format
- -> ContinuationFormat -- ^ The return point format
- -> [CmmStmt]
-pack_continuation allow_header_set
- (ContinuationFormat _ curr_id curr_frame_size _)
- (ContinuationFormat _ cont_id cont_frame_size live_regs)
- = store_live_values ++ set_stack_header where
- -- TODO: only save variables when actually needed
- -- (may be handled by latter pass)
- store_live_values =
- [stack_put spRel (CmmReg (CmmLocal reg)) offset
- | (reg, offset) <- cont_offsets]
- set_stack_header =
- if needs_header_set && allow_header_set
- then [stack_put spRel continuation_function 0]
- else []
-
- -- TODO: factor with function_entry and CmmInfo.hs(?)
- cont_offsets = mkOffsets label_size live_regs
-
- label_size = 1 :: WordOff
-
- mkOffsets size [] = []
- mkOffsets size (Nothing:regs) = mkOffsets (size+1) regs
- mkOffsets size (Just reg:regs) = (reg, size):mkOffsets (size + width) regs
- where
- width = machRepByteWidth (localRegRep reg) `quot` wORD_SIZE
- -- TODO: it would be better if we had a machRepWordWidth
-
- spRel = curr_frame_size - cont_frame_size
- continuation_function = CmmLit $ CmmLabel $ fromJust cont_id
- needs_header_set =
- case (curr_id, cont_id) of
- (Just x, Just y) -> x /= y
- _ -> isJust cont_id
-
--- Lazy adjustment of stack headers assumes all blocks
--- that could branch to eachother (i.e. control blocks)
--- have the same stack format (this causes a problem
--- only for proc-point).
-function_entry :: ContinuationFormat -> [CmmStmt]
-function_entry (ContinuationFormat formals _ _ live_regs)
- = load_live_values ++ load_args where
- -- TODO: only save variables when actually needed
- -- (may be handled by latter pass)
- load_live_values =
- [stack_get 0 reg offset
- | (reg, offset) <- curr_offsets]
- load_args =
- [stack_get 0 reg offset
- | (reg, StackParam offset) <- argument_formats] ++
- [global_get reg global
- | (reg, RegisterParam global) <- argument_formats]
-
- argument_formats = assignArguments (localRegRep) formals
-
- -- TODO: eliminate copy/paste with pack_continuation
- curr_offsets = mkOffsets label_size live_regs
-
- label_size = 1 :: WordOff
-
- mkOffsets size [] = []
- mkOffsets size (Nothing:regs) = mkOffsets (size+1) regs
- mkOffsets size (Just reg:regs) = (reg, size):mkOffsets (size + width) regs
- where
- width = machRepByteWidth (localRegRep reg) `quot` wORD_SIZE
- -- TODO: it would be better if we had a machRepWordWidth
-
------------------------------------------------------------------------------
--- Section: Stack and argument register puts and gets
------------------------------------------------------------------------------
--- TODO: document
-
--- |Construct a 'CmmStmt' that will save a value on the stack
-stack_put :: WordOff -- ^ Offset from the real 'Sp' that 'offset'
- -- is relative to (added to offset)
- -> CmmExpr -- ^ What to store onto the stack
- -> WordOff -- ^ Where on the stack to store it
- -- (positive <=> higher addresses)
- -> CmmStmt
-stack_put spRel expr offset =
- CmmStore (CmmRegOff spReg (wORD_SIZE*(spRel + offset))) expr
-
---------------------------------
--- |Construct a
-stack_get :: WordOff
- -> LocalReg
- -> WordOff
- -> CmmStmt
-stack_get spRel reg offset =
- CmmAssign (CmmLocal reg)
- (CmmLoad (CmmRegOff spReg (wORD_SIZE*(spRel + offset)))
- (localRegRep reg))
-global_put :: CmmExpr -> GlobalReg -> CmmStmt
-global_put expr global = CmmAssign (CmmGlobal global) expr
-global_get :: LocalReg -> GlobalReg -> CmmStmt
-global_get reg global = CmmAssign (CmmLocal reg) (CmmReg (CmmGlobal global))
-
+-- There are two complications here:
+-- 1. We need to compile the procedures in two stages because we need
+-- an analysis of the procedures to tell us what CAFs they use.
+-- The first stage returns a map from procedure labels to CAFs,
+-- along with a closure that will compute SRTs and attach them to
+-- the compiled procedures.
+-- The second stage is to combine the CAF information into a top-level
+-- CAF environment mapping non-static closures to the CAFs they keep live,
+-- then pass that environment to the closures returned in the first
+-- stage of compilation.
+-- 2. We need to thread the module's SRT around when the SRT tables
+-- are computed for each procedure.
+-- The SRT needs to be threaded because it is grown lazily.
+protoCmmCPS :: HscEnv -- Compilation env including
+ -- dynamic flags: -dcmm-lint -ddump-cps-cmm
+ -> (TopSRT, [Cmm]) -- SRT table and accumulating list of compiled procs
+ -> Cmm -- Input C-- with Procedures
+ -> IO (TopSRT, [Cmm]) -- Output CPS transformed C--
+protoCmmCPS hsc_env (topSRT, rst) (Cmm tops) =
+ do let dflags = hsc_dflags hsc_env
+ showPass dflags "CPSZ"
+ (cafEnvs, tops) <- liftM unzip $ mapM (cpsTop hsc_env) tops
+ let topCAFEnv = mkTopCAFInfo (concat cafEnvs)
+ (topSRT, tops) <- foldM (toTops hsc_env topCAFEnv) (topSRT, []) tops
+ let cmms = Cmm (reverse (concat tops))
+ dumpIfSet_dyn dflags Opt_D_dump_cps_cmm "Post CPS Cmm" (ppr cmms)
+ return (topSRT, cmms : rst)
+
+{- [Note global fuel]
+~~~~~~~~~~~~~~~~~~~~~
+The identity and the last pass are stored in
+mutable reference cells in an 'HscEnv' and are
+global to one compiler session.
+-}
+
+cpsTop :: HscEnv -> CmmTop -> IO ([(CLabel, CAFSet)], [(CAFSet, CmmTop)])
+cpsTop _ p@(CmmData {}) = return ([], [(Map.empty, p)])
+cpsTop hsc_env (CmmProc h@(TopInfo {stack_info=StackInfo {arg_space=entry_off}}) l g) =
+ do
+ -- Why bother doing it this early?
+ -- g <- dual_rewrite run Opt_D_dump_cmmz "spills and reloads"
+ -- (dualLivenessWithInsertion callPPs) g
+ -- g <- run $ insertLateReloads g -- Duplicate reloads just before uses
+ -- g <- dual_rewrite runOptimization Opt_D_dump_cmmz "Dead Assignment Elimination"
+ -- (removeDeadAssignmentsAndReloads callPPs) g
+ dump Opt_D_dump_cmmz "Pre common block elimination" g
+ g <- return $ elimCommonBlocks g
+ dump Opt_D_dump_cmmz "Post common block elimination" g
+
+ -- Any work storing block Labels must be performed _after_ elimCommonBlocks
+
+ ----------- Proc points -------------------
+ let callPPs = callProcPoints g
+ procPoints <- run $ minimalProcPointSet callPPs g
+ g <- run $ addProcPointProtocols callPPs procPoints g
+ dump Opt_D_dump_cmmz "Post Proc Points Added" g
+
+ ----------- Spills and reloads -------------------
+ g <-
+ -- pprTrace "pre Spills" (ppr g) $
+ dual_rewrite run Opt_D_dump_cmmz "spills and reloads"
+ (dualLivenessWithInsertion procPoints) g
+ -- Insert spills at defns; reloads at return points
+ g <-
+ -- pprTrace "pre insertLateReloads" (ppr g) $
+ runOptimization $ insertLateReloads g -- Duplicate reloads just before uses
+ dump Opt_D_dump_cmmz "Post late reloads" g
+ g <-
+ -- pprTrace "post insertLateReloads" (ppr g) $
+ dual_rewrite runOptimization Opt_D_dump_cmmz "Dead Assignment Elimination"
+ (removeDeadAssignmentsAndReloads procPoints) g
+ -- Remove redundant reloads (and any other redundant asst)
+
+ ----------- Debug only: add code to put zero in dead stack slots----
+ -- Debugging: stubbing slots on death can cause crashes early
+ g <- -- trace "post dead-assign elim" $
+ if opt_StubDeadValues then run $ stubSlotsOnDeath g else return g
+
+
+ --------------- Stack layout ----------------
+ slotEnv <- run $ liveSlotAnal g
+ let spEntryMap = getSpEntryMap entry_off g
+ mbpprTrace "live slot analysis results: " (ppr slotEnv) $ return ()
+ let areaMap = layout procPoints spEntryMap slotEnv entry_off g
+ mbpprTrace "areaMap" (ppr areaMap) $ return ()
+
+ ------------ Manifest the stack pointer --------
+ g <- run $ manifestSP spEntryMap areaMap entry_off g
+ dump Opt_D_dump_cmmz "after manifestSP" g
+ -- UGH... manifestSP can require updates to the procPointMap.
+ -- We can probably do something quicker here for the update...
+
+ ------------- Split into separate procedures ------------
+ procPointMap <- run $ procPointAnalysis procPoints g
+ dump Opt_D_dump_cmmz "procpoint map" procPointMap
+ gs <- run $ splitAtProcPoints l callPPs procPoints procPointMap
+ (CmmProc h l g)
+ mapM_ (dump Opt_D_dump_cmmz "after splitting") gs
+
+ ------------- More CAFs and foreign calls ------------
+ cafEnv <- run $ cafAnal g
+ let localCAFs = catMaybes $ map (localCAFInfo cafEnv) gs
+ mbpprTrace "localCAFs" (ppr localCAFs) $ return ()
+
+ gs <- run $ mapM (lowerSafeForeignCalls areaMap) gs
+ mapM_ (dump Opt_D_dump_cmmz "after lowerSafeForeignCalls") gs
+
+ -- NO MORE GRAPH TRANSFORMATION AFTER HERE -- JUST MAKING INFOTABLES
+ let gs' = map (setInfoTableStackMap slotEnv areaMap) gs
+ mapM_ (dump Opt_D_dump_cmmz "after setInfoTableStackMap") gs'
+ let gs'' = map (bundleCAFs cafEnv) gs'
+ mapM_ (dump Opt_D_dump_cmmz "after bundleCAFs") gs''
+ return (localCAFs, gs'')
+ where dflags = hsc_dflags hsc_env
+ mbpprTrace x y z = if dopt Opt_D_dump_cmmz dflags then pprTrace x y z else z
+ dump f txt g = dumpIfSet_dyn dflags f txt (ppr g)
+ -- Runs a required transformation/analysis
+ run = runInfiniteFuelIO (hsc_OptFuel hsc_env)
+ -- Runs an optional transformation/analysis (and should
+ -- thus be subject to optimization fuel)
+ runOptimization = runFuelIO (hsc_OptFuel hsc_env)
+
+ -- pass 'run' or 'runOptimization' for 'r'
+ dual_rewrite r flag txt pass g =
+ do dump flag ("Pre " ++ txt) g
+ g <- r $ pass g
+ dump flag ("Post " ++ txt) $ g
+ return g
+
+-- This probably belongs in CmmBuildInfoTables?
+-- We're just finishing the job here: once we know what CAFs are defined
+-- in non-static closures, we can build the SRTs.
+toTops :: HscEnv -> Map CLabel CAFSet -> (TopSRT, [[CmmTop]])
+ -> [(CAFSet, CmmTop)] -> IO (TopSRT, [[CmmTop]])
+toTops hsc_env topCAFEnv (topSRT, tops) gs =
+ do let setSRT (topSRT, rst) g =
+ do (topSRT, gs) <- setInfoTableSRT topCAFEnv topSRT g
+ return (topSRT, gs : rst)
+ (topSRT, gs') <- runFuelIO (hsc_OptFuel hsc_env) $ foldM setSRT (topSRT, []) gs
+ return (topSRT, concat gs' : tops)