-module CmmCPS (cmmCPS) where
+module CmmCPS (
+ -- | Converts C-- with full proceedures and parameters
+ -- to a CPS transformed C-- with the stack made manifest.
+ cmmCPS
+) where
#include "HsVersions.h"
import CmmLint
import PprCmm
-import Dataflow (fixedpoint)
+import Dataflow
import CmmLive
+import CmmBrokenBlock
+import CmmProcPoint
+import CmmCallConv
import MachOp
import ForeignCall
import IO
import Data.List
+-----------------------------------------------------------------------------
+-- |Top level driver for the CPS pass
+-----------------------------------------------------------------------------
+cmmCPS :: DynFlags -- ^ Dynamic flags: -dcmm-lint -ddump-cps-cmm
+ -> [Cmm] -- ^ Input C-- with Proceedures
+ -> IO [Cmm] -- ^ 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
+
+-----------------------------------------------------------------------------
+-- |CPS a single CmmTop (proceedure)
+-- Only 'CmmProc' are transformed 'CmmData' will be left alone.
+-----------------------------------------------------------------------------
+
+cpsProc :: UniqSupply
+ -> CmmTop -- ^Input proceedure
+ -> [CmmTop] -- ^Output proceedure and continuations
+cpsProc uniqSupply x@(CmmData _ _) = [x]
+cpsProc uniqSupply x@(CmmProc info_table ident params blocks) = cps_procs
+ where
+ uniqes :: [[Unique]]
+ uniqes = map uniqsFromSupply $ listSplitUniqSupply uniqSupply
+
+ -- Break the block at each function call.
+ -- The part after the function call will have to become a continuation.
+ broken_blocks :: [BrokenBlock]
+ broken_blocks =
+ concat $ zipWith3 breakBlock uniqes blocks
+ (FunctionEntry ident params:repeat ControlEntry)
+
+ -- 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]
+ continuations = map (gatherBlocksIntoContinuation proc_points block_env)
+ (uniqSetToList proc_points)
+
+ -- Select the stack format on entry to each continuation.
+ --
+ -- This is an association list instead of a UniqFM because
+ -- CLabel's don't have a 'Uniqueable' instance.
+ formats :: [(CLabel, StackFormat)]
+ formats = selectStackFormat live continuations
+
+ -- Do the actual CPS transform.
+ cps_procs :: [CmmTop]
+ cps_procs = map (continuationToProc formats) continuations
+
--------------------------------------------------------------------------------
-- The format for the call to a continuation
-- and heap memory (not sure if that's usefull at all though, but it may
-- be worth exploring the design space).
-data BrokenBlock
- = BrokenBlock {
- brokenBlockId :: BlockId, -- Like a CmmBasicBlock
- brokenBlockEntry :: BlockEntryInfo,
- -- How this block can be entered
-
- brokenBlockStmts :: [CmmStmt],
- -- Like a CmmBasicBlock
- -- (but without the last statement)
-
- brokenBlockTargets :: [BlockId],
- -- Blocks that this block could
- -- branch to one either by conditional
- -- branches or via the last statement
-
- brokenBlockExit :: BlockExitInfo
- -- How the block can be left
- }
-
-
-data BlockEntryInfo
- = FunctionEntry -- Beginning of function
-
- | ContinuationEntry -- Return point of a call
- CmmFormals -- return values
- -- TODO:
- -- | ProcPointEntry -- no return values, but some live might end up as params or possibly in the frame
-
- | ControlEntry -- A label in the input
-
-data BlockExitInfo
- = ControlExit
- BlockId -- next block (must be a ControlEntry)
-
- | ReturnExit
- CmmActuals -- return values
-
- | TailCallExit
- CmmExpr -- the function to call
- CmmActuals -- arguments to call
-
- | CallExit
- BlockId -- next block after call (must be a ContinuationEntry)
- CmmCallTarget -- the function to call
- CmmFormals -- results from call (redundant with ContinuationEntry)
- CmmActuals -- arguments to call
- (Maybe [GlobalReg]) -- registers that must be saved (TODO)
- -- TODO: | ProcPointExit (needed?)
-
+continuationLabel (Continuation _ _ l _ _) = l
+data Continuation =
+ Continuation
+ Bool -- True => Function entry, False => Continuation/return point
+ [CmmStatic] -- Info table, may be empty
+ CLabel -- Used to generate both info & entry labels
+ CmmFormals -- Argument locals live on entry (C-- procedure params)
+ [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.
+
+-- Describes the layout of a stack frame for a continuation
data StackFormat
= StackFormat
- BlockId {- block that is the start of the continuation. may or may not be the current block -}
- WordOff {- total frame size -}
- [(CmmReg, WordOff)] {- local reg offsets from stack top -}
+ (Maybe CLabel) -- The label occupying the top slot
+ WordOff -- Total frame size in words
+ [(CmmReg, WordOff)] -- 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 block =
+ if sizeUniqSet current_targets == sizeUniqSet new_targets
+ then current_targets
+ else foldl (collectNonProcPointTargets proc_points blocks) new_targets targets
+ where
+ block' = lookupWithDefaultUFM blocks (panic "TODO") block
+ 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 $ (mkUniqSet $ brokenBlockTargets block') `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 ::
+ UniqSet BlockId -> BlockEnv BrokenBlock
+ -> BlockId -> Continuation
+gatherBlocksIntoContinuation proc_points blocks start =
+ Continuation is_entry info_table clabel params body
+ where
+ children = (collectNonProcPointTargets proc_points blocks (unitUniqSet start) start) `delOneFromUniqSet` start
+ start_block = lookupWithDefaultUFM blocks (panic "TODO") start
+ children_blocks = map (lookupWithDefaultUFM blocks (panic "TODO")) (uniqSetToList children)
+ body = start_block : children_blocks
+ info_table = [] -- TODO
+ start_block_entry = brokenBlockEntry start_block
+ is_entry = case start_block_entry of
+ FunctionEntry _ _ -> True
+ _ -> False
+ 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 -> [] -- TODO: it's a proc-point, we could pass lives in parameter registers
+
--------------------------------------------------------------------------------
-- For now just select the continuation orders in the order they are in the set with no gaps
-selectStackFormat2 :: BlockEnv CmmLive -> [BrokenBlock] -> BlockEnv StackFormat
-selectStackFormat2 live blocks = fixedpoint dependants update (map brokenBlockId blocks) emptyUFM where
- blocks_ufm = listToUFM $ map (\b -> (brokenBlockId b, b)) blocks
- dependants ident =
- brokenBlockTargets $ lookupWithDefaultUFM blocks_ufm (panic "TODO") ident
- update ident cause formats =
- let BrokenBlock _ entry _ _ _ = lookupWithDefaultUFM blocks_ufm (panic "unknown BlockId in selectStackFormat:live") ident in
- case cause of
- -- Propagate only to blocks entered by branches (not function entry blocks or continuation entry blocks)
- Just cause_name ->
- let cause_format = lookupWithDefaultUFM formats (panic "update signaled for block not in format") cause_name
- in case entry of
- ControlEntry -> Just $ addToUFM formats ident cause_format
- FunctionEntry -> Nothing
- ContinuationEntry _ -> Nothing
- -- Do initial calculates for function blocks
- Nothing ->
- case entry of
- ControlEntry -> Nothing
- FunctionEntry -> Just $ addToUFM formats ident $ StackFormat ident 0 []
- ContinuationEntry _ -> Just $ addToUFM formats ident $ live_to_format ident $ lookupWithDefaultUFM live (panic "TODO") ident
- live_to_format label live =
- foldl extend_format
- (StackFormat label retAddrSizeW [])
- (uniqSetToList live)
- extend_format :: StackFormat -> LocalReg -> StackFormat
- extend_format (StackFormat block size offsets) reg =
- StackFormat block (slot_size reg + size) ((CmmLocal reg, size) : offsets)
-
-slot_size reg = ((machRepByteWidth (localRegRep reg) - 1) `div` wORD_SIZE) + 1
-
-constructContinuation2 :: BlockEnv StackFormat -> BrokenBlock -> CmmBasicBlock
-constructContinuation2 formats (BrokenBlock ident entry stmts _ exit) =
- BasicBlock ident (prefix++stmts++postfix)
+selectStackFormat :: BlockEnv CmmLive -> [Continuation] -> [(CLabel, StackFormat)]
+selectStackFormat live continuations =
+ map (\c -> (continuationLabel c, selectStackFormat' c)) continuations
where
- curr_format = lookupWithDefaultUFM formats (panic $ "format: unknown block " ++ (showSDoc $ ppr $ getUnique ident)) ident
- prefix = case entry of
- ControlEntry -> []
- FunctionEntry -> []
- ContinuationEntry formals -> unpack_continuation curr_format
- postfix = case exit of
- ControlExit next -> [CmmBranch next]
- ReturnExit arguments -> exit_function curr_format (CmmLoad (CmmReg spReg) wordRep) arguments
- TailCallExit target arguments -> exit_function curr_format target arguments
- -- TODO: do something about global saves
- CallExit next (CmmForeignCall target CmmCallConv) results arguments saves ->
- let cont_format = lookupWithDefaultUFM formats (panic $ "format: unknown block " ++ (showSDoc $ ppr $ getUnique next)) next
- in pack_continuation curr_format cont_format ++
- [CmmJump target arguments]
- CallExit next _ results arguments saves -> panic "unimplemented CmmCall"
+ selectStackFormat' (Continuation True info_table label formals blocks) =
+ StackFormat (Just label) 0 []
+ selectStackFormat' (Continuation False info_table 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 live_to_format label formals $ lookupWithDefaultUFM live unknown_block ident
+
+ live_to_format :: CLabel -> CmmFormals -> CmmLive -> StackFormat
+ live_to_format label formals live =
+ foldl extend_format
+ (StackFormat (Just label) retAddrSizeW [])
+ (uniqSetToList (live `minusUniqSet` mkUniqSet (cmmFormalsToLiveLocals formals)))
+
+ extend_format :: StackFormat -> LocalReg -> StackFormat
+ extend_format (StackFormat label size offsets) reg =
+ StackFormat label (slot_size reg + size) ((CmmLocal reg, size) : offsets)
+
+ slot_size :: LocalReg -> Int
+ slot_size reg = ((machRepByteWidth (localRegRep reg) - 1) `div` wORD_SIZE) + 1
+
+ unknown_block = panic "unknown BlockId in selectStackFormat"
+
+continuationToProc :: [(CLabel, StackFormat)] -> Continuation -> CmmTop
+continuationToProc formats (Continuation is_entry info label formals blocks) =
+ CmmProc info label formals (map (continuationToProc' label formats) blocks)
+ where
+ continuationToProc' :: CLabel -> [(CLabel, StackFormat)] -> BrokenBlock
+ -> CmmBasicBlock
+ continuationToProc' curr_ident formats (BrokenBlock ident entry stmts _ exit) =
+ BasicBlock ident (prefix++stmts++postfix)
+ where
+ curr_format = maybe unknown_block id $ lookup curr_ident formats
+ unknown_block = panic "unknown BlockId in continuationToProc"
+ prefix = case entry of
+ ControlEntry -> []
+ FunctionEntry _ _ -> []
+ ContinuationEntry formals ->
+ unpack_continuation curr_format
+ postfix = case exit of
+ FinalBranch next -> [CmmBranch next]
+ FinalSwitch expr targets -> [CmmSwitch expr targets]
+ FinalReturn arguments ->
+ exit_function curr_format
+ (CmmLoad (CmmReg spReg) wordRep)
+ arguments
+ FinalJump target arguments ->
+ exit_function curr_format target arguments
+ FinalCall next (CmmForeignCall target CmmCallConv)
+ results arguments ->
+ pack_continuation curr_format cont_format ++
+ [CmmJump target arguments]
+ where
+ cont_format = maybe unknown_block id $
+ lookup (mkReturnPtLabel $ getUnique next) formats
+ FinalCall next _ results arguments -> panic "unimplemented CmmCall"
--------------------------------------------------------------------------------
-- Functions that generate CmmStmt sequences
exit_function :: StackFormat -> CmmExpr -> CmmActuals -> [CmmStmt]
exit_function (StackFormat curr_id curr_frame_size curr_offsets) target arguments
= adjust_spReg ++ jump where
- adjust_spReg = [
- CmmAssign spReg
- (CmmRegOff spReg (curr_frame_size*wORD_SIZE))]
+ adjust_spReg =
+ if curr_frame_size == 0
+ then []
+ else [CmmAssign spReg
+ (CmmRegOff spReg (curr_frame_size*wORD_SIZE))]
jump = [CmmJump target arguments]
enter_function :: WordOff -> [CmmStmt]
(CmmMachOp (MO_U_Lt $ cmmRegRep spReg)
[CmmRegOff spReg max_frame_size, CmmReg spLimReg])
gc_block]
- gc_block = undefined -- TODO: get stack and heap checks to go to same
+ gc_block = panic "gc_check not implemented" -- TODO: get stack and heap checks to go to same
-- TODO: fix branches to proc point (we have to insert a new block to marshel the continuation)
pack_continuation :: StackFormat -> StackFormat -> [CmmStmt]
spReg (wORD_SIZE*(curr_frame_size - cont_frame_size + offset)))
(CmmReg reg)
| (reg, offset) <- cont_offsets]
- set_stack_header = -- TODO: only set when needed
- [CmmStore (CmmRegOff spReg (wORD_SIZE*(curr_frame_size - cont_frame_size))) continuation_function]
- continuation_function = CmmLit $ CmmLabel $ mkReturnPtLabel $ getUnique cont_id
+ needs_header =
+ case (curr_id, cont_id) of
+ (Just x, Just y) -> x /= y
+ _ -> isJust cont_id
+ set_stack_header =
+ if not needs_header
+ then []
+ else [CmmStore (CmmRegOff spReg (wORD_SIZE*(curr_frame_size - cont_frame_size))) continuation_function]
+ continuation_function = CmmLit $ CmmLabel $ fromJust cont_id
adjust_spReg =
if curr_frame_size == cont_frame_size
then []
(CmmLoad (CmmRegOff spReg (wORD_SIZE*offset)) (cmmRegRep reg))
| (reg, offset) <- curr_offsets]
------------------------------------------------------------------------------
--- Breaking basic blocks on function calls
------------------------------------------------------------------------------
-
------------------------------------------------------------------------------
--- Takes a basic block and returns a list of basic blocks that
--- each have at most 1 CmmCall in them which must occur at the end.
--- Also returns with each basic block, the variables that will
--- be arguments to the continuation of the block once the call (if any)
--- returns.
-
-breakBlock uniques (BasicBlock ident stmts) entry =
- breakBlock' uniques ident entry [] [] stmts where
- breakBlock' uniques current_id entry exits accum_stmts stmts =
- case stmts of
- [] -> panic "block doesn't end in jump, goto or return"
- [CmmJump target arguments] ->
- [BrokenBlock current_id entry accum_stmts exits
- (TailCallExit target arguments)]
- [CmmReturn arguments] ->
- [BrokenBlock current_id entry accum_stmts exits
- (ReturnExit arguments)]
- [CmmBranch target] ->
- [BrokenBlock current_id entry accum_stmts (target:exits)
- (ControlExit target)]
- (CmmJump _ _:_) ->
- panic "jump in middle of block"
- (CmmReturn _:_) ->
- panic "return in middle of block"
- (CmmBranch _:_) ->
- panic "branch in middle of block"
- (CmmSwitch _ _:_) ->
- panic "switch in block not implemented"
- (CmmCall target results arguments saves:stmts) ->
- let new_id = BlockId $ head uniques
- rest = breakBlock' (tail uniques) new_id (ContinuationEntry results) [] [] stmts
- in BrokenBlock current_id entry accum_stmts (new_id:exits)
- (CallExit new_id target results arguments saves) : rest
- (s@(CmmCondBranch test target):stmts) ->
- breakBlock' uniques current_id entry (target:exits) (accum_stmts++[s]) stmts
- (s:stmts) ->
- breakBlock' uniques current_id entry exits (accum_stmts++[s]) stmts
-
------------------------------------------------------------------------------
-cmmBlockFromBrokenBlock :: BrokenBlock -> CmmBasicBlock
-cmmBlockFromBrokenBlock (BrokenBlock ident _ stmts _ exit) = BasicBlock ident (stmts++exit_stmt)
- where
- exit_stmt =
- case exit of
- ControlExit target -> [CmmBranch target]
- ReturnExit arguments -> [CmmReturn arguments]
- TailCallExit target arguments -> [CmmJump target arguments]
- CallExit branch_target call_target results arguments saves -> [CmmCall call_target results arguments saves, CmmBranch branch_target]
-
------------------------------------------------------------------------------
--- CPS a single CmmTop (proceedure)
------------------------------------------------------------------------------
-
-cpsProc :: UniqSupply -> CmmTop -> [CmmTop]
-cpsProc uniqSupply x@(CmmData _ _) = [x]
-cpsProc uniqSupply x@(CmmProc info_table ident params blocks) =
- [CmmProc info_table ident params $ map (constructContinuation2 formats) broken_blocks]
- where
- uniqes :: [[Unique]]
- uniqes = map uniqsFromSupply $ listSplitUniqSupply uniqSupply
-
- broken_blocks :: [BrokenBlock]
- broken_blocks = concat $ zipWith3 breakBlock uniqes blocks (FunctionEntry:repeat ControlEntry)
-
- live :: BlockEntryLiveness
- live = cmmLiveness $ map cmmBlockFromBrokenBlock broken_blocks
-
- -- TODO: branches for proc points
- -- TODO: let blocks_with_live = map (cmmLivenessComment live . snd) broken_blocks
-
- formats :: BlockEnv StackFormat -- Stack format on entry
- formats = selectStackFormat2 live broken_blocks
-
-
---------------------------------------------------------------------------------
-cmmCPS :: DynFlags
- -> [Cmm] -- C-- with Proceedures
- -> IO [Cmm] -- 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: check for use of branches to non-existant blocks
- -- TODO: check for use of Sp, SpLim, R1, R2, etc.
- -- continuationC <- return abstractC
- -- TODO: find out if it is valid to create a new unique source like this
- uniqSupply <- mkSplitUniqSupply 'p'
- let supplies = listSplitUniqSupply uniqSupply
- let continuationC = zipWith (\s (Cmm c) -> Cmm $ concat $ zipWith (cpsProc) (listSplitUniqSupply s) c) supplies abstractC
-
- dumpIfSet_dyn dflags Opt_D_dump_cps_cmm "CPS Cmm" (pprCmms continuationC)
- -- TODO: add option to dump Cmm to file
- return continuationC