CmmBuildInfoTables.hs 22.9 KB
Newer Older
1
{-# OPTIONS_GHC -XGADTs -XNoMonoLocalBinds #-}
2 3 4
-- Norman likes local bindings
-- If this module lives on I'd like to get rid of this flag in due course

5 6 7
-- Todo: remove

{-# OPTIONS_GHC -fno-warn-warnings-deprecations #-}
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
8
module CmmBuildInfoTables
9
    ( CAFSet, CAFEnv, cafAnal, localCAFInfo, mkTopCAFInfo
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
10 11
    , setInfoTableSRT, setInfoTableStackMap
    , TopSRT, emptySRT, srtToData
12
    , bundleCAFs
13 14
    , lowerSafeForeignCalls
    , cafTransfers, liveSlotTransfers)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
15 16 17 18 19 20 21
where

#include "HsVersions.h"

import Constants
import Digraph
import qualified Prelude as P
22
import Prelude hiding (succ)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
23 24 25 26 27
import Util (sortLe)

import BlockId
import Bitmap
import CLabel
28 29 30
import Cmm
import CmmDecl
import CmmExpr
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
31
import CmmStackLayout
32
import Module
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
33 34 35
import FastString
import ForeignCall
import IdInfo
Ian Lynagh's avatar
Ian Lynagh committed
36
import Data.List
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
37
import Maybes
38
import MkGraph as M
Ian Lynagh's avatar
Ian Lynagh committed
39
import Control.Monad
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
40
import Name
41
import OptimizationFuel
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
42 43 44 45 46 47
import Outputable
import SMRep
import StgCmmClosure
import StgCmmForeign
import StgCmmUtils
import UniqSupply
48 49

import Compiler.Hoopl
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
50

51 52 53 54
import Data.Map (Map)
import qualified Data.Map as Map
import qualified FiniteMap as Map

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
----------------------------------------------------------------
-- Building InfoTables


-----------------------------------------------------------------------
-- Stack Maps

-- Given a block ID, we return a representation of the layout of the stack,
-- as suspended before entering that block.
-- (For a return site to a function call, the layout does not include the
--  parameter passing area (or the "return address" on the stack)).
-- If the element is `Nothing`, then it represents a word of the stack that
-- does not contain a live pointer.
-- If the element is `Just` a register, then it represents a live spill slot
-- for a pointer; we assume that a pointer is the size of a word.
-- The head of the list represents the young end of the stack where the infotable
-- pointer for the block `Bid` is stored.
-- The infotable pointer itself is not included in the list.
-- Call areas are also excluded from the list: besides the stuff in the update
-- frame (and the return infotable), call areas should never be live across
-- function calls.

-- RTS Invariant: All pointers must be word-aligned because each bit in the bitmap
-- represents a word. Consequently, we have to be careful when we see a live slot
-- on the stack: if we have packed multiple sub-word values into a word,
-- we have to make sure that we only mark the entire word as a non-pointer.

-- Also, don't forget to stop at the old end of the stack (oldByte),
-- which may differ depending on whether there is an update frame.
84 85 86 87 88 89

type RegSlotInfo
   = ( Int	  -- Offset from oldest byte of Old area
     , LocalReg   -- The register
     , Int)       -- Width of the register

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
90 91
live_ptrs :: ByteOff -> BlockEnv SubAreaSet -> AreaMap -> BlockId -> [Maybe LocalReg]
live_ptrs oldByte slotEnv areaMap bid =
92 93 94 95 96
  -- pprTrace "live_ptrs for" (ppr bid <+> text (show oldByte ++ "-" ++ show youngByte) <+>
  --                           ppr liveSlots) $
  -- pprTrace ("stack layout for " ++ show bid ++ ": ") (ppr res) $ res
  res
  where res = reverse $ slotsToList youngByte liveSlots []
97 98 99 100 101 102 103
 
        slotsToList :: Int -> [RegSlotInfo] -> [Maybe LocalReg] -> [Maybe LocalReg]
        -- n starts at youngByte and is decremented down to oldByte
	-- Returns a list, one element per word, with 
	--    (Just r) meaning 'pointer register r is saved here', 
	--    Nothing  meaning 'non-pointer or empty'

104
        slotsToList n [] results | n == oldByte = results -- at old end of stack frame
105

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
106 107 108
        slotsToList n (s : _) _  | n == oldByte =
          pprPanic "slot left off live_ptrs" (ppr s <+> ppr oldByte <+>
               ppr n <+> ppr liveSlots <+> ppr youngByte)
109

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
110 111
        slotsToList n _ _ | n < oldByte =
          panic "stack slots not allocated on word boundaries?"
112

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
113 114 115 116 117 118 119 120 121 122
        slotsToList n l@((n', r, w) : rst) results =
          if n == (n' + w) then -- slot's young byte is at n
            ASSERT (not (isPtr r) ||
                    (n `mod` wORD_SIZE == 0 && w == wORD_SIZE)) -- ptrs must be aligned
            slotsToList next (dropWhile (non_ptr_younger_than next) rst)
                        (stack_rep : results)
          else slotsToList next (dropWhile (non_ptr_younger_than next) l)
                           (Nothing : results)
          where next = n - wORD_SIZE
                stack_rep = if isPtr r then Just r else Nothing
123

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
124
        slotsToList n [] results = slotsToList (n - wORD_SIZE) [] (Nothing : results)
125

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
126 127 128 129 130
        non_ptr_younger_than next (n', r, w) =
          n' + w > next &&
            ASSERT (not (isPtr r))
            True
        isPtr = isGcPtrType . localRegType
131 132

        liveSlots :: [RegSlotInfo]
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
133
        liveSlots = sortBy (\ (off,_,_) (off',_,_) -> compare off' off)
134
                           (Map.foldRightWithKey (\_ -> flip $ foldl add_slot) [] slots)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
135
                    
136
        add_slot :: [RegSlotInfo] -> SubArea -> [RegSlotInfo]
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
137 138
        add_slot rst (a@(RegSlot r@(LocalReg _ ty)), off, w) = 
          if off == w && widthInBytes (typeWidth ty) == w then
139
            (expectJust "add_slot" (Map.lookup a areaMap), r, w) : rst
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
140
          else panic "live_ptrs: only part of a variable live at a proc point"
141
        add_slot rst (CallArea Old, _, _) =
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
142 143
          rst -- the update frame (or return infotable) should be live
              -- would be nice to check that only that part of the callarea is live...
144
        add_slot rst ((CallArea _), _, _) =
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
145 146 147 148 149 150 151 152
          rst
          -- JD: THIS ISN'T CURRENTLY A CORRECTNESS PROBLEM, BUT WE SHOULD REALLY
          -- MAKE LIVENESS INFO AROUND CALLS MORE PRECISE -- FOR NOW, A 32-BIT
          -- FLOAT PADS OUT TO 64 BITS, BUT WE ASSUME THE WHOLE PARAMETER-PASSING
          -- AREA IS LIVE (WHICH IT ISN'T...).  WE SHOULD JUST PUT THE LIVE AREAS
          -- IN THE CALL NODES, WHICH SHOULD EVENTUALLY HAVE LIVE REGISTER AS WELL,
          -- SO IT'S ALL GOING IN THE SAME DIRECTION.
          -- pprPanic "CallAreas must not be live across function calls" (ppr bid <+> ppr c)
153 154

        slots :: SubAreaSet	 -- The SubAreaSet for 'bid'
155
        slots = expectJust "live_ptrs slots" $ mapLookup bid slotEnv
156
        youngByte = expectJust "live_ptrs bid_pos" $ Map.lookup (CallArea (Young bid)) areaMap
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
157

158 159 160 161
-- Construct the stack maps for a procedure _if_ it needs an infotable.
-- When wouldn't a procedure need an infotable? If it is a procpoint that
-- is not the successor of a call.
setInfoTableStackMap :: SlotEnv -> AreaMap -> CmmTop -> CmmTop
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
162
setInfoTableStackMap slotEnv areaMap
163 164 165
     t@(CmmProc (TopInfo {stack_info=StackInfo {updfr_space = Just updfr_off}}) _ (CmmGraph {g_entry = eid})) =
  updInfo (const (live_ptrs updfr_off slotEnv areaMap eid)) id t
setInfoTableStackMap _ _ t = t
166
                 
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183


-----------------------------------------------------------------------
-- SRTs

-- WE NEED AN EXAMPLE HERE.
-- IN PARTICULAR, WE NEED TO POINT OUT THE DISTINCTION BETWEEN
-- FUNCTIONS WITH STATIC CLOSURES AND THOSE THAT MUST BE CONSTRUCTED
-- DYNAMICALLY (AND HENCE CAN'T BE REFERENCED IN AN SRT).
-- IN THE LATTER CASE, WE HAVE TO TAKE ALL THE CAFs REFERENCED BY
-- THE CLOSURE AND INLINE THEM INTO ANY SRT THAT MAY MENTION THE CLOSURE.
-- (I.E. TAKE THE TRANSITIVE CLOSURE, but only for non-static closures).


-----------------------------------------------------------------------
-- Finding the CAFs used by a procedure

184
type CAFSet = Map CLabel ()
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
185 186 187 188
type CAFEnv = BlockEnv CAFSet

-- First, an analysis to find live CAFs.
cafLattice :: DataflowLattice CAFSet
189 190 191 192 193 194
cafLattice = DataflowLattice "live cafs" Map.empty add
  where add _ (OldFact old) (NewFact new) = case old `Map.union` new of
                                              new' -> (changeIf $ Map.size new' > Map.size old, new')

cafTransfers :: BwdTransfer CmmNode CAFSet
cafTransfers = mkBTransfer3 first middle last
195
  where first  _ live = live
196 197
        middle m live = foldExpDeep addCaf m live
        last   l live = foldExpDeep addCaf l (joinOutFacts cafLattice l live)
198 199 200 201 202
        addCaf e set = case e of
               CmmLit (CmmLabel c)              -> add c set
               CmmLit (CmmLabelOff c _)         -> add c set
               CmmLit (CmmLabelDiffOff c1 c2 _) -> add c1 $ add c2 set
               _ -> set
203
        add l s = if hasCAF l then Map.insert (cvtToClosureLbl l) () s else s
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
204

205 206
cafAnal :: CmmGraph -> FuelUniqSM CAFEnv
cafAnal g = liftM snd $ dataflowPassBwd g [] $ analBwd cafLattice cafTransfers
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
207 208 209 210 211 212 213 214 215

-----------------------------------------------------------------------
-- Building the SRTs

-- Description of the SRT for a given module.
-- Note that this SRT may grow as we greedily add new CAFs to it.
data TopSRT = TopSRT { lbl      :: CLabel
                     , next_elt :: Int -- the next entry in the table
                     , rev_elts :: [CLabel]
216
                     , elt_map  :: Map CLabel Int }
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
217 218 219 220 221 222 223 224
                        -- map: CLabel -> its last entry in the table
instance Outputable TopSRT where
  ppr (TopSRT lbl next elts eltmap) =
    text "TopSRT:" <+> ppr lbl <+> ppr next <+> ppr elts <+> ppr eltmap

emptySRT :: MonadUnique m => m TopSRT
emptySRT =
  do top_lbl <- getUniqueM >>= \ u -> return $ mkSRTLabel (mkFCallName u "srt") NoCafRefs
225
     return TopSRT { lbl = top_lbl, next_elt = 0, rev_elts = [], elt_map = Map.empty }
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
226 227

cafMember :: TopSRT -> CLabel -> Bool
228
cafMember srt lbl = Map.member lbl (elt_map srt)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
229 230

cafOffset :: TopSRT -> CLabel -> Maybe Int
231
cafOffset srt lbl = Map.lookup lbl (elt_map srt)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
232 233 234 235 236

addCAF :: CLabel -> TopSRT -> TopSRT
addCAF caf srt =
  srt { next_elt = last + 1
      , rev_elts = caf : rev_elts srt
237
      , elt_map  = Map.insert caf last (elt_map srt) }
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
238 239
    where last  = next_elt srt

240
srtToData :: TopSRT -> Cmm
241
srtToData srt = Cmm [CmmData RelocatableReadOnlyData (Statics (lbl srt) tbl)]
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
242 243 244 245 246 247 248 249 250 251
    where tbl = map (CmmStaticLit . CmmLabel) (reverse (rev_elts srt))

-- Once we have found the CAFs, we need to do two things:
-- 1. Build a table of all the CAFs used in the procedure.
-- 2. Compute the C_SRT describing the subset of CAFs live at each procpoint.
--
-- When building the local view of the SRT, we first make sure that all the CAFs are 
-- in the SRT. Then, if the number of CAFs is small enough to fit in a bitmap,
-- we make sure they're all close enough to the bottom of the table that the
-- bitmap will be able to cover all of them.
252
buildSRTs :: TopSRT -> Map CLabel CAFSet -> CAFSet ->
253
             FuelUniqSM (TopSRT, Maybe CmmTop, C_SRT)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
254
buildSRTs topSRT topCAFMap cafs =
255
  do let liftCAF lbl () z = -- get CAFs for functions without static closures
256 257
           case Map.lookup lbl topCAFMap of Just cafs -> z `Map.union` cafs
                                            Nothing   -> Map.insert lbl () z
258 259
         -- For each label referring to a function f without a static closure,
         -- replace it with the CAFs that are reachable from f.
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
260
         sub_srt topSRT localCafs =
261
           let cafs = Map.keys (Map.foldRightWithKey liftCAF Map.empty localCafs)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
262 263 264
               mkSRT topSRT =
                 do localSRTs <- procpointSRT (lbl topSRT) (elt_map topSRT) cafs
                    return (topSRT, localSRTs)
265
           in if length cafs > maxBmpSize then
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
266 267 268 269 270 271 272 273 274 275 276
                mkSRT (foldl add_if_missing topSRT cafs)
              else -- make sure all the cafs are near the bottom of the srt
                mkSRT (add_if_too_far topSRT cafs)
         add_if_missing srt caf =
           if cafMember srt caf then srt else addCAF caf srt
         -- If a CAF is more than maxBmpSize entries from the young end of the
         -- SRT, then we add it to the SRT again.
         -- (Note: Not in the SRT => infinitely far.)
         add_if_too_far srt@(TopSRT {elt_map = m}) cafs =
           add srt (sortBy farthestFst cafs)
             where
277
               farthestFst x y = case (Map.lookup x m, Map.lookup y m) of
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
                                   (Nothing, Nothing) -> EQ
                                   (Nothing, Just _)  -> LT
                                   (Just _,  Nothing) -> GT
                                   (Just d, Just d')  -> compare d' d
               add srt [] = srt
               add srt@(TopSRT {next_elt = next}) (caf : rst) =
                 case cafOffset srt caf of
                   Just ix -> if next - ix > maxBmpSize then
                                add (addCAF caf srt) rst
                              else srt
                   Nothing -> add (addCAF caf srt) rst
     (topSRT, subSRTs) <- sub_srt topSRT cafs
     let (sub_tbls, blockSRTs) = subSRTs
     return (topSRT, sub_tbls, blockSRTs)

-- Construct an SRT bitmap.
-- Adapted from simpleStg/SRT.lhs, which expects Id's.
295
procpointSRT :: CLabel -> Map CLabel Int -> [CLabel] ->
296
                FuelUniqSM (Maybe CmmTop, C_SRT)
297
procpointSRT _ _ [] =
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
298 299 300 301 302
 return (Nothing, NoC_SRT)
procpointSRT top_srt top_table entries =
 do (top, srt) <- bitmap `seq` to_SRT top_srt offset len bitmap
    return (top, srt)
  where
303
    ints = map (expectJust "constructSRT" . flip Map.lookup top_table) entries
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
304 305 306 307 308 309 310 311 312 313
    sorted_ints = sortLe (<=) ints
    offset = head sorted_ints
    bitmap_entries = map (subtract offset) sorted_ints
    len = P.last bitmap_entries + 1
    bitmap = intsToBitmap len bitmap_entries

maxBmpSize :: Int
maxBmpSize = widthInBits wordWidth `div` 2

-- Adapted from codeGen/StgCmmUtils, which converts from SRT to C_SRT.
314
to_SRT :: CLabel -> Int -> Int -> Bitmap -> FuelUniqSM (Maybe CmmTop, C_SRT)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
315 316 317 318 319
to_SRT top_srt off len bmp
  | len > maxBmpSize || bmp == [fromIntegral srt_escape]
  = do id <- getUniqueM
       let srt_desc_lbl = mkLargeSRTLabel id
           tbl = CmmData RelocatableReadOnlyData $
320
                   Statics srt_desc_lbl $ map CmmStaticLit
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
321 322 323 324 325 326 327 328 329 330 331 332
                     ( cmmLabelOffW top_srt off
                     : mkWordCLit (fromIntegral len)
                     : map mkWordCLit bmp)
       return (Just tbl, C_SRT srt_desc_lbl 0 srt_escape)
  | otherwise
  = return (Nothing, C_SRT top_srt off (fromIntegral (head bmp)))
	-- The fromIntegral converts to StgHalfWord

-- Gather CAF info for a procedure, but only if the procedure
-- doesn't have a static closure.
-- (If it has a static closure, it will already have an SRT to
--  keep its CAFs live.)
333
-- Any procedure referring to a non-static CAF c must keep live
334
-- any CAF that is reachable from c.
335
localCAFInfo :: CAFEnv -> CmmTop -> Maybe (CLabel, CAFSet)
336
localCAFInfo _      (CmmData _ _) = Nothing
337 338
localCAFInfo cafEnv (CmmProc top_info top_l (CmmGraph {g_entry=entry})) =
  case info_tbl top_info of
339
    CmmInfoTable _ False _ _ _ ->
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
340
      Just (cvtToClosureLbl top_l,
341
            expectJust "maybeBindCAFs" $ mapLookup entry cafEnv)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
342 343 344 345 346 347 348 349 350 351 352 353 354
    _ -> Nothing

-- Once we have the local CAF sets for some (possibly) mutually
-- recursive functions, we can create an environment mapping
-- each function to its set of CAFs. Note that a CAF may
-- be a reference to a function. If that function f does not have
-- a static closure, then we need to refer specifically
-- to the set of CAFs used by f. Of course, the set of CAFs
-- used by f must be included in the local CAF sets that are input to
-- this function. To minimize lookup time later, we return
-- the environment with every reference to f replaced by its set of CAFs.
-- To do this replacement efficiently, we gather strongly connected
-- components, then we sort the components in topological order.
355 356
mkTopCAFInfo :: [(CLabel, CAFSet)] -> Map CLabel CAFSet
mkTopCAFInfo localCAFs = foldl addToTop Map.empty g
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
357
  where addToTop env (AcyclicSCC (l, cafset)) =
358
          Map.insert l (flatten env cafset) env
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
359 360
        addToTop env (CyclicSCC nodes) =
          let (lbls, cafsets) = unzip nodes
361 362 363
              cafset  = lbls `Map.deleteList` foldl Map.union Map.empty cafsets
          in foldl (\env l -> Map.insert l (flatten env cafset) env) env lbls
        flatten env cafset = Map.foldRightWithKey (lookup env) Map.empty cafset
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
364
        lookup env caf () cafset' =
365 366 367
          case Map.lookup caf env of Just cafs -> Map.foldRightWithKey add cafset' cafs
                                     Nothing -> add caf () cafset'
        add caf () cafset' = Map.insert caf () cafset'
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
368
        g = stronglyConnCompFromEdgedVertices
369
              (map (\n@(l, cafs) -> (n, l, Map.keys cafs)) localCAFs)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
370 371 372

type StackLayout = [Maybe LocalReg]

373
-- Bundle the CAFs used at a procpoint.
374 375 376 377
bundleCAFs :: CAFEnv -> CmmTop -> (CAFSet, CmmTop)
bundleCAFs cafEnv t@(CmmProc _ _ (CmmGraph {g_entry=entry})) =
  (expectJust "bundleCAFs" (mapLookup entry cafEnv), t)
bundleCAFs _ t = (Map.empty, t)
378

dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
379
-- Construct the SRTs for the given procedure.
380 381 382
setInfoTableSRT :: Map CLabel CAFSet -> TopSRT -> (CAFSet, CmmTop) ->
                   FuelUniqSM (TopSRT, [CmmTop])
setInfoTableSRT topCAFMap topSRT (cafs, t) =
383 384
  setSRT cafs topCAFMap topSRT t

385
setSRT :: CAFSet -> Map CLabel CAFSet -> TopSRT ->
386
          CmmTop -> FuelUniqSM (TopSRT, [CmmTop])
387 388
setSRT cafs topCAFMap topSRT t =
  do (topSRT, cafTable, srt) <- buildSRTs topSRT topCAFMap cafs
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
389 390
     let t' = updInfo id (const srt) t
     case cafTable of
391
       Just tbl -> return (topSRT, [t', tbl])
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
392 393
       Nothing  -> return (topSRT, [t'])

394 395 396 397 398 399
updInfo :: (StackLayout -> StackLayout) -> (C_SRT -> C_SRT) -> CmmTop -> CmmTop
updInfo toVars toSrt (CmmProc top_info top_l g) =
  CmmProc (top_info {info_tbl=updInfoTbl toVars toSrt (info_tbl top_info)}) top_l g
updInfo _ _ t = t

updInfoTbl :: (StackLayout -> StackLayout) -> (C_SRT -> C_SRT) -> CmmInfoTable -> CmmInfoTable
400 401
updInfoTbl toVars toSrt (CmmInfoTable l s p t typeinfo)
  = CmmInfoTable l s p t typeinfo'
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
402 403 404 405 406 407
    where typeinfo' = case typeinfo of
            t@(ConstrInfo _ _ _)    -> t
            (FunInfo    c s a d e)  -> FunInfo c (toSrt s) a d e
            (ThunkInfo  c s)        -> ThunkInfo c (toSrt s)
            (ThunkSelectorInfo x s) -> ThunkSelectorInfo x (toSrt s)
            (ContInfo v s)          -> ContInfo (toVars v) (toSrt s)
408
updInfoTbl _ _ t@CmmNonInfoTable = t
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426
  
----------------------------------------------------------------
-- Safe foreign calls: We need to insert the code that suspends and resumes
-- the thread before and after a safe foreign call.
-- Why do we do this so late in the pipeline?
-- Because we need this code to appear without interrruption: you can't rely on the
-- value of the stack pointer between the call and resetting the thread state;
-- you need to have an infotable on the young end of the stack both when
-- suspending the thread and making the foreign call.
-- All of this is much easier if we insert the suspend and resume calls here.

-- At the same time, we prepare for the stages of the compiler that
-- build the proc points. We have to do this at the same time because
-- the safe foreign calls need special treatment with respect to infotables.
-- A safe foreign call needs an infotable even though it isn't
-- a procpoint. The following datatype captures the information
-- needed to generate the infotables along with the Cmm data and procedures.

427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442
-- JD: Why not do this while splitting procedures?
lowerSafeForeignCalls :: AreaMap -> CmmTop -> FuelUniqSM CmmTop
lowerSafeForeignCalls _ t@(CmmData _ _) = return t
lowerSafeForeignCalls areaMap (CmmProc info l g@(CmmGraph {g_entry=entry})) = do
  let block b mblocks = mblocks >>= lowerSafeCallBlock entry areaMap b
  blocks <- foldGraphBlocks block (return mapEmpty) g
  return $ CmmProc info l (ofBlockMap entry blocks)

-- If the block ends with a safe call in the block, lower it to an unsafe
-- call (with appropriate saves and restores before and after).
lowerSafeCallBlock :: BlockId -> AreaMap -> CmmBlock -> BlockEnv CmmBlock
                              -> FuelUniqSM (BlockEnv CmmBlock)
lowerSafeCallBlock entry areaMap b blocks =
  case blockToNodeList b of
    (JustC (CmmEntry id), m, JustC l@(CmmForeignCall {})) -> lowerSafeForeignCall entry areaMap blocks id m l
    _                                                    -> return $ insertBlock b blocks
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
443 444 445

-- Late in the code generator, we want to insert the code necessary
-- to lower a safe foreign call to a sequence of unsafe calls.
446 447 448 449 450
lowerSafeForeignCall :: BlockId -> AreaMap -> BlockEnv CmmBlock -> BlockId -> [CmmNode O O] -> CmmNode O C
                                -> FuelUniqSM (BlockEnv CmmBlock)
lowerSafeForeignCall entry areaMap blocks bid m
    (CmmForeignCall {tgt=tgt, res=rs, args=as, succ=succ, updfr = updfr_off, intrbl = intrbl}) =
 do let newTemp rep = getUniqueM >>= \u -> return (LocalReg u rep)
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
451 452 453 454
    -- Both 'id' and 'new_base' are KindNonPtr because they're
    -- RTS-only objects and are not subject to garbage collection
    id <- newTemp bWord
    new_base <- newTemp (cmmRegType (CmmGlobal BaseReg))
455
    let (caller_save, caller_load) = callerSaveVolatileRegs
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
456
    load_tso <- newTemp gcWord -- TODO FIXME NOW
457 458 459 460 461 462 463
    load_stack <- newTemp gcWord -- TODO FIXME NOW
    let (<**>) = (M.<*>)
    let suspendThread = foreignLbl "suspendThread"
        resumeThread  = foreignLbl "resumeThread"
        foreignLbl name = CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit name)))
        suspend = saveThreadState <**>
                  caller_save <**>
dias@eecs.harvard.edu's avatar
dias@eecs.harvard.edu committed
464
                  mkUnsafeCall (ForeignTarget suspendThread
465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495
                                (ForeignConvention CCallConv [AddrHint, NoHint] [AddrHint]))
                               [id] [CmmReg (CmmGlobal BaseReg), CmmLit (CmmInt (fromIntegral (fromEnum intrbl)) wordWidth)]
        midCall = mkUnsafeCall tgt rs as
        resume  = mkUnsafeCall (ForeignTarget resumeThread
                                (ForeignConvention CCallConv [AddrHint] [AddrHint]))
                     [new_base] [CmmReg (CmmLocal id)] <**>
                  -- Assign the result to BaseReg: we
                  -- might now have a different Capability!
                  mkAssign (CmmGlobal BaseReg) (CmmReg (CmmLocal new_base)) <**>
                  caller_load <**>
                  loadThreadState load_tso load_stack
        -- We have to save the return value on the stack because its next use
        -- may appear in a different procedure due to procpoint splitting...
        saveRetVals = foldl (<**>) emptyAGraph $ map (M.mkMiddle . spill) rs
        spill r = CmmStore (regSlot r) (CmmReg $ CmmLocal r)
        regSlot r@(LocalReg _ _) = CmmRegOff (CmmGlobal Sp) (sp_off - offset)
          where offset = w + expectJust "lowerForeign" (Map.lookup (RegSlot r) areaMap)
                sp_off = wORD_SIZE + expectJust "lowerForeign" (Map.lookup (CallArea area) areaMap)
                area = if succ == entry then Old else Young succ
                w = widthInBytes $ typeWidth $ localRegType r
        -- Note: The successor must be a procpoint, and we have already split,
        --       so we use a jump, not a branch.
        succLbl = CmmLit (CmmLabel (infoTblLbl succ))
        jump = CmmCall { cml_target  = succLbl, cml_cont = Nothing
                       , cml_args    = widthInBytes wordWidth ,cml_ret_args = 0
                       , cml_ret_off = updfr_off}
    graph' <- liftUniq $ labelAGraph bid $ catAGraphs (map M.mkMiddle m) <**>
                                           suspend <**> midCall <**>
                                           resume  <**> saveRetVals <**> M.mkLast jump
    return $ blocks `mapUnion` toBlockMap graph'
lowerSafeForeignCall _ _ _ _ _ _ = panic "lowerSafeForeignCall was passed something else"