StgCmmHeap.hs 23.5 KB
Newer Older
1 2 3 4 5 6 7 8 9
-----------------------------------------------------------------------------
--
-- Stg to C--: heap management functions
--
-- (c) The University of Glasgow 2004-2006
--
-----------------------------------------------------------------------------

module StgCmmHeap (
10 11
        getVirtHp, setVirtHp, setRealHp,
        getHpRelOffset, hpRel,
12

13
        entryHeapCheck, altHeapCheck, altHeapCheckReturnsTo,
14

15 16
        mkVirtHeapOffsets, mkVirtConstrOffsets,
        mkStaticClosureFields, mkStaticClosure,
17

18 19
        allocDynClosure, allocDynClosureReg, allocDynClosureCmm,
        emitSetDynHdr
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
    ) where

#include "HsVersions.h"

import StgSyn
import CLabel
import StgCmmLayout
import StgCmmUtils
import StgCmmMonad
import StgCmmProf
import StgCmmTicky
import StgCmmGran
import StgCmmClosure
import StgCmmEnv

35
import MkGraph
36

37
import Hoopl
38
import SMRep
39
import Cmm
40 41 42
import CmmUtils
import CostCentre
import Outputable
43
import IdInfo( CafInfo(..), mayHaveCafRefs )
44
import Module
45
import DynFlags
46
import FastString( mkFastString, fsLit )
47
import Constants
48
import Util
49

50 51
import Control.Monad (when)

52
-----------------------------------------------------------
53
--              Initialise dynamic heap objects
54 55 56
-----------------------------------------------------------

allocDynClosure
Simon Marlow's avatar
Simon Marlow committed
57 58
        :: CmmInfoTable
        -> LambdaFormInfo
59 60 61 62 63 64 65 66 67
        -> CmmExpr              -- Cost Centre to stick in the object
        -> CmmExpr              -- Cost Centre to blame for this alloc
                                -- (usually the same; sometimes "OVERHEAD")

        -> [(NonVoid StgArg, VirtualHpOffset)]  -- Offsets from start of object
                                                -- ie Info ptr has offset zero.
                                                -- No void args in here
        -> FCode (LocalReg, CmmAGraph)

68
allocDynClosureReg
Simon Marlow's avatar
Simon Marlow committed
69
        :: CmmInfoTable -> LambdaFormInfo -> CmmExpr -> CmmExpr
70 71 72
        -> [(CmmExpr, VirtualHpOffset)]
        -> FCode (LocalReg, CmmAGraph)

73 74 75 76 77
allocDynClosureCmm
        :: CmmInfoTable -> LambdaFormInfo -> CmmExpr -> CmmExpr
        -> [(CmmExpr, VirtualHpOffset)]
        -> FCode CmmExpr -- returns Hp+n

78
-- allocDynClosure allocates the thing in the heap,
79
-- and modifies the virtual Hp to account for this.
80 81 82
-- The second return value is the graph that sets the value of the
-- returned LocalReg, which should point to the closure after executing
-- the graph.
83 84 85 86 87

-- Note [Return a LocalReg]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-- allocDynClosure returns a LocalReg, not a (Hp+8) CmmExpr.
-- Reason:
88 89 90 91 92
--      ...allocate object...
--      obj = Hp + 8
--      y = f(z)
--      ...here obj is still valid,
--         but Hp+8 means something quite different...
93 94


Simon Marlow's avatar
Simon Marlow committed
95
allocDynClosure info_tbl lf_info use_cc _blame_cc args_w_offsets
96 97
  = do  { let (args, offsets) = unzip args_w_offsets
        ; cmm_args <- mapM getArgAmode args     -- No void args
98
        ; allocDynClosureReg info_tbl lf_info
Simon Marlow's avatar
Simon Marlow committed
99
                             use_cc _blame_cc (zip cmm_args offsets)
100 101
        }

102 103 104 105 106 107 108 109
allocDynClosureReg  info_tbl lf_info use_cc _blame_cc amodes_w_offsets
  = do  { hp_rel <- allocDynClosureCmm info_tbl lf_info
                                       use_cc _blame_cc amodes_w_offsets

        -- Note [Return a LocalReg]
        ; getCodeR $ assignTemp hp_rel
        }

Simon Marlow's avatar
Simon Marlow committed
110
allocDynClosureCmm info_tbl lf_info use_cc _blame_cc amodes_w_offsets
111 112 113
  = do  { virt_hp <- getVirtHp

        -- SAY WHAT WE ARE ABOUT TO DO
Simon Marlow's avatar
Simon Marlow committed
114 115 116
        ; let rep = cit_rep info_tbl
        ; tickyDynAlloc rep lf_info
        ; profDynAlloc rep use_cc
117 118 119 120 121 122 123 124

        -- FIND THE OFFSET OF THE INFO-PTR WORD
        ; let   info_offset = virt_hp + 1
                -- info_offset is the VirtualHpOffset of the first
                -- word of the new object
                -- Remember, virtHp points to last allocated word,
                -- ie 1 *before* the info-ptr word of new object.

Simon Marlow's avatar
Simon Marlow committed
125
                info_ptr = CmmLit (CmmLabel (cit_lbl info_tbl))
126 127 128

        -- ALLOCATE THE OBJECT
        ; base <- getHpRelOffset info_offset
129
        ; emitComment $ mkFastString "allocDynClosure"
130 131 132 133 134
        ; emitSetDynHdr base info_ptr  use_cc
        ; let (cmm_args, offsets) = unzip amodes_w_offsets
        ; hpStore base cmm_args offsets

        -- BUMP THE VIRTUAL HEAP POINTER
135 136
        ; dflags <- getDynFlags
        ; setVirtHp (virt_hp + heapClosureSize dflags rep)
137

138 139
        ; getHpRelOffset info_offset
        }
140 141

emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
142
emitSetDynHdr base info_ptr ccs
143 144
  = do dflags <- getDynFlags
       hpStore base (header dflags) [0..]
145
  where
146 147
    header :: DynFlags -> [CmmExpr]
    header dflags = [info_ptr] ++ dynProfHdr dflags ccs
148 149 150
        -- ToDo: Gransim stuff
        -- ToDo: Parallel stuff
        -- No ticky header
151 152 153 154 155 156

hpStore :: CmmExpr -> [CmmExpr] -> [VirtualHpOffset] -> FCode ()
-- Store the item (expr,off) in base[off]
hpStore base vals offs
  = emit (catAGraphs (zipWith mk_store vals offs))
  where
157
    mk_store val off = mkStore (cmmOffsetW base off) val
158 159 160


-----------------------------------------------------------
161
--              Layout of static closures
162 163 164 165 166
-----------------------------------------------------------

-- Make a static closure, adding on any extra padding needed for CAFs,
-- and adding a static link field if necessary.

167
mkStaticClosureFields
168 169
        :: DynFlags
        -> CmmInfoTable
170
        -> CostCentreStack
171
        -> CafInfo
172 173
        -> [CmmLit]             -- Payload
        -> [CmmLit]             -- The full closure
174 175
mkStaticClosureFields dflags info_tbl ccs caf_refs payload
  = mkStaticClosure dflags info_lbl ccs payload padding
176
        static_link_field saved_info_field
177
  where
Simon Marlow's avatar
Simon Marlow committed
178
    info_lbl = cit_lbl info_tbl
179 180 181 182 183 184 185 186 187

    -- CAFs must have consistent layout, regardless of whether they
    -- are actually updatable or not.  The layout of a CAF is:
    --
    --        3 saved_info
    --        2 static_link
    --        1 indirectee
    --        0 info ptr
    --
Simon Marlow's avatar
Simon Marlow committed
188 189 190
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:
191

Simon Marlow's avatar
Simon Marlow committed
192
    is_caf = isThunkRep (cit_rep info_tbl)
193

194 195 196
    padding
        | not is_caf = []
        | otherwise  = ASSERT(null payload) [mkIntCLit 0]
197 198

    static_link_field
199
        | is_caf || staticClosureNeedsLink (mayHaveCafRefs caf_refs) info_tbl
Simon Marlow's avatar
Simon Marlow committed
200 201 202
        = [static_link_value]
        | otherwise
        = []
203 204

    saved_info_field
205 206
        | is_caf     = [mkIntCLit 0]
        | otherwise  = []
207

208
        -- For a static constructor which has NoCafRefs, we set the
209 210
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
211
    static_link_value
212 213
        | mayHaveCafRefs caf_refs  = mkIntCLit 0
        | otherwise                = mkIntCLit 1  -- No CAF refs
214 215


216
mkStaticClosure :: DynFlags -> CLabel -> CostCentreStack -> [CmmLit]
217
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
218
mkStaticClosure dflags info_lbl ccs payload padding static_link_field saved_info_field
219 220
  =  [CmmLabel info_lbl]
  ++ variable_header_words
221
  ++ concatMap padLitToWord payload
222
  ++ padding
223 224 225 226
  ++ static_link_field
  ++ saved_info_field
  where
    variable_header_words
227 228
        =  staticGranHdr
        ++ staticParHdr
229
        ++ staticProfHdr dflags ccs
230
        ++ staticTickyHdr
231

232 233
-- JD: Simon had ellided this padding, but without it the C back end asserts
-- failure. Maybe it's a bad assertion, and this padding is indeed unnecessary?
234 235 236 237 238 239 240 241 242 243 244
padLitToWord :: CmmLit -> [CmmLit]
padLitToWord lit = lit : padding pad_length
  where width = typeWidth (cmmLitType lit)
        pad_length = wORD_SIZE - widthInBytes width :: Int

        padding n | n <= 0 = []
                  | n `rem` 2 /= 0 = CmmInt 0 W8  : padding (n-1)
                  | n `rem` 4 /= 0 = CmmInt 0 W16 : padding (n-2)
                  | n `rem` 8 /= 0 = CmmInt 0 W32 : padding (n-4)
                  | otherwise      = CmmInt 0 W64 : padding (n-8)

245
-----------------------------------------------------------
246
--              Heap overflow checking
247 248 249 250 251 252 253 254 255 256 257 258
-----------------------------------------------------------

{- Note [Heap checks]
   ~~~~~~~~~~~~~~~~~~
Heap checks come in various forms.  We provide the following entry
points to the runtime system, all of which use the native C-- entry
convention.

  * gc() performs garbage collection and returns
    nothing to its caller

  * A series of canned entry points like
259
        r = gc_1p( r )
260 261
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.
262

263
  * A series of canned entry points like
264
        gcfun_2p( f, x, y )
265 266 267 268 269 270 271 272 273
    where f is a function closure of arity 2
    This performs garbage collection, keeping alive the
    three argument ptrs, and then tail-calls f(x,y)

These are used in the following circumstances

* entryHeapCheck: Function entry
    (a) With a canned GC entry sequence
        f( f_clo, x:ptr, y:ptr ) {
274 275 276
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
277 278 279
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
280
     it should do no register shuffling
281 282 283

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
284 285 286
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
287
          L: HpAlloc = 8
288 289
             call gc()  -- Needs an info table
             goto T }
290 291

* altHeapCheck: Immediately following an eval
292 293
  Started as
        case f x y of r { (p,q) -> rhs }
294 295 296
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
297 298 299 300 301 302
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
303

304 305
        L: r = gc_1p( r )
           goto K }
306

307 308 309 310
        Here, the info table needed by the call
        to gc_1p should be the *same* as the
        one for the call to f; the C-- optimiser
        spots this sharing opportunity)
311 312 313

   (b) No canned sequence for results of f
       Note second info table
314 315 316 317 318 319
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
320

321 322
        L: call gc()    -- Extra info table here
           goto K
323 324 325

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
326
       case branch *not* following eval,
327 328 329
       or let-no-escape
  Exactly the same as the previous case:

330 331 332 333
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...
334

335 336
        L: call gc()
           goto K
337 338 339 340 341
-}

--------------------------------------------------------------
-- A heap/stack check at a function or thunk entry point.

342 343 344 345 346 347 348
entryHeapCheck :: ClosureInfo
               -> Int            -- Arg Offset
               -> Maybe LocalReg -- Function (closure environment)
               -> Int            -- Arity -- not same as len args b/c of voids
               -> [LocalReg]     -- Non-void args (empty for thunk)
               -> FCode ()
               -> FCode ()
349

350
entryHeapCheck cl_info offset nodeSet arity args code
351 352
  = do dflags <- getDynFlags
       let is_thunk = arity == 0
353 354 355 356 357 358
           is_fastf = case closureFunInfo cl_info of
                           Just (_, ArgGen _) -> False
                           _otherwise         -> True

           args' = map (CmmReg . CmmLocal) args
           setN = case nodeSet of
Simon Marlow's avatar
Simon Marlow committed
359
                          Just _  -> mkNop -- No need to assign R1, it already
360
                                           -- points to the closure
361
                          Nothing -> mkAssign nodeReg $
362
                              CmmLit (CmmLabel $ staticClosureLabel cl_info)
363

364
           {- Thunks:          jump GCEnter1
365 366 367 368
              Function (fast): Set R1 = node, jump GCFun
              Function (slow): Set R1 = node, call generic_gc -}
           gc_call upd = setN <*> gc_lbl upd
           gc_lbl upd
369 370 371
               | is_thunk  = mkDirectJump dflags (CmmReg $ CmmGlobal GCEnter1) [] sp
               | is_fastf  = mkDirectJump dflags (CmmReg $ CmmGlobal GCFun) [] sp
               | otherwise = mkForeignJump dflags Slow (CmmReg $ CmmGlobal GCFun) args' upd
372 373 374 375 376 377 378
               where sp = max offset upd
           {- DT (12/08/10) This is a little fishy, mainly the sp fix up amount.
            - This is since the ncg inserts spills before the stack/heap check.
            - This should be fixed up and then we won't need to fix up the Sp on
            - GC calls, but until then this fishy code works -}

       updfr_sz <- getUpdFrameOff
379 380 381 382

       loop_id <- newLabelC
       emitLabel loop_id
       heapCheck True (gc_call updfr_sz <*> mkBranch loop_id) code
383 384 385 386 387 388 389 390 391

{-
    -- This code is slightly outdated now and we could easily keep the above
    -- GC methods. However, there may be some performance gains to be made by
    -- using more specialised GC entry points. Since the semi generic GCFun
    -- entry needs to check the node and figure out what registers to save...
    -- if we provided and used more specialised GC entry points then these
    -- runtime decisions could be turned into compile time decisions.

392 393
    args'     = case fun of Just f  -> f : args
                            Nothing -> args
394
    arg_exprs = map (CmmReg . CmmLocal) args'
395
    gc_call updfr_sz
396
        | arity == 0 = mkJumpGC (CmmReg (CmmGlobal GCEnter1)) arg_exprs updfr_sz
397 398 399 400 401 402
        | otherwise =
            case gc_lbl args' of
                Just _lbl -> panic "StgCmmHeap.entryHeapCheck: not finished"
                            -- mkJumpGC (CmmLit (CmmLabel (mkRtsCodeLabel lbl)))
                            --         arg_exprs updfr_sz
                Nothing  -> mkCall generic_gc (GC, GC) [] [] updfr_sz
403

404
    gc_lbl :: [LocalReg] -> Maybe FastString
405
    gc_lbl [reg]
406 407 408 409 410 411 412 413 414 415 416
        | isGcPtrType ty  = Just (sLit "stg_gc_unpt_r1") -- "stg_gc_fun_1p"
        | isFloatType ty  = case width of
                              W32 -> Just (sLit "stg_gc_f1")
                              W64 -> Just (sLit "stg_gc_d1")
                              _other -> Nothing
        | width == wordWidth = Just (mkGcLabel "stg_gc_unbx_r1")
        | width == W64       = Just (mkGcLabel "stg_gc_l1")
        | otherwise          = Nothing
        where
          ty = localRegType reg
          width = typeWidth ty
417 418 419

    gc_lbl regs = gc_lbl_ptrs (map (isGcPtrType . localRegType) regs)

420
    gc_lbl_ptrs :: [Bool] -> Maybe FastString
421
    -- JD: TEMPORARY -- UNTIL THESE FUNCTIONS EXIST...
422 423 424
    --gc_lbl_ptrs [True,True]      = Just (sLit "stg_gc_fun_2p")
    --gc_lbl_ptrs [True,True,True] = Just (sLit "stg_gc_fun_3p")
    gc_lbl_ptrs _ = Nothing
425 426 427
-}


428 429
-- ------------------------------------------------------------
-- A heap/stack check in a case alternative
430

431 432 433 434 435 436 437 438 439 440 441 442 443 444 445

-- If there are multiple alts and we need to GC, but don't have a
-- continuation already (the scrut was simple), then we should
-- pre-generate the continuation.  (if there are multiple alts it is
-- always a canned GC point).

-- altHeapCheck:
-- If we have a return continuation,
--   then if it is a canned GC pattern,
--           then we do mkJumpReturnsTo
--           else we do a normal call to stg_gc_noregs
--   else if it is a canned GC pattern,
--           then generate the continuation and do mkCallReturnsTo
--           else we do a normal call to stg_gc_noregs

446 447
altHeapCheck :: [LocalReg] -> FCode a -> FCode a
altHeapCheck regs code
448 449 450
  = case cannedGCEntryPoint regs of
      Nothing -> genericGC code
      Just gc -> do
451
        dflags <- getDynFlags
452
        lret <- newLabelC
453
        let (off, copyin) = copyInOflow dflags NativeReturn (Young lret) regs
454 455 456 457 458 459 460 461 462 463 464 465 466 467 468
        lcont <- newLabelC
        emitOutOfLine lret (copyin <*> mkBranch lcont)
        emitLabel lcont
        cannedGCReturnsTo False gc regs lret off code

altHeapCheckReturnsTo :: [LocalReg] -> Label -> ByteOff -> FCode a -> FCode a
altHeapCheckReturnsTo regs lret off code
  = case cannedGCEntryPoint regs of
      Nothing -> genericGC code
      Just gc -> cannedGCReturnsTo True gc regs lret off code

cannedGCReturnsTo :: Bool -> CmmExpr -> [LocalReg] -> Label -> ByteOff
                  -> FCode a
                  -> FCode a
cannedGCReturnsTo cont_on_stack gc regs lret off code
469 470 471
  = do dflags <- getDynFlags
       updfr_sz <- getUpdFrameOff
       heapCheck False (gc_call dflags gc updfr_sz) code
472 473
  where
    reg_exprs = map (CmmReg . CmmLocal) regs
474
      -- Note [stg_gc arguments]
475

476 477 478
    gc_call dflags label sp
      | cont_on_stack = mkJumpReturnsTo dflags label GC reg_exprs lret off sp
      | otherwise     = mkCallReturnsTo dflags label GC reg_exprs lret off sp (0,[])
479

480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505
genericGC :: FCode a -> FCode a
genericGC code
  = do updfr_sz <- getUpdFrameOff
       lretry <- newLabelC
       emitLabel lretry
       call <- mkCall generic_gc (GC, GC) [] [] updfr_sz (0,[])
       heapCheck False (call <*> mkBranch lretry) code

cannedGCEntryPoint :: [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint regs
  = case regs of
      []  -> Just (mkGcLabel "stg_gc_noregs")
      [reg]
          | isGcPtrType ty -> Just (mkGcLabel "stg_gc_unpt_r1")
          | isFloatType ty -> case width of
                                  W32       -> Just (mkGcLabel "stg_gc_f1")
                                  W64       -> Just (mkGcLabel "stg_gc_d1")
                                  _         -> Nothing
        
          | width == wordWidth -> Just (mkGcLabel "stg_gc_unbx_r1")
          | width == W64       -> Just (mkGcLabel "stg_gc_l1")
          | otherwise          -> Nothing
          where
              ty = localRegType reg
              width = typeWidth ty
      _otherwise -> Nothing
506

507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
-- Note [stg_gc arguments]
-- It might seem that we could avoid passing the arguments to the
-- stg_gc function, because they are already in the right registers.
-- While this is usually the case, it isn't always.  Sometimes the
-- code generator has cleverly avoided the eval in a case, e.g. in
-- ffi/should_run/4221.hs we found
--
--   case a_r1mb of z
--     FunPtr x y -> ...
--
-- where a_r1mb is bound a top-level constructor, and is known to be
-- evaluated.  The codegen just assigns x, y and z, and continues;
-- R1 is never assigned.
--
-- So we'll have to rely on optimisations to eliminatethese
-- assignments where possible.

524

525 526
-- | The generic GC procedure; no params, no results
generic_gc :: CmmExpr
527
generic_gc = mkGcLabel "stg_gc_noregs"
528 529

-- | Create a CLabel for calling a garbage collector entry point
530 531
mkGcLabel :: String -> CmmExpr
mkGcLabel s = CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit s)))
532 533

-------------------------------
534 535
heapCheck :: Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck checkStack do_gc code
536
  = getHeapUsage $ \ hpHw ->
537 538
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
539
    do  { codeOnly $ do_checks checkStack hpHw do_gc
540 541 542 543
        ; tickyAllocHeap hpHw
        ; doGranAllocate hpHw
        ; setRealHp hpHw
        ; code }
544

545
do_checks :: Bool       -- Should we check the stack?
546 547
          -> WordOff    -- Heap headroom
          -> CmmAGraph  -- What to do on failure
548 549 550 551
          -> FCode ()
do_checks checkStack alloc do_gc = do
  gc_id <- newLabelC

552 553
  when checkStack $
     emit =<< mkCmmIfGoto sp_oflo gc_id
554

555 556 557
  when (alloc /= 0) $ do
     emitAssign hpReg bump_hp
     emit =<< mkCmmIfThen hp_oflo (alloc_n <*> mkBranch gc_id)
558 559

  emitOutOfLine gc_id $
560 561
     do_gc -- this is expected to jump back somewhere

562 563 564 565 566 567
                -- Test for stack pointer exhaustion, then
                -- bump heap pointer, and test for heap exhaustion
                -- Note that we don't move the heap pointer unless the
                -- stack check succeeds.  Otherwise we might end up
                -- with slop at the end of the current block, which can
                -- confuse the LDV profiler.
568
  where
569
    alloc_lit = CmmLit (mkIntCLit (alloc*wORD_SIZE)) -- Bytes
570 571
    bump_hp   = cmmOffsetExprB (CmmReg hpReg) alloc_lit

572 573 574
    -- Sp overflow if (Sp - CmmHighStack < SpLim)
    sp_oflo = CmmMachOp mo_wordULt
                  [CmmMachOp (MO_Sub (typeWidth (cmmRegType spReg)))
575 576
                             [CmmReg spReg, CmmLit CmmHighStackMark],
                   CmmReg spLimReg]
577

578 579 580 581 582 583 584
    -- Hp overflow if (Hp > HpLim)
    -- (Hp has been incremented by now)
    -- HpLim points to the LAST WORD of valid allocation space.
    hp_oflo = CmmMachOp mo_wordUGt
                  [CmmReg hpReg, CmmReg (CmmGlobal HpLim)]

    alloc_n = mkAssign (CmmGlobal HpAlloc) alloc_lit
585 586 587 588 589 590 591 592 593 594

{-

{- Unboxed tuple alternatives and let-no-escapes (the two most annoying
constructs to generate code for!)  For unboxed tuple returns, there
are an arbitrary number of possibly unboxed return values, some of
which will be in registers, and the others will be on the stack.  We
always organise the stack-resident fields into pointers &
non-pointers, and pass the number of each to the heap check code. -}

595 596 597 598 599 600 601
unbxTupleHeapCheck
        :: [(Id, GlobalReg)]    -- Live registers
        -> WordOff      -- no. of stack slots containing ptrs
        -> WordOff      -- no. of stack slots containing nonptrs
        -> CmmAGraph    -- code to insert in the failure path
        -> FCode ()
        -> FCode ()
602 603

unbxTupleHeapCheck regs ptrs nptrs fail_code code
604
  -- We can't manage more than 255 pointers/non-pointers
605 606
  -- in a generic heap check.
  | ptrs > 255 || nptrs > 255 = panic "altHeapCheck"
607
  | otherwise
608
  = initHeapUsage $ \ hpHw -> do
609 610 611 612 613
        { codeOnly $ do { do_checks 0 {- no stack check -} hpHw
                                    full_fail_code rts_label
                        ; tickyAllocHeap hpHw }
        ; setRealHp hpHw
        ; code }
614 615
  where
    full_fail_code  = fail_code `plusStmts` oneStmt assign_liveness
616 617 618 619
    assign_liveness = CmmAssign (CmmGlobal (VanillaReg 9))      -- Ho ho ho!
                                (CmmLit (mkWordCLit liveness))
    liveness        = mkRegLiveness regs ptrs nptrs
    rts_label       = CmmLit (CmmLabel (mkRtsCodeLabel (sLit "stg_gc_ut")))
620 621


622
{- Old Gransim com -- I have no idea whether it still makes sense (SLPJ Sep07)
623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641
For GrAnSim the code for doing a heap check and doing a context switch
has been separated. Especially, the HEAP_CHK macro only performs a
heap check. THREAD_CONTEXT_SWITCH should be used for doing a context
switch. GRAN_FETCH_AND_RESCHEDULE must be put at the beginning of
every slow entry code in order to simulate the fetching of
closures. If fetching is necessary (i.e. current closure is not local)
then an automatic context switch is done. -}


When failing a check, we save a return address on the stack and
jump to a pre-compiled code fragment that saves the live registers
and returns to the scheduler.

The return address in most cases will be the beginning of the basic
block in which the check resides, since we need to perform the check
again on re-entry because someone else might have stolen the resource
in the meantime.

%************************************************************************
642
%*                                                                      *
643
     Generic Heap/Stack Checks - used in the RTS
644
%*                                                                      *
645 646 647 648 649 650 651 652
%************************************************************************

\begin{code}
hpChkGen :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
hpChkGen bytes liveness reentry
  = do_checks' bytes True assigns stg_gc_gen
  where
    assigns = mkStmts [
653 654 655
                CmmAssign (CmmGlobal (VanillaReg 9))  liveness,
                CmmAssign (CmmGlobal (VanillaReg 10)) reentry
                ]
656 657 658 659 660 661 662 663 664 665 666 667

-- a heap check where R1 points to the closure to enter on return, and
-- we want to assign to Sp[0] on failure (used in AutoApply.cmm:BUILD_PAP).
hpChkNodePointsAssignSp0 :: CmmExpr -> CmmExpr -> FCode ()
hpChkNodePointsAssignSp0 bytes sp0
  = do_checks' bytes True assign stg_gc_enter1
  where assign = oneStmt (CmmStore (CmmReg spReg) sp0)

stg_gc_gen    = CmmLit (CmmLabel (mkRtsCodeLabel (sLit "stg_gc_gen")))
\end{code}

-}