StgCmmHeap.hs 23.3 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
        allocDynClosure, allocDynClosureCmm,
19
        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
        -> 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
66
        -> FCode CmmExpr -- returns Hp+n
67

68 69 70 71 72
allocDynClosureCmm
        :: CmmInfoTable -> LambdaFormInfo -> CmmExpr -> CmmExpr
        -> [(CmmExpr, VirtualHpOffset)]
        -> FCode CmmExpr -- returns Hp+n

73
-- allocDynClosure allocates the thing in the heap,
74
-- and modifies the virtual Hp to account for this.
75 76 77
-- 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.
78

79 80 81 82 83 84 85 86 87 88
-- allocDynClosure returns an (Hp+8) CmmExpr, and hence the result is
-- only valid until Hp is changed.  The caller should assign the
-- result to a LocalReg if it is required to remain live.
--
-- The reason we don't assign it to a LocalReg here is that the caller
-- is often about to call regIdInfo, which immediately assigns the
-- result of allocDynClosure to a new temp in order to add the tag.
-- So by not generating a LocalReg here we avoid a common source of
-- new temporaries and save some compile time.  This can be quite
-- significant - see test T4801.
89 90


Simon Marlow's avatar
Simon Marlow committed
91
allocDynClosure info_tbl lf_info use_cc _blame_cc args_w_offsets
92 93
  = do  { let (args, offsets) = unzip args_w_offsets
        ; cmm_args <- mapM getArgAmode args     -- No void args
94
        ; allocDynClosureCmm info_tbl lf_info
Simon Marlow's avatar
Simon Marlow committed
95
                             use_cc _blame_cc (zip cmm_args offsets)
96 97
        }

Simon Marlow's avatar
Simon Marlow committed
98
allocDynClosureCmm info_tbl lf_info use_cc _blame_cc amodes_w_offsets
99 100 101
  = do  { virt_hp <- getVirtHp

        -- SAY WHAT WE ARE ABOUT TO DO
Simon Marlow's avatar
Simon Marlow committed
102 103 104
        ; let rep = cit_rep info_tbl
        ; tickyDynAlloc rep lf_info
        ; profDynAlloc rep use_cc
105 106 107 108 109 110 111 112

        -- 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
113
                info_ptr = CmmLit (CmmLabel (cit_lbl info_tbl))
114 115 116

        -- ALLOCATE THE OBJECT
        ; base <- getHpRelOffset info_offset
117
        ; emitComment $ mkFastString "allocDynClosure"
118 119 120 121 122
        ; emitSetDynHdr base info_ptr  use_cc
        ; let (cmm_args, offsets) = unzip amodes_w_offsets
        ; hpStore base cmm_args offsets

        -- BUMP THE VIRTUAL HEAP POINTER
123 124
        ; dflags <- getDynFlags
        ; setVirtHp (virt_hp + heapClosureSize dflags rep)
125

126 127
        ; getHpRelOffset info_offset
        }
128 129

emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
130
emitSetDynHdr base info_ptr ccs
131 132
  = do dflags <- getDynFlags
       hpStore base (header dflags) [0..]
133
  where
134 135
    header :: DynFlags -> [CmmExpr]
    header dflags = [info_ptr] ++ dynProfHdr dflags ccs
136 137 138
        -- ToDo: Gransim stuff
        -- ToDo: Parallel stuff
        -- No ticky header
139 140 141 142

hpStore :: CmmExpr -> [CmmExpr] -> [VirtualHpOffset] -> FCode ()
-- Store the item (expr,off) in base[off]
hpStore base vals offs
143 144 145
  = do dflags <- getDynFlags
       let mk_store val off = mkStore (cmmOffsetW dflags base off) val
       emit (catAGraphs (zipWith mk_store vals offs))
146 147 148


-----------------------------------------------------------
149
--              Layout of static closures
150 151 152 153 154
-----------------------------------------------------------

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

155
mkStaticClosureFields
156 157
        :: DynFlags
        -> CmmInfoTable
158
        -> CostCentreStack
159
        -> CafInfo
160 161
        -> [CmmLit]             -- Payload
        -> [CmmLit]             -- The full closure
162 163
mkStaticClosureFields dflags info_tbl ccs caf_refs payload
  = mkStaticClosure dflags info_lbl ccs payload padding
164
        static_link_field saved_info_field
165
  where
Simon Marlow's avatar
Simon Marlow committed
166
    info_lbl = cit_lbl info_tbl
167 168 169 170 171 172 173 174 175

    -- 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
176 177 178
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:
179

Simon Marlow's avatar
Simon Marlow committed
180
    is_caf = isThunkRep (cit_rep info_tbl)
181

182 183 184
    padding
        | not is_caf = []
        | otherwise  = ASSERT(null payload) [mkIntCLit 0]
185 186

    static_link_field
187
        | is_caf || staticClosureNeedsLink (mayHaveCafRefs caf_refs) info_tbl
Simon Marlow's avatar
Simon Marlow committed
188 189 190
        = [static_link_value]
        | otherwise
        = []
191 192

    saved_info_field
193 194
        | is_caf     = [mkIntCLit 0]
        | otherwise  = []
195

196
        -- For a static constructor which has NoCafRefs, we set the
197 198
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
199
    static_link_value
200 201
        | mayHaveCafRefs caf_refs  = mkIntCLit 0
        | otherwise                = mkIntCLit 1  -- No CAF refs
202 203


204
mkStaticClosure :: DynFlags -> CLabel -> CostCentreStack -> [CmmLit]
205
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
206
mkStaticClosure dflags info_lbl ccs payload padding static_link_field saved_info_field
207 208
  =  [CmmLabel info_lbl]
  ++ variable_header_words
209
  ++ concatMap (padLitToWord dflags) payload
210
  ++ padding
211 212 213 214
  ++ static_link_field
  ++ saved_info_field
  where
    variable_header_words
215 216
        =  staticGranHdr
        ++ staticParHdr
217
        ++ staticProfHdr dflags ccs
218
        ++ staticTickyHdr
219

220 221
-- 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?
222 223 224
padLitToWord :: DynFlags -> CmmLit -> [CmmLit]
padLitToWord dflags lit = lit : padding pad_length
  where width = typeWidth (cmmLitType dflags lit)
225 226 227 228 229 230 231 232
        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)

233
-----------------------------------------------------------
234
--              Heap overflow checking
235 236 237 238 239 240 241 242 243 244 245 246
-----------------------------------------------------------

{- 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
247
        r = gc_1p( r )
248 249
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.
250

251
  * A series of canned entry points like
252
        gcfun_2p( f, x, y )
253 254 255 256 257 258 259 260 261
    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 ) {
262 263 264
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
265 266 267
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
268
     it should do no register shuffling
269 270 271

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
272 273 274
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
275
          L: HpAlloc = 8
276 277
             call gc()  -- Needs an info table
             goto T }
278 279

* altHeapCheck: Immediately following an eval
280 281
  Started as
        case f x y of r { (p,q) -> rhs }
282 283 284
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
285 286 287 288 289 290
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
291

292 293
        L: r = gc_1p( r )
           goto K }
294

295 296 297 298
        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)
299 300 301

   (b) No canned sequence for results of f
       Note second info table
302 303 304 305 306 307
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
308

309 310
        L: call gc()    -- Extra info table here
           goto K
311 312 313

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
314
       case branch *not* following eval,
315 316 317
       or let-no-escape
  Exactly the same as the previous case:

318 319 320 321
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...
322

323 324
        L: call gc()
           goto K
325 326 327 328 329
-}

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

330 331 332 333 334 335
entryHeapCheck :: ClosureInfo
               -> 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 ()
336

337
entryHeapCheck cl_info nodeSet arity args code
338 339
  = do dflags <- getDynFlags
       let is_thunk = arity == 0
340 341 342 343 344
           is_fastf = case closureFunInfo cl_info of
                           Just (_, ArgGen _) -> False
                           _otherwise         -> True

           args' = map (CmmReg . CmmLocal) args
345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369
           node = case nodeSet of
                      Just r  -> CmmReg (CmmLocal r)
                      Nothing -> CmmLit (CmmLabel $ staticClosureLabel cl_info)
           stg_gc_fun    = CmmReg (CmmGlobal GCFun)
           stg_gc_enter1 = CmmReg (CmmGlobal GCEnter1)

           {- Thunks:          jump stg_gc_enter_1

              Function (fast): call (NativeNode) stg_gc_fun(fun, args)

              Function (slow): R1 = fun
                               call (slow) stg_gc_fun(args)
               XXX: this is a bit naughty, we should really pass R1 as an
               argument and use a special calling convention.
           -}
           gc_call upd
               | is_thunk
                 = mkJump dflags stg_gc_enter1 [node] upd

               | is_fastf
                 = mkJump dflags stg_gc_fun (node : args') upd

               | otherwise
                 = mkAssign nodeReg node <*>
                   mkForeignJump dflags Slow stg_gc_fun args' upd
370 371

       updfr_sz <- getUpdFrameOff
372 373 374 375

       loop_id <- newLabelC
       emitLabel loop_id
       heapCheck True (gc_call updfr_sz <*> mkBranch loop_id) code
376 377 378 379 380 381 382 383 384

{-
    -- 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.

385 386
    args'     = case fun of Just f  -> f : args
                            Nothing -> args
387
    arg_exprs = map (CmmReg . CmmLocal) args'
388
    gc_call updfr_sz
389
        | arity == 0 = mkJumpGC (CmmReg (CmmGlobal GCEnter1)) arg_exprs updfr_sz
390 391 392 393 394 395
        | 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
396

397
    gc_lbl :: [LocalReg] -> Maybe FastString
398
    gc_lbl [reg]
399 400 401 402 403 404 405 406 407 408 409
        | 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
410 411 412

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

413
    gc_lbl_ptrs :: [Bool] -> Maybe FastString
414
    -- JD: TEMPORARY -- UNTIL THESE FUNCTIONS EXIST...
415 416 417
    --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
418 419 420
-}


421 422
-- ------------------------------------------------------------
-- A heap/stack check in a case alternative
423

424 425 426 427 428 429 430 431 432 433 434 435 436 437 438

-- 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

439 440
altHeapCheck :: [LocalReg] -> FCode a -> FCode a
altHeapCheck regs code
441 442 443
  = case cannedGCEntryPoint regs of
      Nothing -> genericGC code
      Just gc -> do
444
        dflags <- getDynFlags
445
        lret <- newLabelC
446
        let (off, copyin) = copyInOflow dflags NativeReturn (Young lret) regs
447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
        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
462 463 464
  = do dflags <- getDynFlags
       updfr_sz <- getUpdFrameOff
       heapCheck False (gc_call dflags gc updfr_sz) code
465 466
  where
    reg_exprs = map (CmmReg . CmmLocal) regs
467
      -- Note [stg_gc arguments]
468

469 470 471
    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,[])
472

473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498
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
499

500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516
-- 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.

517

518 519
-- | The generic GC procedure; no params, no results
generic_gc :: CmmExpr
520
generic_gc = mkGcLabel "stg_gc_noregs"
521 522

-- | Create a CLabel for calling a garbage collector entry point
523 524
mkGcLabel :: String -> CmmExpr
mkGcLabel s = CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit s)))
525 526

-------------------------------
527 528
heapCheck :: Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck checkStack do_gc code
529
  = getHeapUsage $ \ hpHw ->
530 531
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
532
    do  { codeOnly $ do_checks checkStack hpHw do_gc
533 534 535 536
        ; tickyAllocHeap hpHw
        ; doGranAllocate hpHw
        ; setRealHp hpHw
        ; code }
537

538
do_checks :: Bool       -- Should we check the stack?
539 540
          -> WordOff    -- Heap headroom
          -> CmmAGraph  -- What to do on failure
541 542 543 544
          -> FCode ()
do_checks checkStack alloc do_gc = do
  gc_id <- newLabelC

545 546 547
  when checkStack $ do
     dflags <- getDynFlags
     emit =<< mkCmmIfGoto (sp_oflo dflags) gc_id
548

549
  when (alloc /= 0) $ do
550 551
     dflags <- getDynFlags
     emitAssign hpReg (bump_hp dflags)
552
     emit =<< mkCmmIfThen hp_oflo (alloc_n <*> mkBranch gc_id)
553 554

  emitOutOfLine gc_id $
555 556
     do_gc -- this is expected to jump back somewhere

557 558 559 560 561 562
                -- 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.
563
  where
564
    alloc_lit = mkIntExpr (alloc*wORD_SIZE) -- Bytes
565
    bump_hp dflags = cmmOffsetExprB dflags (CmmReg hpReg) alloc_lit
566

567
    -- Sp overflow if (Sp - CmmHighStack < SpLim)
568 569 570
    sp_oflo dflags
            = CmmMachOp mo_wordULt
                  [CmmMachOp (MO_Sub (typeWidth (cmmRegType dflags spReg)))
571 572
                             [CmmReg spReg, CmmLit CmmHighStackMark],
                   CmmReg spLimReg]
573

574 575 576 577 578 579 580
    -- 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
581 582 583 584 585 586 587 588 589 590

{-

{- 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. -}

591 592 593 594 595 596 597
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 ()
598 599

unbxTupleHeapCheck regs ptrs nptrs fail_code code
600
  -- We can't manage more than 255 pointers/non-pointers
601 602
  -- in a generic heap check.
  | ptrs > 255 || nptrs > 255 = panic "altHeapCheck"
603
  | otherwise
604
  = initHeapUsage $ \ hpHw -> do
605 606 607 608 609
        { codeOnly $ do { do_checks 0 {- no stack check -} hpHw
                                    full_fail_code rts_label
                        ; tickyAllocHeap hpHw }
        ; setRealHp hpHw
        ; code }
610 611
  where
    full_fail_code  = fail_code `plusStmts` oneStmt assign_liveness
612 613 614 615
    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")))
616 617


618
{- Old Gransim com -- I have no idea whether it still makes sense (SLPJ Sep07)
619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637
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.

%************************************************************************
638
%*                                                                      *
639
     Generic Heap/Stack Checks - used in the RTS
640
%*                                                                      *
641 642 643 644 645 646 647 648
%************************************************************************

\begin{code}
hpChkGen :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
hpChkGen bytes liveness reentry
  = do_checks' bytes True assigns stg_gc_gen
  where
    assigns = mkStmts [
649 650 651
                CmmAssign (CmmGlobal (VanillaReg 9))  liveness,
                CmmAssign (CmmGlobal (VanillaReg 10)) reentry
                ]
652 653 654 655 656 657 658 659 660 661 662 663

-- 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}

-}