StgCmmHeap.hs 24.8 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
        getVirtHp, setVirtHp, setRealHp,
11
        getHpRelOffset,
12

13
        entryHeapCheck, altHeapCheck, noEscapeHeapCheck, altHeapCheckReturnsTo,
14 15
        heapStackCheckGen,
        entryHeapCheck',
16

17
        mkStaticClosureFields, mkStaticClosure,
18

19
        allocDynClosure, allocDynClosureCmm, allocHeapClosure,
20
        emitSetDynHdr
21 22
    ) where

23 24
import GhcPrelude hiding ((<*>))

25 26 27 28 29
import StgSyn
import CLabel
import StgCmmLayout
import StgCmmUtils
import StgCmmMonad
30
import StgCmmProf (profDynAlloc, dynProfHdr, staticProfHdr)
31 32 33 34
import StgCmmTicky
import StgCmmClosure
import StgCmmEnv

35
import MkGraph
36

37
import Hoopl.Label
38
import SMRep
39
import BlockId
40
import Cmm
41 42
import CmmUtils
import CostCentre
43
import IdInfo( CafInfo(..), mayHaveCafRefs )
44
import Id ( Id )
45
import Module
46
import DynFlags
47
import FastString( mkFastString, fsLit )
48
import Panic( sorry )
49

50
import Control.Monad (when)
51
import Data.Maybe (isJust)
52

53
-----------------------------------------------------------
54
--              Initialise dynamic heap objects
55 56 57
-----------------------------------------------------------

allocDynClosure
58 59
        :: Maybe Id
        -> CmmInfoTable
Simon Marlow's avatar
Simon Marlow committed
60
        -> LambdaFormInfo
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
68
        -> FCode CmmExpr -- returns Hp+n
69

70
allocDynClosureCmm
71
        :: Maybe Id -> CmmInfoTable -> LambdaFormInfo -> CmmExpr -> CmmExpr
72
        -> [(CmmExpr, ByteOff)]
73 74
        -> FCode CmmExpr -- returns Hp+n

75
-- allocDynClosure allocates the thing in the heap,
76
-- and modifies the virtual Hp to account for this.
77 78 79
-- 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.
80

81 82 83 84 85 86 87 88 89 90
-- 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.
91 92


93 94 95 96 97
allocDynClosure mb_id info_tbl lf_info use_cc _blame_cc args_w_offsets = do
  let (args, offsets) = unzip args_w_offsets
  cmm_args <- mapM getArgAmode args     -- No void args
  allocDynClosureCmm mb_id info_tbl lf_info
                     use_cc _blame_cc (zip cmm_args offsets)
98 99


100 101 102 103 104 105
allocDynClosureCmm mb_id info_tbl lf_info use_cc _blame_cc amodes_w_offsets = do
  -- SAY WHAT WE ARE ABOUT TO DO
  let rep = cit_rep info_tbl
  tickyDynAlloc mb_id rep lf_info
  let info_ptr = CmmLit (CmmLabel (cit_lbl info_tbl))
  allocHeapClosure rep info_ptr use_cc amodes_w_offsets
106 107


108 109 110 111 112
-- | Low-level heap object allocation.
allocHeapClosure
  :: SMRep                            -- ^ representation of the object
  -> CmmExpr                          -- ^ info pointer
  -> CmmExpr                          -- ^ cost centre
113
  -> [(CmmExpr,ByteOff)]              -- ^ payload
114 115
  -> FCode CmmExpr                    -- ^ returns the address of the object
allocHeapClosure rep info_ptr use_cc payload = do
116 117
  profDynAlloc rep use_cc

118
  virt_hp <- getVirtHp
119

120 121 122 123 124 125
  -- 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.
126

127
  base <- getHpRelOffset info_offset
128
  emitComment $ mkFastString "allocHeapClosure"
129 130 131 132 133 134 135 136 137 138
  emitSetDynHdr base info_ptr use_cc

  -- Fill in the fields
  hpStore base payload

  -- Bump the virtual heap pointer
  dflags <- getDynFlags
  setVirtHp (virt_hp + heapClosureSizeW dflags rep)

  return base
139

140 141

emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
142
emitSetDynHdr base info_ptr ccs
143
  = do dflags <- getDynFlags
144
       hpStore base (zip (header dflags) [0, wORD_SIZE dflags ..])
145
  where
146 147
    header :: DynFlags -> [CmmExpr]
    header dflags = [info_ptr] ++ dynProfHdr dflags ccs
Gabor Greif's avatar
Gabor Greif committed
148
        -- ToDo: Parallel stuff
149
        -- No ticky header
150 151

-- Store the item (expr,off) in base[off]
152
hpStore :: CmmExpr -> [(CmmExpr, ByteOff)] -> FCode ()
153 154 155
hpStore base vals = do
  dflags <- getDynFlags
  sequence_ $
156
    [ emitStore (cmmOffsetB dflags base off) val | (val,off) <- vals ]
157 158

-----------------------------------------------------------
159
--              Layout of static closures
160 161 162 163 164
-----------------------------------------------------------

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

165
mkStaticClosureFields
166 167
        :: DynFlags
        -> CmmInfoTable
168
        -> CostCentreStack
169
        -> CafInfo
170 171
        -> [CmmLit]             -- Payload
        -> [CmmLit]             -- The full closure
172 173
mkStaticClosureFields dflags info_tbl ccs caf_refs payload
  = mkStaticClosure dflags info_lbl ccs payload padding
174
        static_link_field saved_info_field
175
  where
176
    info_lbl = cit_lbl info_tbl
177 178 179 180 181 182 183 184 185

    -- 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
    --
186 187 188
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:
189

190
    is_caf = isThunkRep (cit_rep info_tbl)
191

192
    padding
193 194
        | is_caf && null payload = [mkIntCLit dflags 0]
        | otherwise = []
195 196

    static_link_field
197
        | is_caf || staticClosureNeedsLink (mayHaveCafRefs caf_refs) info_tbl
Simon Marlow's avatar
Simon Marlow committed
198 199 200
        = [static_link_value]
        | otherwise
        = []
201 202

    saved_info_field
203
        | is_caf     = [mkIntCLit dflags 0]
204
        | otherwise  = []
205

206
        -- For a static constructor which has NoCafRefs, we set the
207 208
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
209
    static_link_value
210
        | mayHaveCafRefs caf_refs  = mkIntCLit dflags 0
211 212 213
        | otherwise                = mkIntCLit dflags 3  -- No CAF refs
                                      -- See Note [STATIC_LINK fields]
                                      -- in rts/sm/Storage.h
214

215
mkStaticClosure :: DynFlags -> CLabel -> CostCentreStack -> [CmmLit]
216
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
217
mkStaticClosure dflags info_lbl ccs payload padding static_link_field saved_info_field
218
  =  [CmmLabel info_lbl]
Jan Stolarek's avatar
Jan Stolarek committed
219
  ++ staticProfHdr dflags ccs
220
  ++ payload
221
  ++ padding
222 223 224 225
  ++ static_link_field
  ++ saved_info_field

-----------------------------------------------------------
226
--              Heap overflow checking
227 228 229 230 231 232 233 234 235 236 237 238
-----------------------------------------------------------

{- 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
239
        r = gc_1p( r )
240 241
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.
242

243
  * A series of canned entry points like
244
        gcfun_2p( f, x, y )
245 246 247 248 249 250 251 252 253
    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 ) {
254 255 256
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
257 258 259
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
260
     it should do no register shuffling
261 262 263

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
264 265 266
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
267
          L: HpAlloc = 8
268 269
             call gc()  -- Needs an info table
             goto T }
270 271

* altHeapCheck: Immediately following an eval
272 273
  Started as
        case f x y of r { (p,q) -> rhs }
274 275 276
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
277 278 279 280 281 282
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
283

284 285
        L: r = gc_1p( r )
           goto K }
286

287 288 289 290
        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)
291 292 293

   (b) No canned sequence for results of f
       Note second info table
294 295 296 297 298 299
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
300

301 302
        L: call gc()    -- Extra info table here
           goto K
303 304 305

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
306
       case branch *not* following eval,
307 308 309
       or let-no-escape
  Exactly the same as the previous case:

310 311 312 313
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...
314

315 316
        L: call gc()
           goto K
317 318 319 320 321
-}

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

322 323 324 325 326 327
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 ()
328

329
entryHeapCheck cl_info nodeSet arity args code
330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347
  = entryHeapCheck' is_fastf node arity args code
  where
    node = case nodeSet of
              Just r  -> CmmReg (CmmLocal r)
              Nothing -> CmmLit (CmmLabel $ staticClosureLabel cl_info)

    is_fastf = case closureFunInfo cl_info of
                 Just (_, ArgGen _) -> False
                 _otherwise         -> True

-- | lower-level version for CmmParse
entryHeapCheck' :: Bool           -- is a known function pattern
                -> CmmExpr        -- expression for the closure pointer
                -> Int            -- Arity -- not same as len args b/c of voids
                -> [LocalReg]     -- Non-void args (empty for thunk)
                -> FCode ()
                -> FCode ()
entryHeapCheck' is_fastf node arity args code
348 349
  = do dflags <- getDynFlags
       let is_thunk = arity == 0
350

351
           args' = map (CmmReg . CmmLocal) args
352 353 354 355 356 357 358
           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)

359
              Function (slow): call (slow) stg_gc_fun(fun, args)
360 361 362
           -}
           gc_call upd
               | is_thunk
363
                 = mkJump dflags NativeNodeCall stg_gc_enter1 [node] upd
364 365

               | is_fastf
366
                 = mkJump dflags NativeNodeCall stg_gc_fun (node : args') upd
367 368

               | otherwise
369
                 = mkJump dflags Slow stg_gc_fun (node : args') upd
370 371

       updfr_sz <- getUpdFrameOff
372

373
       loop_id <- newBlockId
374
       emitLabel loop_id
375
       heapCheck True True (gc_call updfr_sz <*> mkBranch loop_id) code
376

377 378
-- ------------------------------------------------------------
-- A heap/stack check in a case alternative
379

380 381 382 383 384 385 386 387 388 389 390 391 392 393 394

-- 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

395
altHeapCheck :: [LocalReg] -> FCode a -> FCode a
396 397 398 399
altHeapCheck regs code = altOrNoEscapeHeapCheck False regs code

altOrNoEscapeHeapCheck :: Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck checkYield regs code = do
400 401
    dflags <- getDynFlags
    case cannedGCEntryPoint dflags regs of
402
      Nothing -> genericGC checkYield code
403
      Just gc -> do
404
        lret <- newBlockId
405
        let (off, _, copyin) = copyInOflow dflags NativeReturn (Young lret) regs []
406
        lcont <- newBlockId
Peter Wortmann's avatar
Peter Wortmann committed
407 408
        tscope <- getTickScope
        emitOutOfLine lret (copyin <*> mkBranch lcont, tscope)
409
        emitLabel lcont
410
        cannedGCReturnsTo checkYield False gc regs lret off code
411 412 413

altHeapCheckReturnsTo :: [LocalReg] -> Label -> ByteOff -> FCode a -> FCode a
altHeapCheckReturnsTo regs lret off code
414 415
  = do dflags <- getDynFlags
       case cannedGCEntryPoint dflags regs of
416 417 418 419 420 421 422 423
           Nothing -> genericGC False code
           Just gc -> cannedGCReturnsTo False True gc regs lret off code

-- noEscapeHeapCheck is implemented identically to altHeapCheck (which
-- is more efficient), but cannot be optimized away in the non-allocating
-- case because it may occur in a loop
noEscapeHeapCheck :: [LocalReg] -> FCode a -> FCode a
noEscapeHeapCheck regs code = altOrNoEscapeHeapCheck True regs code
424

425
cannedGCReturnsTo :: Bool -> Bool -> CmmExpr -> [LocalReg] -> Label -> ByteOff
426 427
                  -> FCode a
                  -> FCode a
428
cannedGCReturnsTo checkYield cont_on_stack gc regs lret off code
429 430
  = do dflags <- getDynFlags
       updfr_sz <- getUpdFrameOff
431
       heapCheck False checkYield (gc_call dflags gc updfr_sz) code
432
  where
433
    reg_exprs = map (CmmReg . CmmLocal) regs
434
      -- Note [stg_gc arguments]
435

436 437 438 439
      -- NB. we use the NativeReturn convention for passing arguments
      -- to the canned heap-check routines, because we are in a case
      -- alternative and hence the [LocalReg] was passed to us in the
      -- NativeReturn convention.
440
    gc_call dflags label sp
441 442 443 444
      | cont_on_stack
      = mkJumpReturnsTo dflags label NativeReturn reg_exprs lret off sp
      | otherwise
      = mkCallReturnsTo dflags label NativeReturn reg_exprs lret off sp []
445

446 447
genericGC :: Bool -> FCode a -> FCode a
genericGC checkYield code
448
  = do updfr_sz <- getUpdFrameOff
449
       lretry <- newBlockId
450
       emitLabel lretry
451
       call <- mkCall generic_gc (GC, GC) [] [] updfr_sz []
452
       heapCheck False checkYield (call <*> mkBranch lretry) code
453

454 455
cannedGCEntryPoint :: DynFlags -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint dflags regs
456
  = case map localRegType regs of
457
      []  -> Just (mkGcLabel "stg_gc_noregs")
458
      [ty]
459 460 461 462 463
          | 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
464

465 466 467
          | width == wordWidth dflags -> Just (mkGcLabel "stg_gc_unbx_r1")
          | width == W64              -> Just (mkGcLabel "stg_gc_l1")
          | otherwise                 -> Nothing
468 469
          where
              width = typeWidth ty
470 471 472 473 474 475 476 477 478 479 480 481
      [ty1,ty2]
          |  isGcPtrType ty1
          && isGcPtrType ty2 -> Just (mkGcLabel "stg_gc_pp")
      [ty1,ty2,ty3]
          |  isGcPtrType ty1
          && isGcPtrType ty2
          && isGcPtrType ty3 -> Just (mkGcLabel "stg_gc_ppp")
      [ty1,ty2,ty3,ty4]
          |  isGcPtrType ty1
          && isGcPtrType ty2
          && isGcPtrType ty3
          && isGcPtrType ty4 -> Just (mkGcLabel "stg_gc_pppp")
482
      _otherwise -> Nothing
483

484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500
-- 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.

501

502 503
-- | The generic GC procedure; no params, no results
generic_gc :: CmmExpr
504
generic_gc = mkGcLabel "stg_gc_noregs"
505 506

-- | Create a CLabel for calling a garbage collector entry point
507
mkGcLabel :: String -> CmmExpr
508
mkGcLabel s = CmmLit (CmmLabel (mkCmmCodeLabel rtsUnitId (fsLit s)))
509 510

-------------------------------
511 512
heapCheck :: Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck checkStack checkYield do_gc code
513
  = getHeapUsage $ \ hpHw ->
514 515
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
516 517
    do  { dflags <- getDynFlags
        ; let mb_alloc_bytes
518 519 520 521
                 | hpHw > mBLOCK_SIZE = sorry $ unlines
                    [" Trying to allocate more than "++show mBLOCK_SIZE++" bytes.",
                     "",
                     "This is currently not possible due to a limitation of GHC's code generator.",
522
                     "See https://gitlab.haskell.org/ghc/ghc/issues/4505 for details.",
523 524
                     "Suggestion: read data from a file instead of having large static data",
                     "structures in code."]
525 526
                 | hpHw > 0  = Just (mkIntExpr dflags (hpHw * (wORD_SIZE dflags)))
                 | otherwise = Nothing
527
                 where mBLOCK_SIZE = bLOCKS_PER_MBLOCK dflags * bLOCK_SIZE_W dflags
528 529 530
              stk_hwm | checkStack = Just (CmmLit CmmHighStackMark)
                      | otherwise  = Nothing
        ; codeOnly $ do_checks stk_hwm checkYield mb_alloc_bytes do_gc
nfrisby's avatar
nfrisby committed
531
        ; tickyAllocHeap True hpHw
532 533
        ; setRealHp hpHw
        ; code }
534

535 536 537
heapStackCheckGen :: Maybe CmmExpr -> Maybe CmmExpr -> FCode ()
heapStackCheckGen stk_hwm mb_bytes
  = do updfr_sz <- getUpdFrameOff
538
       lretry <- newBlockId
539 540
       emitLabel lretry
       call <- mkCall generic_gc (GC, GC) [] [] updfr_sz []
541
       do_checks stk_hwm False mb_bytes (call <*> mkBranch lretry)
542

543 544
-- Note [Single stack check]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~
545 546 547
-- When compiling a function we can determine how much stack space it
-- will use. We therefore need to perform only a single stack check at
-- the beginning of a function to see if we have enough stack space.
548
--
549 550 551 552 553
-- The check boils down to comparing Sp-N with SpLim, where N is the
-- amount of stack space needed (see Note [Stack usage] below).  *BUT*
-- at this stage of the pipeline we are not supposed to refer to Sp
-- itself, because the stack is not yet manifest, so we don't quite
-- know where Sp pointing.
554 555 556 557 558

-- So instead of referring directly to Sp - as we used to do in the
-- past - the code generator uses (old + 0) in the stack check. That
-- is the address of the first word of the old area, so if we add N
-- we'll get the address of highest used word.
559
--
560 561 562 563 564 565 566 567 568 569 570 571
-- This makes the check robust.  For example, while we need to perform
-- only one stack check for each function, we could in theory place
-- more stack checks later in the function. They would be redundant,
-- but not incorrect (in a sense that they should not change program
-- behaviour). We need to make sure however that a stack check
-- inserted after incrementing the stack pointer checks for a
-- respectively smaller stack space. This would not be the case if the
-- code generator produced direct references to Sp. By referencing
-- (old + 0) we make sure that we always check for a correct amount of
-- stack: when converting (old + 0) to Sp the stack layout phase takes
-- into account changes already made to stack pointer. The idea for
-- this change came from observations made while debugging #8275.
572

573 574 575 576 577 578
-- Note [Stack usage]
-- ~~~~~~~~~~~~~~~~~~
-- At the moment we convert from STG to Cmm we don't know N, the
-- number of bytes of stack that the function will use, so we use a
-- special late-bound CmmLit, namely
--       CmmHighStackMark
579
-- to stand for the number of bytes needed. When the stack is made
580 581 582
-- manifest, the number of bytes needed is calculated, and used to
-- replace occurrences of CmmHighStackMark
--
583
-- The (Maybe CmmExpr) passed to do_checks is usually
584 585 586 587 588
--     Just (CmmLit CmmHighStackMark)
-- but can also (in certain hand-written RTS functions)
--     Just (CmmLit 8)  or some other fixed valuet
-- If it is Nothing, we don't generate a stack check at all.

589
do_checks :: Maybe CmmExpr    -- Should we check the stack?
590 591
                              -- See Note [Stack usage]
          -> Bool             -- Should we check for preemption?
592
          -> Maybe CmmExpr    -- Heap headroom (bytes)
593
          -> CmmAGraph        -- What to do on failure
594
          -> FCode ()
595
do_checks mb_stk_hwm checkYield mb_alloc_lit do_gc = do
596
  dflags <- getDynFlags
597
  gc_id <- newBlockId
598

599
  let
600 601
    Just alloc_lit = mb_alloc_lit

602
    bump_hp   = cmmOffsetExprB dflags hpExpr alloc_lit
603

604 605 606
    -- Sp overflow if ((old + 0) - CmmHighStack < SpLim)
    -- At the beginning of a function old + 0 = Sp
    -- See Note [Single stack check]
607 608
    sp_oflo sp_hwm =
         CmmMachOp (mo_wordULt dflags)
609
                  [CmmMachOp (MO_Sub (typeWidth (cmmRegType dflags spReg)))
610
                             [CmmStackSlot Old 0, sp_hwm],
611 612 613 614 615
                   CmmReg spLimReg]

    -- Hp overflow if (Hp > HpLim)
    -- (Hp has been incremented by now)
    -- HpLim points to the LAST WORD of valid allocation space.
616
    hp_oflo = CmmMachOp (mo_wordUGt dflags) [hpExpr, hpLimExpr]
617

618
    alloc_n = mkAssign hpAllocReg alloc_lit
619

620 621
  case mb_stk_hwm of
    Nothing -> return ()
622 623
    Just stk_hwm -> tickyStackCheck
      >> (emit =<< mkCmmIfGoto' (sp_oflo stk_hwm) gc_id (Just False) )
624

625 626 627 628 629 630 631 632 633
  -- Emit new label that might potentially be a header
  -- of a self-recursive tail call.
  -- See Note [Self-recursive loop header].
  self_loop_info <- getSelfLoop
  case self_loop_info of
    Just (_, loop_header_id, _)
        | checkYield && isJust mb_stk_hwm -> emitLabel loop_header_id
    _otherwise -> return ()

634
  if (isJust mb_alloc_lit)
635
    then do
636
     tickyHeapCheck
637
     emitAssign hpReg bump_hp
638
     emit =<< mkCmmIfThen' hp_oflo (alloc_n <*> mkBranch gc_id) (Just False)
639
    else do
640
      when (checkYield && not (gopt Opt_OmitYields dflags)) $ do
641 642
         -- Yielding if HpLim == 0
         let yielding = CmmMachOp (mo_wordEq dflags)
643
                                  [CmmReg hpLimReg,
644
                                   CmmLit (zeroCLit dflags)]
645
         emit =<< mkCmmIfGoto' yielding gc_id (Just False)
646

Peter Wortmann's avatar
Peter Wortmann committed
647 648 649
  tscope <- getTickScope
  emitOutOfLine gc_id
   (do_gc, tscope) -- this is expected to jump back somewhere
650

651 652 653 654 655 656
                -- 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.
657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680

-- Note [Self-recursive loop header]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
-- Self-recursive loop header is required by loopification optimization (See
-- Note [Self-recursive tail calls] in StgCmmExpr). We emit it if:
--
--  1. There is information about self-loop in the FCode environment. We don't
--     check the binder (first component of the self_loop_info) because we are
--     certain that if the self-loop info is present then we are compiling the
--     binder body. Reason: the only possible way to get here with the
--     self_loop_info present is from closureCodeBody.
--
--  2. checkYield && isJust mb_stk_hwm. checkYield tells us that it is possible
--     to preempt the heap check (see #367 for motivation behind this check). It
--     is True for heap checks placed at the entry to a function and
--     let-no-escape heap checks but false for other heap checks (eg. in case
--     alternatives or created from hand-written high-level Cmm). The second
--     check (isJust mb_stk_hwm) is true for heap checks at the entry to a
--     function and some heap checks created in hand-written Cmm. Otherwise it
--     is Nothing. In other words the only situation when both conditions are
--     true is when compiling stack and heap checks at the entry to a
--     function. This is the only situation when we want to emit a self-loop
--     label.