StgCmmHeap.hs 24.7 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, noEscapeHeapCheck, altHeapCheckReturnsTo,
14 15
        heapStackCheckGen,
        entryHeapCheck',
16

17
        mkStaticClosureFields, mkStaticClosure,
18

19
        allocDynClosure, allocDynClosureCmm, allocHeapClosure,
20
        emitSetDynHdr
21 22 23 24 25 26 27 28 29
    ) where

#include "HsVersions.h"

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
38
import SMRep
39
import Cmm
40 41
import CmmUtils
import CostCentre
42
import IdInfo( CafInfo(..), mayHaveCafRefs )
43
import Id ( Id )
44
import Module
45
import DynFlags
46
import FastString( mkFastString, fsLit )
47

48
import Control.Monad (when)
49
import Data.Maybe (isJust)
50

51
-----------------------------------------------------------
52
--              Initialise dynamic heap objects
53 54 55
-----------------------------------------------------------

allocDynClosure
56 57
        :: Maybe Id
        -> CmmInfoTable
Simon Marlow's avatar
Simon Marlow committed
58
        -> 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
allocDynClosureCmm
69
        :: Maybe Id -> CmmInfoTable -> LambdaFormInfo -> CmmExpr -> CmmExpr
70
        -> [(CmmExpr, ByteOff)]
71 72
        -> 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


91 92 93 94 95
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)
96 97


98 99 100 101 102 103 104
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
  profDynAlloc rep use_cc
  let info_ptr = CmmLit (CmmLabel (cit_lbl info_tbl))
  allocHeapClosure rep info_ptr use_cc amodes_w_offsets
105 106


107 108 109 110 111 112 113 114 115
-- | Low-level heap object allocation.
allocHeapClosure
  :: SMRep                            -- ^ representation of the object
  -> CmmExpr                          -- ^ info pointer
  -> CmmExpr                          -- ^ cost centre
  -> [(CmmExpr,ByteOff)]              -- ^ payload
  -> FCode CmmExpr                    -- ^ returns the address of the object
allocHeapClosure rep info_ptr use_cc payload = do
  virt_hp <- getVirtHp
116

117 118 119 120 121 122
  -- 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.
123

124 125 126 127 128 129 130 131 132 133 134 135
  base <- getHpRelOffset info_offset
  emitComment $ mkFastString "allocDynClosure"
  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
136

137 138

emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
139
emitSetDynHdr base info_ptr ccs
140
  = do dflags <- getDynFlags
141
       hpStore base (zip (header dflags) [0, wORD_SIZE dflags ..])
142
  where
143 144
    header :: DynFlags -> [CmmExpr]
    header dflags = [info_ptr] ++ dynProfHdr dflags ccs
Jan Stolarek's avatar
Jan Stolarek committed
145
        -- ToDof: Parallel stuff
146
        -- No ticky header
147 148

-- Store the item (expr,off) in base[off]
149 150 151 152 153
hpStore :: CmmExpr -> [(CmmExpr, ByteOff)] -> FCode ()
hpStore base vals = do
  dflags <- getDynFlags
  sequence_ $
    [ emitStore (cmmOffsetB dflags base off) val | (val,off) <- vals ]
154 155

-----------------------------------------------------------
156
--              Layout of static closures
157 158 159 160 161
-----------------------------------------------------------

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

162
mkStaticClosureFields
163 164
        :: DynFlags
        -> CmmInfoTable
165
        -> CostCentreStack
166
        -> CafInfo
167 168
        -> [CmmLit]             -- Payload
        -> [CmmLit]             -- The full closure
169 170
mkStaticClosureFields dflags info_tbl ccs caf_refs payload
  = mkStaticClosure dflags info_lbl ccs payload padding
171
        static_link_field saved_info_field
172
  where
Simon Marlow's avatar
Simon Marlow committed
173
    info_lbl = cit_lbl info_tbl
174 175 176 177 178 179 180 181 182

    -- 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
183 184 185
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:
186

Simon Marlow's avatar
Simon Marlow committed
187
    is_caf = isThunkRep (cit_rep info_tbl)
188

189
    padding
190 191
        | is_caf && null payload = [mkIntCLit dflags 0]
        | otherwise = []
192 193

    static_link_field
194
        | is_caf || staticClosureNeedsLink (mayHaveCafRefs caf_refs) info_tbl
Simon Marlow's avatar
Simon Marlow committed
195 196 197
        = [static_link_value]
        | otherwise
        = []
198 199

    saved_info_field
200
        | is_caf     = [mkIntCLit dflags 0]
201
        | otherwise  = []
202

203
        -- For a static constructor which has NoCafRefs, we set the
204 205
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
206
    static_link_value
207 208
        | mayHaveCafRefs caf_refs  = mkIntCLit dflags 0
        | otherwise                = mkIntCLit dflags 1  -- No CAF refs
209 210


211
mkStaticClosure :: DynFlags -> CLabel -> CostCentreStack -> [CmmLit]
212
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
213
mkStaticClosure dflags info_lbl ccs payload padding static_link_field saved_info_field
214
  =  [CmmLabel info_lbl]
Jan Stolarek's avatar
Jan Stolarek committed
215
  ++ staticProfHdr dflags ccs
216
  ++ concatMap (padLitToWord dflags) payload
217
  ++ padding
218 219 220
  ++ static_link_field
  ++ saved_info_field

221 222
-- 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?
223 224 225
padLitToWord :: DynFlags -> CmmLit -> [CmmLit]
padLitToWord dflags lit = lit : padding pad_length
  where width = typeWidth (cmmLitType dflags lit)
226
        pad_length = wORD_SIZE dflags - widthInBytes width :: Int
227 228 229 230 231 232 233

        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)

234
-----------------------------------------------------------
235
--              Heap overflow checking
236 237 238 239 240 241 242 243 244 245 246 247
-----------------------------------------------------------

{- 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
248
        r = gc_1p( r )
249 250
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.
251

252
  * A series of canned entry points like
253
        gcfun_2p( f, x, y )
254 255 256 257 258 259 260 261 262
    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 ) {
263 264 265
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
266 267 268
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
269
     it should do no register shuffling
270 271 272

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
273 274 275
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
276
          L: HpAlloc = 8
277 278
             call gc()  -- Needs an info table
             goto T }
279 280

* altHeapCheck: Immediately following an eval
281 282
  Started as
        case f x y of r { (p,q) -> rhs }
283 284 285
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
286 287 288 289 290 291
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
292

293 294
        L: r = gc_1p( r )
           goto K }
295

296 297 298 299
        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)
300 301 302

   (b) No canned sequence for results of f
       Note second info table
303 304 305 306 307 308
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
309

310 311
        L: call gc()    -- Extra info table here
           goto K
312 313 314

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
315
       case branch *not* following eval,
316 317 318
       or let-no-escape
  Exactly the same as the previous case:

319 320 321 322
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...
323

324 325
        L: call gc()
           goto K
326 327 328 329 330
-}

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

331 332 333 334 335 336
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 ()
337

338
entryHeapCheck cl_info nodeSet arity args code
339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
  = 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
357 358
  = do dflags <- getDynFlags
       let is_thunk = arity == 0
359 360

           args' = map (CmmReg . CmmLocal) args
361 362 363 364 365 366 367
           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)

368
              Function (slow): call (slow) stg_gc_fun(fun, args)
369 370 371
           -}
           gc_call upd
               | is_thunk
372
                 = mkJump dflags NativeNodeCall stg_gc_enter1 [node] upd
373 374

               | is_fastf
375
                 = mkJump dflags NativeNodeCall stg_gc_fun (node : args') upd
376 377

               | otherwise
378
                 = mkJump dflags Slow stg_gc_fun (node : args') upd
379 380

       updfr_sz <- getUpdFrameOff
381 382 383

       loop_id <- newLabelC
       emitLabel loop_id
384
       heapCheck True True (gc_call updfr_sz <*> mkBranch loop_id) code
385

386 387
-- ------------------------------------------------------------
-- A heap/stack check in a case alternative
388

389 390 391 392 393 394 395 396 397 398 399 400 401 402 403

-- 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

404
altHeapCheck :: [LocalReg] -> FCode a -> FCode a
405 406 407 408
altHeapCheck regs code = altOrNoEscapeHeapCheck False regs code

altOrNoEscapeHeapCheck :: Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck checkYield regs code = do
409 410
    dflags <- getDynFlags
    case cannedGCEntryPoint dflags regs of
411
      Nothing -> genericGC checkYield code
412 413
      Just gc -> do
        lret <- newLabelC
414
        let (off, _, copyin) = copyInOflow dflags NativeReturn (Young lret) regs []
415 416 417
        lcont <- newLabelC
        emitOutOfLine lret (copyin <*> mkBranch lcont)
        emitLabel lcont
418
        cannedGCReturnsTo checkYield False gc regs lret off code
419 420 421

altHeapCheckReturnsTo :: [LocalReg] -> Label -> ByteOff -> FCode a -> FCode a
altHeapCheckReturnsTo regs lret off code
422 423
  = do dflags <- getDynFlags
       case cannedGCEntryPoint dflags regs of
424 425 426 427 428 429 430 431
           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
432

433
cannedGCReturnsTo :: Bool -> Bool -> CmmExpr -> [LocalReg] -> Label -> ByteOff
434 435
                  -> FCode a
                  -> FCode a
436
cannedGCReturnsTo checkYield cont_on_stack gc regs lret off code
437 438
  = do dflags <- getDynFlags
       updfr_sz <- getUpdFrameOff
439
       heapCheck False checkYield (gc_call dflags gc updfr_sz) code
440 441
  where
    reg_exprs = map (CmmReg . CmmLocal) regs
442
      -- Note [stg_gc arguments]
443

444 445 446 447
      -- 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.
448
    gc_call dflags label sp
449 450 451 452
      | cont_on_stack
      = mkJumpReturnsTo dflags label NativeReturn reg_exprs lret off sp
      | otherwise
      = mkCallReturnsTo dflags label NativeReturn reg_exprs lret off sp []
453

454 455
genericGC :: Bool -> FCode a -> FCode a
genericGC checkYield code
456 457 458
  = do updfr_sz <- getUpdFrameOff
       lretry <- newLabelC
       emitLabel lretry
459
       call <- mkCall generic_gc (GC, GC) [] [] updfr_sz []
460
       heapCheck False checkYield (call <*> mkBranch lretry) code
461

462 463
cannedGCEntryPoint :: DynFlags -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint dflags regs
464
  = case map localRegType regs of
465
      []  -> Just (mkGcLabel "stg_gc_noregs")
466
      [ty]
467 468 469 470 471
          | 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
472

473 474 475
          | width == wordWidth dflags -> Just (mkGcLabel "stg_gc_unbx_r1")
          | width == W64              -> Just (mkGcLabel "stg_gc_l1")
          | otherwise                 -> Nothing
476 477
          where
              width = typeWidth ty
478 479 480 481 482 483 484 485 486 487 488 489
      [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")
490
      _otherwise -> Nothing
491

492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508
-- 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.

509

510 511
-- | The generic GC procedure; no params, no results
generic_gc :: CmmExpr
512
generic_gc = mkGcLabel "stg_gc_noregs"
513 514

-- | Create a CLabel for calling a garbage collector entry point
515 516
mkGcLabel :: String -> CmmExpr
mkGcLabel s = CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit s)))
517 518

-------------------------------
519 520
heapCheck :: Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck checkStack checkYield do_gc code
521
  = getHeapUsage $ \ hpHw ->
522 523
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
524 525 526 527 528 529 530
    do  { dflags <- getDynFlags
        ; let mb_alloc_bytes
                 | hpHw > 0  = Just (mkIntExpr dflags (hpHw * (wORD_SIZE dflags)))
                 | otherwise = Nothing
              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

535 536 537 538 539 540
heapStackCheckGen :: Maybe CmmExpr -> Maybe CmmExpr -> FCode ()
heapStackCheckGen stk_hwm mb_bytes
  = do updfr_sz <- getUpdFrameOff
       lretry <- newLabelC
       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
597 598
  gc_id <- newLabelC

599
  let
600 601 602
    Just alloc_lit = mb_alloc_lit

    bump_hp   = cmmOffsetExprB dflags (CmmReg hpReg) 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],
611 612 613 614 615 616
                   CmmReg spLimReg]

    -- 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 dflags)
617
                  [CmmReg hpReg, CmmReg (CmmGlobal HpLim)]
618

619
    alloc_n = mkAssign (CmmGlobal HpAlloc) alloc_lit
620

621 622
  case mb_stk_hwm of
    Nothing -> return ()
623
    Just stk_hwm -> tickyStackCheck >> (emit =<< mkCmmIfGoto (sp_oflo stk_hwm) gc_id)
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
637 638
     emitAssign hpReg bump_hp
     emit =<< mkCmmIfThen hp_oflo (alloc_n <*> mkBranch gc_id)
639
    else do
640
      when (checkYield && not (gopt Opt_OmitYields dflags)) $ do
641 642 643 644 645
         -- Yielding if HpLim == 0
         let yielding = CmmMachOp (mo_wordEq dflags)
                                  [CmmReg (CmmGlobal HpLim),
                                   CmmLit (zeroCLit dflags)]
         emit =<< mkCmmIfGoto yielding gc_id
646 647

  emitOutOfLine gc_id $
648 649
     do_gc -- this is expected to jump back somewhere

650 651 652 653 654 655
                -- 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.
656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679

-- 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.