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
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
104 105


106 107 108 109 110 111 112 113
-- | 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
114 115
  profDynAlloc rep use_cc

116
  virt_hp <- getVirtHp
117

118 119 120 121 122 123
  -- 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.
124

125
  base <- getHpRelOffset info_offset
126
  emitComment $ mkFastString "allocHeapClosure"
127 128 129 130 131 132 133 134 135 136
  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
137

138 139

emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
140
emitSetDynHdr base info_ptr ccs
141
  = do dflags <- getDynFlags
142
       hpStore base (zip (header dflags) [0, wORD_SIZE dflags ..])
143
  where
144 145
    header :: DynFlags -> [CmmExpr]
    header dflags = [info_ptr] ++ dynProfHdr dflags ccs
Jan Stolarek's avatar
Jan Stolarek committed
146
        -- ToDof: Parallel stuff
147
        -- No ticky header
148 149

-- Store the item (expr,off) in base[off]
150 151 152 153 154
hpStore :: CmmExpr -> [(CmmExpr, ByteOff)] -> FCode ()
hpStore base vals = do
  dflags <- getDynFlags
  sequence_ $
    [ emitStore (cmmOffsetB dflags base off) val | (val,off) <- vals ]
155 156

-----------------------------------------------------------
157
--              Layout of static closures
158 159 160 161 162
-----------------------------------------------------------

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

163
mkStaticClosureFields
164 165
        :: DynFlags
        -> CmmInfoTable
166
        -> CostCentreStack
167
        -> CafInfo
168 169
        -> [CmmLit]             -- Payload
        -> [CmmLit]             -- The full closure
170 171
mkStaticClosureFields dflags info_tbl ccs caf_refs payload
  = mkStaticClosure dflags info_lbl ccs payload padding
172
        static_link_field saved_info_field
173
  where
Simon Marlow's avatar
Simon Marlow committed
174
    info_lbl = cit_lbl info_tbl
175 176 177 178 179 180 181 182 183

    -- 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
184 185 186
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:
187

Simon Marlow's avatar
Simon Marlow committed
188
    is_caf = isThunkRep (cit_rep info_tbl)
189

190
    padding
191 192
        | is_caf && null payload = [mkIntCLit dflags 0]
        | otherwise = []
193 194

    static_link_field
195
        | is_caf || staticClosureNeedsLink (mayHaveCafRefs caf_refs) info_tbl
Simon Marlow's avatar
Simon Marlow committed
196 197 198
        = [static_link_value]
        | otherwise
        = []
199 200

    saved_info_field
201
        | is_caf     = [mkIntCLit dflags 0]
202
        | otherwise  = []
203

204
        -- For a static constructor which has NoCafRefs, we set the
205 206
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
207
    static_link_value
208 209
        | mayHaveCafRefs caf_refs  = mkIntCLit dflags 0
        | otherwise                = mkIntCLit dflags 1  -- No CAF refs
210 211


212
mkStaticClosure :: DynFlags -> CLabel -> CostCentreStack -> [CmmLit]
213
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
214
mkStaticClosure dflags info_lbl ccs payload padding static_link_field saved_info_field
215
  =  [CmmLabel info_lbl]
Jan Stolarek's avatar
Jan Stolarek committed
216
  ++ staticProfHdr dflags ccs
217
  ++ concatMap (padLitToWord dflags) payload
218
  ++ padding
219 220 221
  ++ static_link_field
  ++ saved_info_field

222 223
-- 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?
224 225 226
padLitToWord :: DynFlags -> CmmLit -> [CmmLit]
padLitToWord dflags lit = lit : padding pad_length
  where width = typeWidth (cmmLitType dflags lit)
227
        pad_length = wORD_SIZE dflags - widthInBytes width :: Int
228 229 230 231 232 233 234

        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)

235
-----------------------------------------------------------
236
--              Heap overflow checking
237 238 239 240 241 242 243 244 245 246 247 248
-----------------------------------------------------------

{- 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
249
        r = gc_1p( r )
250 251
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.
252

253
  * A series of canned entry points like
254
        gcfun_2p( f, x, y )
255 256 257 258 259 260 261 262 263
    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 ) {
264 265 266
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
267 268 269
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
270
     it should do no register shuffling
271 272 273

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
274 275 276
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
277
          L: HpAlloc = 8
278 279
             call gc()  -- Needs an info table
             goto T }
280 281

* altHeapCheck: Immediately following an eval
282 283
  Started as
        case f x y of r { (p,q) -> rhs }
284 285 286
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
287 288 289 290 291 292
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
293

294 295
        L: r = gc_1p( r )
           goto K }
296

297 298 299 300
        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)
301 302 303

   (b) No canned sequence for results of f
       Note second info table
304 305 306 307 308 309
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
310

311 312
        L: call gc()    -- Extra info table here
           goto K
313 314 315

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
316
       case branch *not* following eval,
317 318 319
       or let-no-escape
  Exactly the same as the previous case:

320 321 322 323
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...
324

325 326
        L: call gc()
           goto K
327 328 329 330 331
-}

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

332 333 334 335 336 337
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 ()
338

339
entryHeapCheck cl_info nodeSet arity args code
340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357
  = 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
358 359
  = do dflags <- getDynFlags
       let is_thunk = arity == 0
360 361

           args' = map (CmmReg . CmmLocal) args
362 363 364 365 366 367 368
           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)

369
              Function (slow): call (slow) stg_gc_fun(fun, args)
370 371 372
           -}
           gc_call upd
               | is_thunk
373
                 = mkJump dflags NativeNodeCall stg_gc_enter1 [node] upd
374 375

               | is_fastf
376
                 = mkJump dflags NativeNodeCall stg_gc_fun (node : args') upd
377 378

               | otherwise
379
                 = mkJump dflags Slow stg_gc_fun (node : args') upd
380 381

       updfr_sz <- getUpdFrameOff
382 383 384

       loop_id <- newLabelC
       emitLabel loop_id
385
       heapCheck True True (gc_call updfr_sz <*> mkBranch loop_id) code
386

387 388
-- ------------------------------------------------------------
-- A heap/stack check in a case alternative
389

390 391 392 393 394 395 396 397 398 399 400 401 402 403 404

-- 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

405
altHeapCheck :: [LocalReg] -> FCode a -> FCode a
406 407 408 409
altHeapCheck regs code = altOrNoEscapeHeapCheck False regs code

altOrNoEscapeHeapCheck :: Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck checkYield regs code = do
410 411
    dflags <- getDynFlags
    case cannedGCEntryPoint dflags regs of
412
      Nothing -> genericGC checkYield code
413 414
      Just gc -> do
        lret <- newLabelC
415
        let (off, _, copyin) = copyInOflow dflags NativeReturn (Young lret) regs []
416 417 418
        lcont <- newLabelC
        emitOutOfLine lret (copyin <*> mkBranch lcont)
        emitLabel lcont
419
        cannedGCReturnsTo checkYield False gc regs lret off code
420 421 422

altHeapCheckReturnsTo :: [LocalReg] -> Label -> ByteOff -> FCode a -> FCode a
altHeapCheckReturnsTo regs lret off code
423 424
  = do dflags <- getDynFlags
       case cannedGCEntryPoint dflags regs of
425 426 427 428 429 430 431 432
           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
433

434
cannedGCReturnsTo :: Bool -> Bool -> CmmExpr -> [LocalReg] -> Label -> ByteOff
435 436
                  -> FCode a
                  -> FCode a
437
cannedGCReturnsTo checkYield cont_on_stack gc regs lret off code
438 439
  = do dflags <- getDynFlags
       updfr_sz <- getUpdFrameOff
440
       heapCheck False checkYield (gc_call dflags gc updfr_sz) code
441 442
  where
    reg_exprs = map (CmmReg . CmmLocal) regs
443
      -- Note [stg_gc arguments]
444

445 446 447 448
      -- 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.
449
    gc_call dflags label sp
450 451 452 453
      | cont_on_stack
      = mkJumpReturnsTo dflags label NativeReturn reg_exprs lret off sp
      | otherwise
      = mkCallReturnsTo dflags label NativeReturn reg_exprs lret off sp []
454

455 456
genericGC :: Bool -> FCode a -> FCode a
genericGC checkYield code
457 458 459
  = do updfr_sz <- getUpdFrameOff
       lretry <- newLabelC
       emitLabel lretry
460
       call <- mkCall generic_gc (GC, GC) [] [] updfr_sz []
461
       heapCheck False checkYield (call <*> mkBranch lretry) code
462

463 464
cannedGCEntryPoint :: DynFlags -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint dflags regs
465
  = case map localRegType regs of
466
      []  -> Just (mkGcLabel "stg_gc_noregs")
467
      [ty]
468 469 470 471 472
          | 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
473

474 475 476
          | width == wordWidth dflags -> Just (mkGcLabel "stg_gc_unbx_r1")
          | width == W64              -> Just (mkGcLabel "stg_gc_l1")
          | otherwise                 -> Nothing
477 478
          where
              width = typeWidth ty
479 480 481 482 483 484 485 486 487 488 489 490
      [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")
491
      _otherwise -> Nothing
492

493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509
-- 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.

510

511 512
-- | The generic GC procedure; no params, no results
generic_gc :: CmmExpr
513
generic_gc = mkGcLabel "stg_gc_noregs"
514 515

-- | Create a CLabel for calling a garbage collector entry point
516 517
mkGcLabel :: String -> CmmExpr
mkGcLabel s = CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit s)))
518 519

-------------------------------
520 521
heapCheck :: Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck checkStack checkYield do_gc code
522
  = getHeapUsage $ \ hpHw ->
523 524
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
525 526 527 528 529 530 531
    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
532
        ; tickyAllocHeap True hpHw
533 534
        ; setRealHp hpHw
        ; code }
535

536 537 538 539 540 541
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 []
542
       do_checks stk_hwm False mb_bytes (call <*> mkBranch lretry)
543

544 545
-- Note [Single stack check]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~
546 547 548
-- 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.
549
--
550 551 552 553 554
-- 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.
555 556 557 558 559

-- 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.
560
--
561 562 563 564 565 566 567 568 569 570 571 572
-- 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.
573

574 575 576 577 578 579
-- 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
580
-- to stand for the number of bytes needed. When the stack is made
581 582 583
-- manifest, the number of bytes needed is calculated, and used to
-- replace occurrences of CmmHighStackMark
--
584
-- The (Maybe CmmExpr) passed to do_checks is usually
585 586 587 588 589
--     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.

590
do_checks :: Maybe CmmExpr    -- Should we check the stack?
591 592
                              -- See Note [Stack usage]
          -> Bool             -- Should we check for preemption?
593
          -> Maybe CmmExpr    -- Heap headroom (bytes)
594
          -> CmmAGraph        -- What to do on failure
595
          -> FCode ()
596
do_checks mb_stk_hwm checkYield mb_alloc_lit do_gc = do
597
  dflags <- getDynFlags
598 599
  gc_id <- newLabelC

600
  let
601 602 603
    Just alloc_lit = mb_alloc_lit

    bump_hp   = cmmOffsetExprB dflags (CmmReg hpReg) alloc_lit
604

605 606 607
    -- Sp overflow if ((old + 0) - CmmHighStack < SpLim)
    -- At the beginning of a function old + 0 = Sp
    -- See Note [Single stack check]
608 609
    sp_oflo sp_hwm =
         CmmMachOp (mo_wordULt dflags)
610
                  [CmmMachOp (MO_Sub (typeWidth (cmmRegType dflags spReg)))
611
                             [CmmStackSlot Old 0, sp_hwm],
612 613 614 615 616 617
                   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)
618
                  [CmmReg hpReg, CmmReg (CmmGlobal HpLim)]
619

620
    alloc_n = mkAssign (CmmGlobal HpAlloc) alloc_lit
621

622 623
  case mb_stk_hwm of
    Nothing -> return ()
624
    Just stk_hwm -> tickyStackCheck >> (emit =<< mkCmmIfGoto (sp_oflo stk_hwm) gc_id)
625

626 627 628 629 630 631 632 633 634
  -- 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 ()

635
  if (isJust mb_alloc_lit)
636
    then do
637
     tickyHeapCheck
638 639
     emitAssign hpReg bump_hp
     emit =<< mkCmmIfThen hp_oflo (alloc_n <*> mkBranch gc_id)
640
    else do
641
      when (checkYield && not (gopt Opt_OmitYields dflags)) $ do
642 643 644 645 646
         -- Yielding if HpLim == 0
         let yielding = CmmMachOp (mo_wordEq dflags)
                                  [CmmReg (CmmGlobal HpLim),
                                   CmmLit (zeroCLit dflags)]
         emit =<< mkCmmIfGoto yielding gc_id
647 648

  emitOutOfLine gc_id $
649 650
     do_gc -- this is expected to jump back somewhere

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.