StgCmmHeap.hs 25.6 KB
Newer Older
1 2
{-# LANGUAGE CPP #-}

3 4 5 6 7 8 9 10 11
-----------------------------------------------------------------------------
--
-- Stg to C--: heap management functions
--
-- (c) The University of Glasgow 2004-2006
--
-----------------------------------------------------------------------------

module StgCmmHeap (
12
        getVirtHp, setVirtHp, setRealHp,
13
        getHpRelOffset,
14

15
        entryHeapCheck, altHeapCheck, noEscapeHeapCheck, altHeapCheckReturnsTo,
16 17
        heapStackCheckGen,
        entryHeapCheck',
18

19
        mkStaticClosureFields, mkStaticClosure,
20

21
        allocDynClosure, allocDynClosureCmm, allocHeapClosure,
22
        emitSetDynHdr
23 24 25 26 27 28 29 30 31
    ) where

#include "HsVersions.h"

import StgSyn
import CLabel
import StgCmmLayout
import StgCmmUtils
import StgCmmMonad
32
import StgCmmProf (profDynAlloc, dynProfHdr, staticProfHdr)
33 34 35 36
import StgCmmTicky
import StgCmmClosure
import StgCmmEnv

37
import MkGraph
38

39
import Hoopl
40
import SMRep
41
import Cmm
42 43
import CmmUtils
import CostCentre
44
import IdInfo( CafInfo(..), mayHaveCafRefs )
45
import Id ( Id )
46
import Module
47
import DynFlags
48
import FastString( mkFastString, fsLit )
49
import Panic( sorry )
50

51 52 53 54
#if __GLASGOW_HASKELL__ >= 709
import Prelude hiding ((<*>))
#endif

55
import Control.Monad (when)
56
import Data.Maybe (isJust)
57

58
-----------------------------------------------------------
59
--              Initialise dynamic heap objects
60 61 62
-----------------------------------------------------------

allocDynClosure
63 64
        :: Maybe Id
        -> CmmInfoTable
Simon Marlow's avatar
Simon Marlow committed
65
        -> LambdaFormInfo
66 67 68 69 70 71 72
        -> 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
73
        -> FCode CmmExpr -- returns Hp+n
74

75
allocDynClosureCmm
76
        :: Maybe Id -> CmmInfoTable -> LambdaFormInfo -> CmmExpr -> CmmExpr
77
        -> [(CmmExpr, ByteOff)]
78 79
        -> FCode CmmExpr -- returns Hp+n

80
-- allocDynClosure allocates the thing in the heap,
81
-- and modifies the virtual Hp to account for this.
82 83 84
-- 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.
85

86 87 88 89 90 91 92 93 94 95
-- 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.
96 97


98 99 100 101 102
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)
103 104


105 106 107 108 109 110
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
111 112


113 114 115 116 117 118 119 120
-- | 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
121 122
  profDynAlloc rep use_cc

123
  virt_hp <- getVirtHp
124

125 126 127 128 129 130
  -- 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.
131

132
  base <- getHpRelOffset info_offset
133
  emitComment $ mkFastString "allocHeapClosure"
134 135 136 137 138 139 140 141 142 143
  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
144

145 146

emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
147
emitSetDynHdr base info_ptr ccs
148
  = do dflags <- getDynFlags
149
       hpStore base (zip (header dflags) [0, wORD_SIZE dflags ..])
150
  where
151 152
    header :: DynFlags -> [CmmExpr]
    header dflags = [info_ptr] ++ dynProfHdr dflags ccs
Jan Stolarek's avatar
Jan Stolarek committed
153
        -- ToDof: Parallel stuff
154
        -- No ticky header
155 156

-- Store the item (expr,off) in base[off]
157 158 159 160 161
hpStore :: CmmExpr -> [(CmmExpr, ByteOff)] -> FCode ()
hpStore base vals = do
  dflags <- getDynFlags
  sequence_ $
    [ emitStore (cmmOffsetB dflags base off) val | (val,off) <- vals ]
162 163

-----------------------------------------------------------
164
--              Layout of static closures
165 166 167 168 169
-----------------------------------------------------------

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

170
mkStaticClosureFields
171 172
        :: DynFlags
        -> CmmInfoTable
173
        -> CostCentreStack
174
        -> CafInfo
175 176
        -> [CmmLit]             -- Payload
        -> [CmmLit]             -- The full closure
177 178
mkStaticClosureFields dflags info_tbl ccs caf_refs payload
  = mkStaticClosure dflags info_lbl ccs payload padding
179
        static_link_field saved_info_field
180
  where
Simon Marlow's avatar
Simon Marlow committed
181
    info_lbl = cit_lbl info_tbl
182 183 184 185 186 187 188 189 190

    -- 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
191 192 193
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:
194

Simon Marlow's avatar
Simon Marlow committed
195
    is_caf = isThunkRep (cit_rep info_tbl)
196

197
    padding
198 199
        | is_caf && null payload = [mkIntCLit dflags 0]
        | otherwise = []
200 201

    static_link_field
202
        | is_caf || staticClosureNeedsLink (mayHaveCafRefs caf_refs) info_tbl
Simon Marlow's avatar
Simon Marlow committed
203 204 205
        = [static_link_value]
        | otherwise
        = []
206 207

    saved_info_field
208
        | is_caf     = [mkIntCLit dflags 0]
209
        | otherwise  = []
210

211
        -- For a static constructor which has NoCafRefs, we set the
212 213
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
214
    static_link_value
215
        | mayHaveCafRefs caf_refs  = mkIntCLit dflags 0
216 217 218
        | otherwise                = mkIntCLit dflags 3  -- No CAF refs
                                      -- See Note [STATIC_LINK fields]
                                      -- in rts/sm/Storage.h
219

220
mkStaticClosure :: DynFlags -> CLabel -> CostCentreStack -> [CmmLit]
221
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
222
mkStaticClosure dflags info_lbl ccs payload padding static_link_field saved_info_field
223
  =  [CmmLabel info_lbl]
Jan Stolarek's avatar
Jan Stolarek committed
224
  ++ staticProfHdr dflags ccs
225
  ++ concatMap (padLitToWord dflags) payload
226
  ++ padding
227 228 229
  ++ static_link_field
  ++ saved_info_field

230 231
-- 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?
232 233 234
padLitToWord :: DynFlags -> CmmLit -> [CmmLit]
padLitToWord dflags lit = lit : padding pad_length
  where width = typeWidth (cmmLitType dflags lit)
235
        pad_length = wORD_SIZE dflags - widthInBytes width :: Int
236 237 238 239 240 241 242

        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)

243
-----------------------------------------------------------
244
--              Heap overflow checking
245 246 247 248 249 250 251 252 253 254 255 256
-----------------------------------------------------------

{- 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
257
        r = gc_1p( r )
258 259
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.
260

261
  * A series of canned entry points like
262
        gcfun_2p( f, x, y )
263 264 265 266 267 268 269 270 271
    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 ) {
272 273 274
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
275 276 277
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
278
     it should do no register shuffling
279 280 281

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
282 283 284
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
285
          L: HpAlloc = 8
286 287
             call gc()  -- Needs an info table
             goto T }
288 289

* altHeapCheck: Immediately following an eval
290 291
  Started as
        case f x y of r { (p,q) -> rhs }
292 293 294
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
295 296 297 298 299 300
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
301

302 303
        L: r = gc_1p( r )
           goto K }
304

305 306 307 308
        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)
309 310 311

   (b) No canned sequence for results of f
       Note second info table
312 313 314 315 316 317
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
318

319 320
        L: call gc()    -- Extra info table here
           goto K
321 322 323

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
324
       case branch *not* following eval,
325 326 327
       or let-no-escape
  Exactly the same as the previous case:

328 329 330 331
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...
332

333 334
        L: call gc()
           goto K
335 336 337 338 339
-}

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

340 341 342 343 344 345
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 ()
346

347
entryHeapCheck cl_info nodeSet arity args code
348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365
  = 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
366 367
  = do dflags <- getDynFlags
       let is_thunk = arity == 0
368 369

           args' = map (CmmReg . CmmLocal) args
370 371 372 373 374 375 376
           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)

377
              Function (slow): call (slow) stg_gc_fun(fun, args)
378 379 380
           -}
           gc_call upd
               | is_thunk
381
                 = mkJump dflags NativeNodeCall stg_gc_enter1 [node] upd
382 383

               | is_fastf
384
                 = mkJump dflags NativeNodeCall stg_gc_fun (node : args') upd
385 386

               | otherwise
387
                 = mkJump dflags Slow stg_gc_fun (node : args') upd
388 389

       updfr_sz <- getUpdFrameOff
390 391 392

       loop_id <- newLabelC
       emitLabel loop_id
393
       heapCheck True True (gc_call updfr_sz <*> mkBranch loop_id) code
394

395 396
-- ------------------------------------------------------------
-- A heap/stack check in a case alternative
397

398 399 400 401 402 403 404 405 406 407 408 409 410 411 412

-- 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

413
altHeapCheck :: [LocalReg] -> FCode a -> FCode a
414 415 416 417
altHeapCheck regs code = altOrNoEscapeHeapCheck False regs code

altOrNoEscapeHeapCheck :: Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck checkYield regs code = do
418 419
    dflags <- getDynFlags
    case cannedGCEntryPoint dflags regs of
420
      Nothing -> genericGC checkYield code
421 422
      Just gc -> do
        lret <- newLabelC
423
        let (off, _, copyin) = copyInOflow dflags NativeReturn (Young lret) regs []
424
        lcont <- newLabelC
Peter Wortmann's avatar
Peter Wortmann committed
425 426
        tscope <- getTickScope
        emitOutOfLine lret (copyin <*> mkBranch lcont, tscope)
427
        emitLabel lcont
428
        cannedGCReturnsTo checkYield False gc regs lret off code
429 430 431

altHeapCheckReturnsTo :: [LocalReg] -> Label -> ByteOff -> FCode a -> FCode a
altHeapCheckReturnsTo regs lret off code
432 433
  = do dflags <- getDynFlags
       case cannedGCEntryPoint dflags regs of
434 435 436 437 438 439 440 441
           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
442

443
cannedGCReturnsTo :: Bool -> Bool -> CmmExpr -> [LocalReg] -> Label -> ByteOff
444 445
                  -> FCode a
                  -> FCode a
446
cannedGCReturnsTo checkYield cont_on_stack gc regs lret off code
447 448
  = do dflags <- getDynFlags
       updfr_sz <- getUpdFrameOff
449
       heapCheck False checkYield (gc_call dflags gc updfr_sz) code
450 451
  where
    reg_exprs = map (CmmReg . CmmLocal) regs
452
      -- Note [stg_gc arguments]
453

454 455 456 457
      -- 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.
458
    gc_call dflags label sp
459 460 461 462
      | cont_on_stack
      = mkJumpReturnsTo dflags label NativeReturn reg_exprs lret off sp
      | otherwise
      = mkCallReturnsTo dflags label NativeReturn reg_exprs lret off sp []
463

464 465
genericGC :: Bool -> FCode a -> FCode a
genericGC checkYield code
466 467 468
  = do updfr_sz <- getUpdFrameOff
       lretry <- newLabelC
       emitLabel lretry
469
       call <- mkCall generic_gc (GC, GC) [] [] updfr_sz []
470
       heapCheck False checkYield (call <*> mkBranch lretry) code
471

472 473
cannedGCEntryPoint :: DynFlags -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint dflags regs
474
  = case map localRegType regs of
475
      []  -> Just (mkGcLabel "stg_gc_noregs")
476
      [ty]
477 478 479 480 481
          | 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
482

483 484 485
          | width == wordWidth dflags -> Just (mkGcLabel "stg_gc_unbx_r1")
          | width == W64              -> Just (mkGcLabel "stg_gc_l1")
          | otherwise                 -> Nothing
486 487
          where
              width = typeWidth ty
488 489 490 491 492 493 494 495 496 497 498 499
      [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")
500
      _otherwise -> Nothing
501

502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
-- 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.

519

520 521
-- | The generic GC procedure; no params, no results
generic_gc :: CmmExpr
522
generic_gc = mkGcLabel "stg_gc_noregs"
523 524

-- | Create a CLabel for calling a garbage collector entry point
525
mkGcLabel :: String -> CmmExpr
526
mkGcLabel s = CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageKey (fsLit s)))
527 528

-------------------------------
529 530
heapCheck :: Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck checkStack checkYield do_gc code
531
  = getHeapUsage $ \ hpHw ->
532 533
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
534 535
    do  { dflags <- getDynFlags
        ; let mb_alloc_bytes
536 537 538 539 540 541 542
                 | 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.",
                     "See http://hackage.haskell.org/trac/ghc/ticket/4505 for details.",
                     "Suggestion: read data from a file instead of having large static data",
                     "structures in code."]
543 544
                 | hpHw > 0  = Just (mkIntExpr dflags (hpHw * (wORD_SIZE dflags)))
                 | otherwise = Nothing
545
                 where mBLOCK_SIZE = bLOCKS_PER_MBLOCK dflags * bLOCK_SIZE_W dflags
546 547 548
              stk_hwm | checkStack = Just (CmmLit CmmHighStackMark)
                      | otherwise  = Nothing
        ; codeOnly $ do_checks stk_hwm checkYield mb_alloc_bytes do_gc
nfrisby's avatar
nfrisby committed
549
        ; tickyAllocHeap True hpHw
550 551
        ; setRealHp hpHw
        ; code }
552

553 554 555 556 557 558
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 []
559
       do_checks stk_hwm False mb_bytes (call <*> mkBranch lretry)
560

561 562
-- Note [Single stack check]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~
563 564 565
-- 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.
566
--
567 568 569 570 571
-- 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.
572 573 574 575 576

-- 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.
577
--
578 579 580 581 582 583 584 585 586 587 588 589
-- 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.
590

591 592 593 594 595 596
-- 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
597
-- to stand for the number of bytes needed. When the stack is made
598 599 600
-- manifest, the number of bytes needed is calculated, and used to
-- replace occurrences of CmmHighStackMark
--
601
-- The (Maybe CmmExpr) passed to do_checks is usually
602 603 604 605 606
--     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.

607
do_checks :: Maybe CmmExpr    -- Should we check the stack?
608 609
                              -- See Note [Stack usage]
          -> Bool             -- Should we check for preemption?
610
          -> Maybe CmmExpr    -- Heap headroom (bytes)
611
          -> CmmAGraph        -- What to do on failure
612
          -> FCode ()
613
do_checks mb_stk_hwm checkYield mb_alloc_lit do_gc = do
614
  dflags <- getDynFlags
615 616
  gc_id <- newLabelC

617
  let
618 619 620
    Just alloc_lit = mb_alloc_lit

    bump_hp   = cmmOffsetExprB dflags (CmmReg hpReg) alloc_lit
621

622 623 624
    -- Sp overflow if ((old + 0) - CmmHighStack < SpLim)
    -- At the beginning of a function old + 0 = Sp
    -- See Note [Single stack check]
625 626
    sp_oflo sp_hwm =
         CmmMachOp (mo_wordULt dflags)
627
                  [CmmMachOp (MO_Sub (typeWidth (cmmRegType dflags spReg)))
628
                             [CmmStackSlot Old 0, sp_hwm],
629 630 631 632 633 634
                   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)
635
                  [CmmReg hpReg, CmmReg (CmmGlobal HpLim)]
636

637
    alloc_n = mkAssign (CmmGlobal HpAlloc) alloc_lit
638

639 640
  case mb_stk_hwm of
    Nothing -> return ()
641
    Just stk_hwm -> tickyStackCheck >> (emit =<< mkCmmIfGoto (sp_oflo stk_hwm) gc_id)
642

643 644 645 646 647 648 649 650 651
  -- 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 ()

652
  if (isJust mb_alloc_lit)
653
    then do
654
     tickyHeapCheck
655 656
     emitAssign hpReg bump_hp
     emit =<< mkCmmIfThen hp_oflo (alloc_n <*> mkBranch gc_id)
657
    else do
658
      when (checkYield && not (gopt Opt_OmitYields dflags)) $ do
659 660 661 662 663
         -- Yielding if HpLim == 0
         let yielding = CmmMachOp (mo_wordEq dflags)
                                  [CmmReg (CmmGlobal HpLim),
                                   CmmLit (zeroCLit dflags)]
         emit =<< mkCmmIfGoto yielding gc_id
664

Peter Wortmann's avatar
Peter Wortmann committed
665 666 667
  tscope <- getTickScope
  emitOutOfLine gc_id
   (do_gc, tscope) -- this is expected to jump back somewhere
668

669 670 671 672 673 674
                -- 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.
675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698

-- 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.