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
import Prelude hiding ((<*>))

53
import Control.Monad (when)
54
import Data.Maybe (isJust)
55

56
-----------------------------------------------------------
57
--              Initialise dynamic heap objects
58 59 60
-----------------------------------------------------------

allocDynClosure
61 62
        :: Maybe Id
        -> CmmInfoTable
Simon Marlow's avatar
Simon Marlow committed
63
        -> LambdaFormInfo
64 65 66 67 68 69 70
        -> 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
71
        -> FCode CmmExpr -- returns Hp+n
72

73
allocDynClosureCmm
74
        :: Maybe Id -> CmmInfoTable -> LambdaFormInfo -> CmmExpr -> CmmExpr
75
        -> [(CmmExpr, ByteOff)]
76 77
        -> FCode CmmExpr -- returns Hp+n

78
-- allocDynClosure allocates the thing in the heap,
79
-- and modifies the virtual Hp to account for this.
80 81 82
-- 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.
83

84 85 86 87 88 89 90 91 92 93
-- 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.
94 95


96 97 98 99 100
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)
101 102


103 104 105 106 107 108
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
109 110


111 112 113 114 115
-- | Low-level heap object allocation.
allocHeapClosure
  :: SMRep                            -- ^ representation of the object
  -> CmmExpr                          -- ^ info pointer
  -> CmmExpr                          -- ^ cost centre
116
  -> [(CmmExpr,ByteOff)]              -- ^ payload
117 118
  -> FCode CmmExpr                    -- ^ returns the address of the object
allocHeapClosure rep info_ptr use_cc payload = do
119 120
  profDynAlloc rep use_cc

121
  virt_hp <- getVirtHp
122

123 124 125 126 127 128
  -- 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.
129

130
  base <- getHpRelOffset info_offset
131
  emitComment $ mkFastString "allocHeapClosure"
132 133 134 135 136 137 138 139 140 141
  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
142

143 144

emitSetDynHdr :: CmmExpr -> CmmExpr -> CmmExpr -> FCode ()
145
emitSetDynHdr base info_ptr ccs
146
  = do dflags <- getDynFlags
147
       hpStore base (zip (header dflags) [0, wORD_SIZE dflags ..])
148
  where
149 150
    header :: DynFlags -> [CmmExpr]
    header dflags = [info_ptr] ++ dynProfHdr dflags ccs
Jan Stolarek's avatar
Jan Stolarek committed
151
        -- ToDof: Parallel stuff
152
        -- No ticky header
153 154

-- Store the item (expr,off) in base[off]
155
hpStore :: CmmExpr -> [(CmmExpr, ByteOff)] -> FCode ()
156 157 158
hpStore base vals = do
  dflags <- getDynFlags
  sequence_ $
159
    [ emitStore (cmmOffsetB dflags base off) val | (val,off) <- vals ]
160 161

-----------------------------------------------------------
162
--              Layout of static closures
163 164 165 166 167
-----------------------------------------------------------

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

168
mkStaticClosureFields
169 170
        :: DynFlags
        -> CmmInfoTable
171
        -> CostCentreStack
172
        -> CafInfo
173 174
        -> [CmmLit]             -- Payload
        -> [CmmLit]             -- The full closure
175 176
mkStaticClosureFields dflags info_tbl ccs caf_refs payload
  = mkStaticClosure dflags info_lbl ccs payload padding
177
        static_link_field saved_info_field
178
  where
Simon Marlow's avatar
Simon Marlow committed
179
    info_lbl = cit_lbl info_tbl
180 181 182 183 184 185 186 187 188

    -- 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
189 190 191
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:
192

Simon Marlow's avatar
Simon Marlow committed
193
    is_caf = isThunkRep (cit_rep info_tbl)
194

195
    padding
196 197
        | is_caf && null payload = [mkIntCLit dflags 0]
        | otherwise = []
198 199

    static_link_field
200
        | is_caf || staticClosureNeedsLink (mayHaveCafRefs caf_refs) info_tbl
Simon Marlow's avatar
Simon Marlow committed
201 202 203
        = [static_link_value]
        | otherwise
        = []
204 205

    saved_info_field
206
        | is_caf     = [mkIntCLit dflags 0]
207
        | otherwise  = []
208

209
        -- For a static constructor which has NoCafRefs, we set the
210 211
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
212
    static_link_value
213
        | mayHaveCafRefs caf_refs  = mkIntCLit dflags 0
214 215 216
        | otherwise                = mkIntCLit dflags 3  -- No CAF refs
                                      -- See Note [STATIC_LINK fields]
                                      -- in rts/sm/Storage.h
217

218
mkStaticClosure :: DynFlags -> CLabel -> CostCentreStack -> [CmmLit]
219
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
220
mkStaticClosure dflags info_lbl ccs payload padding static_link_field saved_info_field
221
  =  [CmmLabel info_lbl]
Jan Stolarek's avatar
Jan Stolarek committed
222
  ++ staticProfHdr dflags ccs
223
  ++ concatMap (padLitToWord dflags) payload
224
  ++ padding
225 226 227
  ++ static_link_field
  ++ saved_info_field

228 229
-- 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?
230 231 232
padLitToWord :: DynFlags -> CmmLit -> [CmmLit]
padLitToWord dflags lit = lit : padding pad_length
  where width = typeWidth (cmmLitType dflags lit)
233
        pad_length = wORD_SIZE dflags - widthInBytes width :: Int
234 235 236 237 238 239 240

        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)

241
-----------------------------------------------------------
242
--              Heap overflow checking
243 244 245 246 247 248 249 250 251 252 253 254
-----------------------------------------------------------

{- 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
255
        r = gc_1p( r )
256 257
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.
258

259
  * A series of canned entry points like
260
        gcfun_2p( f, x, y )
261 262 263 264 265 266 267 268 269
    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 ) {
270 271 272
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
273 274 275
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
276
     it should do no register shuffling
277 278 279

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
280 281 282
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
283
          L: HpAlloc = 8
284 285
             call gc()  -- Needs an info table
             goto T }
286 287

* altHeapCheck: Immediately following an eval
288 289
  Started as
        case f x y of r { (p,q) -> rhs }
290 291 292
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
293 294 295 296 297 298
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
299

300 301
        L: r = gc_1p( r )
           goto K }
302

303 304 305 306
        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)
307 308 309

   (b) No canned sequence for results of f
       Note second info table
310 311 312 313 314 315
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
316

317 318
        L: call gc()    -- Extra info table here
           goto K
319 320 321

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
322
       case branch *not* following eval,
323 324 325
       or let-no-escape
  Exactly the same as the previous case:

326 327 328 329
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...
330

331 332
        L: call gc()
           goto K
333 334 335 336 337
-}

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

338 339 340 341 342 343
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 ()
344

345
entryHeapCheck cl_info nodeSet arity args code
346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363
  = 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
364 365
  = do dflags <- getDynFlags
       let is_thunk = arity == 0
366

367
           args' = map (CmmReg . CmmLocal) args
368 369 370 371 372 373 374
           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)

375
              Function (slow): call (slow) stg_gc_fun(fun, args)
376 377 378
           -}
           gc_call upd
               | is_thunk
379
                 = mkJump dflags NativeNodeCall stg_gc_enter1 [node] upd
380 381

               | is_fastf
382
                 = mkJump dflags NativeNodeCall stg_gc_fun (node : args') upd
383 384

               | otherwise
385
                 = mkJump dflags Slow stg_gc_fun (node : args') upd
386 387

       updfr_sz <- getUpdFrameOff
388 389 390

       loop_id <- newLabelC
       emitLabel loop_id
391
       heapCheck True True (gc_call updfr_sz <*> mkBranch loop_id) code
392

393 394
-- ------------------------------------------------------------
-- A heap/stack check in a case alternative
395

396 397 398 399 400 401 402 403 404 405 406 407 408 409 410

-- 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

411
altHeapCheck :: [LocalReg] -> FCode a -> FCode a
412 413 414 415
altHeapCheck regs code = altOrNoEscapeHeapCheck False regs code

altOrNoEscapeHeapCheck :: Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck checkYield regs code = do
416 417
    dflags <- getDynFlags
    case cannedGCEntryPoint dflags regs of
418
      Nothing -> genericGC checkYield code
419 420
      Just gc -> do
        lret <- newLabelC
421
        let (off, _, copyin) = copyInOflow dflags NativeReturn (Young lret) regs []
422
        lcont <- newLabelC
Peter Wortmann's avatar
Peter Wortmann committed
423 424
        tscope <- getTickScope
        emitOutOfLine lret (copyin <*> mkBranch lcont, tscope)
425
        emitLabel lcont
426
        cannedGCReturnsTo checkYield False gc regs lret off code
427 428 429

altHeapCheckReturnsTo :: [LocalReg] -> Label -> ByteOff -> FCode a -> FCode a
altHeapCheckReturnsTo regs lret off code
430 431
  = do dflags <- getDynFlags
       case cannedGCEntryPoint dflags regs of
432 433 434 435 436 437 438 439
           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
440

441
cannedGCReturnsTo :: Bool -> Bool -> CmmExpr -> [LocalReg] -> Label -> ByteOff
442 443
                  -> FCode a
                  -> FCode a
444
cannedGCReturnsTo checkYield cont_on_stack gc regs lret off code
445 446
  = do dflags <- getDynFlags
       updfr_sz <- getUpdFrameOff
447
       heapCheck False checkYield (gc_call dflags gc updfr_sz) code
448
  where
449
    reg_exprs = map (CmmReg . CmmLocal) regs
450
      -- Note [stg_gc arguments]
451

452 453 454 455
      -- 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.
456
    gc_call dflags label sp
457 458 459 460
      | cont_on_stack
      = mkJumpReturnsTo dflags label NativeReturn reg_exprs lret off sp
      | otherwise
      = mkCallReturnsTo dflags label NativeReturn reg_exprs lret off sp []
461

462 463
genericGC :: Bool -> FCode a -> FCode a
genericGC checkYield code
464 465 466
  = do updfr_sz <- getUpdFrameOff
       lretry <- newLabelC
       emitLabel lretry
467
       call <- mkCall generic_gc (GC, GC) [] [] updfr_sz []
468
       heapCheck False checkYield (call <*> mkBranch lretry) code
469

470 471
cannedGCEntryPoint :: DynFlags -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint dflags regs
472
  = case map localRegType regs of
473
      []  -> Just (mkGcLabel "stg_gc_noregs")
474
      [ty]
475 476 477 478 479
          | 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
480

481 482 483
          | width == wordWidth dflags -> Just (mkGcLabel "stg_gc_unbx_r1")
          | width == W64              -> Just (mkGcLabel "stg_gc_l1")
          | otherwise                 -> Nothing
484 485
          where
              width = typeWidth ty
486 487 488 489 490 491 492 493 494 495 496 497
      [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")
498
      _otherwise -> Nothing
499

500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516
-- Note [stg_gc arguments]
-- It might seem that we could avoid passing the arguments to the
-- stg_gc function, because they are already in the right registers.
-- While this is usually the case, it isn't always.  Sometimes the
-- code generator has cleverly avoided the eval in a case, e.g. in
-- ffi/should_run/4221.hs we found
--
--   case a_r1mb of z
--     FunPtr x y -> ...
--
-- where a_r1mb is bound a top-level constructor, and is known to be
-- evaluated.  The codegen just assigns x, y and z, and continues;
-- R1 is never assigned.
--
-- So we'll have to rely on optimisations to eliminatethese
-- assignments where possible.

517

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

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

-------------------------------
527 528
heapCheck :: Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck checkStack checkYield do_gc code
529
  = getHeapUsage $ \ hpHw ->
530 531
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
532 533
    do  { dflags <- getDynFlags
        ; let mb_alloc_bytes
534 535 536 537 538 539 540
                 | 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."]
541 542
                 | hpHw > 0  = Just (mkIntExpr dflags (hpHw * (wORD_SIZE dflags)))
                 | otherwise = Nothing
543
                 where mBLOCK_SIZE = bLOCKS_PER_MBLOCK dflags * bLOCK_SIZE_W dflags
544 545 546
              stk_hwm | checkStack = Just (CmmLit CmmHighStackMark)
                      | otherwise  = Nothing
        ; codeOnly $ do_checks stk_hwm checkYield mb_alloc_bytes do_gc
nfrisby's avatar
nfrisby committed
547
        ; tickyAllocHeap True hpHw
548 549
        ; setRealHp hpHw
        ; code }
550

551 552 553 554 555 556
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 []
557
       do_checks stk_hwm False mb_bytes (call <*> mkBranch lretry)
558

559 560
-- Note [Single stack check]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~
561 562 563
-- 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.
564
--
565 566 567 568 569
-- 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.
570 571 572 573 574

-- 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.
575
--
576 577 578 579 580 581 582 583 584 585 586 587
-- 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.
588

589 590 591 592 593 594
-- 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
595
-- to stand for the number of bytes needed. When the stack is made
596 597 598
-- manifest, the number of bytes needed is calculated, and used to
-- replace occurrences of CmmHighStackMark
--
599
-- The (Maybe CmmExpr) passed to do_checks is usually
600 601 602 603 604
--     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.

605
do_checks :: Maybe CmmExpr    -- Should we check the stack?
606 607
                              -- See Note [Stack usage]
          -> Bool             -- Should we check for preemption?
608
          -> Maybe CmmExpr    -- Heap headroom (bytes)
609
          -> CmmAGraph        -- What to do on failure
610
          -> FCode ()
611
do_checks mb_stk_hwm checkYield mb_alloc_lit do_gc = do
612
  dflags <- getDynFlags
613 614
  gc_id <- newLabelC

615
  let
616 617 618
    Just alloc_lit = mb_alloc_lit

    bump_hp   = cmmOffsetExprB dflags (CmmReg hpReg) alloc_lit
619

620 621 622
    -- Sp overflow if ((old + 0) - CmmHighStack < SpLim)
    -- At the beginning of a function old + 0 = Sp
    -- See Note [Single stack check]
623 624
    sp_oflo sp_hwm =
         CmmMachOp (mo_wordULt dflags)
625
                  [CmmMachOp (MO_Sub (typeWidth (cmmRegType dflags spReg)))
626
                             [CmmStackSlot Old 0, sp_hwm],
627 628 629 630 631 632
                   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)
633
                  [CmmReg hpReg, CmmReg (CmmGlobal HpLim)]
634

635
    alloc_n = mkAssign (CmmGlobal HpAlloc) alloc_lit
636

637 638
  case mb_stk_hwm of
    Nothing -> return ()
639 640
    Just stk_hwm -> tickyStackCheck
      >> (emit =<< mkCmmIfGoto' (sp_oflo stk_hwm) gc_id (Just False) )
641

642 643 644 645 646 647 648 649 650
  -- 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 ()

651
  if (isJust mb_alloc_lit)
652
    then do
653
     tickyHeapCheck
654
     emitAssign hpReg bump_hp
655
     emit =<< mkCmmIfThen' hp_oflo (alloc_n <*> mkBranch gc_id) (Just False)
656
    else do
657
      when (checkYield && not (gopt Opt_OmitYields dflags)) $ do
658 659 660 661
         -- Yielding if HpLim == 0
         let yielding = CmmMachOp (mo_wordEq dflags)
                                  [CmmReg (CmmGlobal HpLim),
                                   CmmLit (zeroCLit dflags)]
662
         emit =<< mkCmmIfGoto' yielding gc_id (Just False)
663

Peter Wortmann's avatar
Peter Wortmann committed
664 665 666
  tscope <- getTickScope
  emitOutOfLine gc_id
   (do_gc, tscope) -- this is expected to jump back somewhere
667

668 669 670 671 672 673
                -- 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.
674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697

-- 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.