StgCmmHeap.hs 24.7 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

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

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

allocDynClosure
58 59
        :: Maybe Id
        -> CmmInfoTable
Simon Marlow's avatar
Simon Marlow committed
60
        -> LambdaFormInfo
61 62 63 64 65 66 67
        -> CmmExpr              -- Cost Centre to stick in the object
        -> CmmExpr              -- Cost Centre to blame for this alloc
                                -- (usually the same; sometimes "OVERHEAD")

        -> [(NonVoid StgArg, VirtualHpOffset)]  -- Offsets from start of object
                                                -- ie Info ptr has offset zero.
                                                -- No void args in here
68
        -> FCode CmmExpr -- returns Hp+n
69

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

75
-- allocDynClosure allocates the thing in the heap,
76
-- and modifies the virtual Hp to account for this.
77 78 79
-- The second return value is the graph that sets the value of the
-- returned LocalReg, which should point to the closure after executing
-- the graph.
80

81 82 83 84 85 86 87 88 89 90
-- allocDynClosure returns an (Hp+8) CmmExpr, and hence the result is
-- only valid until Hp is changed.  The caller should assign the
-- result to a LocalReg if it is required to remain live.
--
-- The reason we don't assign it to a LocalReg here is that the caller
-- is often about to call regIdInfo, which immediately assigns the
-- result of allocDynClosure to a new temp in order to add the tag.
-- So by not generating a LocalReg here we avoid a common source of
-- new temporaries and save some compile time.  This can be quite
-- significant - see test T4801.
91 92


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


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


108 109 110 111 112 113 114 115
-- | 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
116 117
  profDynAlloc rep use_cc

118
  virt_hp <- getVirtHp
119

120 121 122 123 124 125
  -- Find the offset of the info-ptr word
  let info_offset = virt_hp + 1
            -- info_offset is the VirtualHpOffset of the first
            -- word of the new object
            -- Remember, virtHp points to last allocated word,
            -- ie 1 *before* the info-ptr word of new object.
126

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

  -- Fill in the fields
  hpStore base payload

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

  return base
139

140 141

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

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

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

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

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

    -- CAFs must have consistent layout, regardless of whether they
    -- are actually updatable or not.  The layout of a CAF is:
    --
    --        3 saved_info
    --        2 static_link
    --        1 indirectee
    --        0 info ptr
    --
Simon Marlow's avatar
Simon Marlow committed
186 187 188
    -- the static_link and saved_info fields must always be in the
    -- same place.  So we use isThunkRep rather than closureUpdReqd
    -- here:
189

Simon Marlow's avatar
Simon Marlow committed
190
    is_caf = isThunkRep (cit_rep info_tbl)
191

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

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

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

206
        -- For a static constructor which has NoCafRefs, we set the
207 208
        -- static link field to a non-zero value so the garbage
        -- collector will ignore it.
209
    static_link_value
210 211
        | mayHaveCafRefs caf_refs  = mkIntCLit dflags 0
        | otherwise                = mkIntCLit dflags 1  -- No CAF refs
212 213


214
mkStaticClosure :: DynFlags -> CLabel -> CostCentreStack -> [CmmLit]
215
  -> [CmmLit] -> [CmmLit] -> [CmmLit] -> [CmmLit]
216
mkStaticClosure dflags info_lbl ccs payload padding static_link_field saved_info_field
217
  =  [CmmLabel info_lbl]
Jan Stolarek's avatar
Jan Stolarek committed
218
  ++ staticProfHdr dflags ccs
219
  ++ concatMap (padLitToWord dflags) payload
220
  ++ padding
221 222 223
  ++ static_link_field
  ++ saved_info_field

224 225
-- 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?
226 227 228
padLitToWord :: DynFlags -> CmmLit -> [CmmLit]
padLitToWord dflags lit = lit : padding pad_length
  where width = typeWidth (cmmLitType dflags lit)
229
        pad_length = wORD_SIZE dflags - widthInBytes width :: Int
230 231 232 233 234 235 236

        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)

237
-----------------------------------------------------------
238
--              Heap overflow checking
239 240 241 242 243 244 245 246 247 248 249 250
-----------------------------------------------------------

{- 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
251
        r = gc_1p( r )
252 253
    where r is a pointer.  This performs gc, and
    then returns its argument r to its caller.
254

255
  * A series of canned entry points like
256
        gcfun_2p( f, x, y )
257 258 259 260 261 262 263 264 265
    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 ) {
266 267 268
             Hp = Hp+8
             if Hp > HpLim goto L
             ...
269 270 271
          L: HpAlloc = 8
             jump gcfun_2p( f_clo, x, y ) }
     Note the tail call to the garbage collector;
272
     it should do no register shuffling
273 274 275

    (b) No canned sequence
        f( f_clo, x:ptr, y:ptr, ...etc... ) {
276 277 278
          T: Hp = Hp+8
             if Hp > HpLim goto L
             ...
279
          L: HpAlloc = 8
280 281
             call gc()  -- Needs an info table
             goto T }
282 283

* altHeapCheck: Immediately following an eval
284 285
  Started as
        case f x y of r { (p,q) -> rhs }
286 287 288
  (a) With a canned sequence for the results of f
       (which is the very common case since
       all boxed cases return just one pointer
289 290 291 292 293 294
           ...
           r = f( x, y )
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
295

296 297
        L: r = gc_1p( r )
           goto K }
298

299 300 301 302
        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)
303 304 305

   (b) No canned sequence for results of f
       Note second info table
306 307 308 309 310 311
           ...
           (r1,r2,r3) = call f( x, y )
        K:
           Hp = Hp+8
           if Hp > HpLim goto L
           ...code for rhs...
312

313 314
        L: call gc()    -- Extra info table here
           goto K
315 316 317

* generalHeapCheck: Anywhere else
  e.g. entry to thunk
318
       case branch *not* following eval,
319 320 321
       or let-no-escape
  Exactly the same as the previous case:

322 323 324 325
        K:      -- K needs an info table
           Hp = Hp+8
           if Hp > HpLim goto L
           ...
326

327 328
        L: call gc()
           goto K
329 330 331 332 333
-}

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

334 335 336 337 338 339
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 ()
340

341
entryHeapCheck cl_info nodeSet arity args code
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359
  = 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
360 361
  = do dflags <- getDynFlags
       let is_thunk = arity == 0
362 363

           args' = map (CmmReg . CmmLocal) args
364 365 366 367 368 369 370
           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)

371
              Function (slow): call (slow) stg_gc_fun(fun, args)
372 373 374
           -}
           gc_call upd
               | is_thunk
375
                 = mkJump dflags NativeNodeCall stg_gc_enter1 [node] upd
376 377

               | is_fastf
378
                 = mkJump dflags NativeNodeCall stg_gc_fun (node : args') upd
379 380

               | otherwise
381
                 = mkJump dflags Slow stg_gc_fun (node : args') upd
382 383

       updfr_sz <- getUpdFrameOff
384 385 386

       loop_id <- newLabelC
       emitLabel loop_id
387
       heapCheck True True (gc_call updfr_sz <*> mkBranch loop_id) code
388

389 390
-- ------------------------------------------------------------
-- A heap/stack check in a case alternative
391

392 393 394 395 396 397 398 399 400 401 402 403 404 405 406

-- 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

407
altHeapCheck :: [LocalReg] -> FCode a -> FCode a
408 409 410 411
altHeapCheck regs code = altOrNoEscapeHeapCheck False regs code

altOrNoEscapeHeapCheck :: Bool -> [LocalReg] -> FCode a -> FCode a
altOrNoEscapeHeapCheck checkYield regs code = do
412 413
    dflags <- getDynFlags
    case cannedGCEntryPoint dflags regs of
414
      Nothing -> genericGC checkYield code
415 416
      Just gc -> do
        lret <- newLabelC
417
        let (off, _, copyin) = copyInOflow dflags NativeReturn (Young lret) regs []
418 419 420
        lcont <- newLabelC
        emitOutOfLine lret (copyin <*> mkBranch lcont)
        emitLabel lcont
421
        cannedGCReturnsTo checkYield False gc regs lret off code
422 423 424

altHeapCheckReturnsTo :: [LocalReg] -> Label -> ByteOff -> FCode a -> FCode a
altHeapCheckReturnsTo regs lret off code
425 426
  = do dflags <- getDynFlags
       case cannedGCEntryPoint dflags regs of
427 428 429 430 431 432 433 434
           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
435

436
cannedGCReturnsTo :: Bool -> Bool -> CmmExpr -> [LocalReg] -> Label -> ByteOff
437 438
                  -> FCode a
                  -> FCode a
439
cannedGCReturnsTo checkYield cont_on_stack gc regs lret off code
440 441
  = do dflags <- getDynFlags
       updfr_sz <- getUpdFrameOff
442
       heapCheck False checkYield (gc_call dflags gc updfr_sz) code
443 444
  where
    reg_exprs = map (CmmReg . CmmLocal) regs
445
      -- Note [stg_gc arguments]
446

447 448 449 450
      -- 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.
451
    gc_call dflags label sp
452 453 454 455
      | cont_on_stack
      = mkJumpReturnsTo dflags label NativeReturn reg_exprs lret off sp
      | otherwise
      = mkCallReturnsTo dflags label NativeReturn reg_exprs lret off sp []
456

457 458
genericGC :: Bool -> FCode a -> FCode a
genericGC checkYield code
459 460 461
  = do updfr_sz <- getUpdFrameOff
       lretry <- newLabelC
       emitLabel lretry
462
       call <- mkCall generic_gc (GC, GC) [] [] updfr_sz []
463
       heapCheck False checkYield (call <*> mkBranch lretry) code
464

465 466
cannedGCEntryPoint :: DynFlags -> [LocalReg] -> Maybe CmmExpr
cannedGCEntryPoint dflags regs
467
  = case map localRegType regs of
468
      []  -> Just (mkGcLabel "stg_gc_noregs")
469
      [ty]
470 471 472 473 474
          | 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
475

476 477 478
          | width == wordWidth dflags -> Just (mkGcLabel "stg_gc_unbx_r1")
          | width == W64              -> Just (mkGcLabel "stg_gc_l1")
          | otherwise                 -> Nothing
479 480
          where
              width = typeWidth ty
481 482 483 484 485 486 487 488 489 490 491 492
      [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")
493
      _otherwise -> Nothing
494

495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511
-- 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.

512

513 514
-- | The generic GC procedure; no params, no results
generic_gc :: CmmExpr
515
generic_gc = mkGcLabel "stg_gc_noregs"
516 517

-- | Create a CLabel for calling a garbage collector entry point
518 519
mkGcLabel :: String -> CmmExpr
mkGcLabel s = CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit s)))
520 521

-------------------------------
522 523
heapCheck :: Bool -> Bool -> CmmAGraph -> FCode a -> FCode a
heapCheck checkStack checkYield do_gc code
524
  = getHeapUsage $ \ hpHw ->
525 526
    -- Emit heap checks, but be sure to do it lazily so
    -- that the conditionals on hpHw don't cause a black hole
527 528 529 530 531 532 533
    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
534
        ; tickyAllocHeap True hpHw
535 536
        ; setRealHp hpHw
        ; code }
537

538 539 540 541 542 543
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 []
544
       do_checks stk_hwm False mb_bytes (call <*> mkBranch lretry)
545

546 547
-- Note [Single stack check]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~
548 549 550
-- 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.
551
--
552 553 554 555 556
-- 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.
557 558 559 560 561

-- 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.
562
--
563 564 565 566 567 568 569 570 571 572 573 574
-- 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.
575

576 577 578 579 580 581
-- 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
582
-- to stand for the number of bytes needed. When the stack is made
583 584 585
-- manifest, the number of bytes needed is calculated, and used to
-- replace occurrences of CmmHighStackMark
--
586
-- The (Maybe CmmExpr) passed to do_checks is usually
587 588 589 590 591
--     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.

592
do_checks :: Maybe CmmExpr    -- Should we check the stack?
593 594
                              -- See Note [Stack usage]
          -> Bool             -- Should we check for preemption?
595
          -> Maybe CmmExpr    -- Heap headroom (bytes)
596
          -> CmmAGraph        -- What to do on failure
597
          -> FCode ()
598
do_checks mb_stk_hwm checkYield mb_alloc_lit do_gc = do
599
  dflags <- getDynFlags
600 601
  gc_id <- newLabelC

602
  let
603 604 605
    Just alloc_lit = mb_alloc_lit

    bump_hp   = cmmOffsetExprB dflags (CmmReg hpReg) alloc_lit
606

607 608 609
    -- Sp overflow if ((old + 0) - CmmHighStack < SpLim)
    -- At the beginning of a function old + 0 = Sp
    -- See Note [Single stack check]
610 611
    sp_oflo sp_hwm =
         CmmMachOp (mo_wordULt dflags)
612
                  [CmmMachOp (MO_Sub (typeWidth (cmmRegType dflags spReg)))
613
                             [CmmStackSlot Old 0, sp_hwm],
614 615 616 617 618 619
                   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)
620
                  [CmmReg hpReg, CmmReg (CmmGlobal HpLim)]
621

622
    alloc_n = mkAssign (CmmGlobal HpAlloc) alloc_lit
623

624 625
  case mb_stk_hwm of
    Nothing -> return ()
626
    Just stk_hwm -> tickyStackCheck >> (emit =<< mkCmmIfGoto (sp_oflo stk_hwm) gc_id)
627

628 629 630 631 632 633 634 635 636
  -- 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 ()

637
  if (isJust mb_alloc_lit)
638
    then do
639
     tickyHeapCheck
640 641
     emitAssign hpReg bump_hp
     emit =<< mkCmmIfThen hp_oflo (alloc_n <*> mkBranch gc_id)
642
    else do
643
      when (checkYield && not (gopt Opt_OmitYields dflags)) $ do
644 645 646 647 648
         -- Yielding if HpLim == 0
         let yielding = CmmMachOp (mo_wordEq dflags)
                                  [CmmReg (CmmGlobal HpLim),
                                   CmmLit (zeroCLit dflags)]
         emit =<< mkCmmIfGoto yielding gc_id
649 650

  emitOutOfLine gc_id $
651 652
     do_gc -- this is expected to jump back somewhere

653 654 655 656 657 658
                -- 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.
659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682

-- 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.