PrimOps.cmm 66 KB
Newer Older
1
/* -*- tab-width: 8 -*- */
2 3
/* -----------------------------------------------------------------------------
 *
4
 * (c) The GHC Team, 1998-2012
5 6 7 8 9 10 11 12 13
 *
 * Out-of-line primitive operations
 *
 * This file contains the implementations of all the primitive
 * operations ("primops") which are not expanded inline.  See
 * ghc/compiler/prelude/primops.txt.pp for a list of all the primops;
 * this file contains code for most of those with the attribute
 * out_of_line=True.
 *
14 15 16
 * Entry convention: the entry convention for a primop is the
 * NativeNodeCall convention, and the return convention is
 * NativeReturn.  (see compiler/cmm/CmmCallConv.hs)
17 18 19 20 21 22 23 24 25
 *
 * This file is written in a subset of C--, extended with various
 * features specific to GHC.  It is compiled by GHC directly.  For the
 * syntax of .cmm files, see the parser in ghc/compiler/cmm/CmmParse.y.
 *
 * ---------------------------------------------------------------------------*/

#include "Cmm.h"

26
#ifdef __PIC__
27 28
import pthread_mutex_lock;
import pthread_mutex_unlock;
29
#endif
30
import base_ControlziExceptionziBase_nestedAtomically_closure;
31 32
import EnterCriticalSection;
import LeaveCriticalSection;
Ian Lynagh's avatar
Ian Lynagh committed
33
import ghczmprim_GHCziTypes_False_closure;
34
#if defined(USE_MINIINTERPRETER) || !defined(mingw32_HOST_OS)
35
import sm_mutex;
36
#endif
37

38 39 40 41 42 43 44 45 46 47 48 49 50
/*-----------------------------------------------------------------------------
  Array Primitives

  Basically just new*Array - the others are all inline macros.

  The slow entry point is for returning from a heap check, the saved
  size argument must be re-loaded from the stack.
  -------------------------------------------------------------------------- */

/* for objects that are *less* than the size of a word, make sure we
 * round up to the nearest word for the size of the array.
 */

51
stg_newByteArrayzh ( W_ n )
52
{
53 54 55 56 57
    W_ words, payload_words;
    gcptr p;

    MAYBE_GC_N(stg_newByteArrayzh, n);

58 59
    payload_words = ROUNDUP_BYTES_TO_WDS(n);
    words = BYTES_TO_WDS(SIZEOF_StgArrWords) + payload_words;
60
    ("ptr" p) = ccall allocate(MyCapability() "ptr",words);
61
    TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);
62
    SET_HDR(p, stg_ARR_WORDS_info, CCCS);
63
    StgArrWords_bytes(p) = n;
64
    return (p);
65 66
}

Simon Marlow's avatar
Simon Marlow committed
67 68 69
#define BA_ALIGN 16
#define BA_MASK  (BA_ALIGN-1)

70
stg_newPinnedByteArrayzh ( W_ n )
71
{
72 73 74 75
    W_ words, bytes, payload_words;
    gcptr p;

    MAYBE_GC_N(stg_newPinnedByteArrayzh, n);
76

77
    bytes = n;
78 79 80 81 82 83 84 85 86 87
    /* payload_words is what we will tell the profiler we had to allocate */
    payload_words = ROUNDUP_BYTES_TO_WDS(bytes);
    /* When we actually allocate memory, we need to allow space for the
       header: */
    bytes = bytes + SIZEOF_StgArrWords;
    /* And we want to align to BA_ALIGN bytes, so we need to allow space
       to shift up to BA_ALIGN - 1 bytes: */
    bytes = bytes + BA_ALIGN - 1;
    /* Now we convert to a number of words: */
    words = ROUNDUP_BYTES_TO_WDS(bytes);
Simon Marlow's avatar
Simon Marlow committed
88

89
    ("ptr" p) = ccall allocatePinned(MyCapability() "ptr", words);
Simon Marlow's avatar
Simon Marlow committed
90 91
    TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);

92 93
    /* Now we need to move p forward so that the payload is aligned
       to BA_ALIGN bytes: */
Simon Marlow's avatar
Simon Marlow committed
94 95
    p = p + ((-p - SIZEOF_StgArrWords) & BA_MASK);

96
    SET_HDR(p, stg_ARR_WORDS_info, CCCS);
97
    StgArrWords_bytes(p) = n;
98
    return (p);
Simon Marlow's avatar
Simon Marlow committed
99 100
}

101
stg_newAlignedPinnedByteArrayzh ( W_ n, W_ alignment )
Simon Marlow's avatar
Simon Marlow committed
102
{
103 104
    W_ words, bytes, payload_words;
    gcptr p;
Simon Marlow's avatar
Simon Marlow committed
105

106
    again: MAYBE_GC(again);
Simon Marlow's avatar
Simon Marlow committed
107

108 109 110 111 112
    /* we always supply at least word-aligned memory, so there's no
       need to allow extra space for alignment if the requirement is less
       than a word.  This also prevents mischief with alignment == 0. */
    if (alignment <= SIZEOF_W) { alignment = 1; }

113 114
    bytes = n;

115 116
    /* payload_words is what we will tell the profiler we had to allocate */
    payload_words = ROUNDUP_BYTES_TO_WDS(bytes);
Simon Marlow's avatar
Simon Marlow committed
117

118 119 120 121 122 123 124 125
    /* When we actually allocate memory, we need to allow space for the
       header: */
    bytes = bytes + SIZEOF_StgArrWords;
    /* And we want to align to <alignment> bytes, so we need to allow space
       to shift up to <alignment - 1> bytes: */
    bytes = bytes + alignment - 1;
    /* Now we convert to a number of words: */
    words = ROUNDUP_BYTES_TO_WDS(bytes);
126

127
    ("ptr" p) = ccall allocatePinned(MyCapability() "ptr", words);
128 129
    TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);

130 131 132 133
    /* Now we need to move p forward so that the payload is aligned
       to <alignment> bytes. Note that we are assuming that
       <alignment> is a power of 2, which is technically not guaranteed */
    p = p + ((-p - SIZEOF_StgArrWords) & (alignment - 1));
134

135
    SET_HDR(p, stg_ARR_WORDS_info, CCCS);
136
    StgArrWords_bytes(p) = n;
137
    return (p);
138 139
}

140 141 142 143 144
// RRN: This one does not use the "ticketing" approach because it
// deals in unboxed scalars, not heap pointers.
stg_casIntArrayzh( gcptr arr, W_ ind, W_ old, W_ new )
/* MutableByteArray# s -> Int# -> Int# -> Int# -> State# s -> (# State# s, Int# #) */
{
tibbe's avatar
tibbe committed
145
    W_ p, h;
146 147 148 149 150 151 152

    p = arr + SIZEOF_StgArrWords + WDS(ind);
    (h) = ccall cas(p, old, new);

    return(h);
}

153 154 155 156

stg_fetchAddIntArrayzh( gcptr arr, W_ ind, W_ incr )
/* MutableByteArray# s -> Int# -> Int# -> State# s -> (# State# s, Int# #) */
{
tibbe's avatar
tibbe committed
157
    W_ p, h;
158 159

    p = arr + SIZEOF_StgArrWords + WDS(ind);
160
    (h) = ccall atomic_inc(p, incr);
161 162 163 164 165

    return(h);
}


166
stg_newArrayzh ( W_ n /* words */, gcptr init )
167
{
tibbe's avatar
tibbe committed
168 169
    W_ words, size, p;
    gcptr arr;
170

171
    again: MAYBE_GC(again);
172

173 174 175 176 177
    // the mark area contains one byte for each 2^MUT_ARR_PTRS_CARD_BITS words
    // in the array, making sure we round up, and then rounding up to a whole
    // number of words.
    size = n + mutArrPtrsCardWords(n);
    words = BYTES_TO_WDS(SIZEOF_StgMutArrPtrs) + size;
178
    ("ptr" arr) = ccall allocate(MyCapability() "ptr",words);
tibbe's avatar
tibbe committed
179
    TICK_ALLOC_PRIM(SIZEOF_StgMutArrPtrs, WDS(size), 0);
180

181
    SET_HDR(arr, stg_MUT_ARR_PTRS_DIRTY_info, CCCS);
182
    StgMutArrPtrs_ptrs(arr) = n;
183
    StgMutArrPtrs_size(arr) = size;
184 185 186 187

    // Initialise all elements of the the array with the value in R2
    p = arr + SIZEOF_StgMutArrPtrs;
  for:
tibbe's avatar
tibbe committed
188
    if (p < arr + SIZEOF_StgMutArrPtrs + WDS(n)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
189 190 191
        W_[p] = init;
        p = p + WDS(1);
        goto for;
192 193
    }

194
    return (arr);
195 196
}

197
stg_unsafeThawArrayzh ( gcptr arr )
198 199 200
{
  // SUBTLETY TO DO WITH THE OLD GEN MUTABLE LIST
  //
ian@well-typed.com's avatar
ian@well-typed.com committed
201
  // A MUT_ARR_PTRS lives on the mutable list, but a MUT_ARR_PTRS_FROZEN
202 203
  // normally doesn't.  However, when we freeze a MUT_ARR_PTRS, we leave
  // it on the mutable list for the GC to remove (removing something from
204
  // the mutable list is not easy).
ian@well-typed.com's avatar
ian@well-typed.com committed
205
  //
206
  // So that we can tell whether a MUT_ARR_PTRS_FROZEN is on the mutable list,
207 208 209
  // when we freeze it we set the info ptr to be MUT_ARR_PTRS_FROZEN0
  // to indicate that it is still on the mutable list.
  //
210 211
  // So, when we thaw a MUT_ARR_PTRS_FROZEN, we must cope with two cases:
  // either it is on a mut_list, or it isn't.  We adopt the convention that
212
  // the closure type is MUT_ARR_PTRS_FROZEN0 if it is on the mutable list,
213 214 215
  // and MUT_ARR_PTRS_FROZEN otherwise.  In fact it wouldn't matter if
  // we put it on the mutable list more than once, but it would get scavenged
  // multiple times during GC, which would be unnecessarily slow.
216
  //
217 218 219
  if (StgHeader_info(arr) != stg_MUT_ARR_PTRS_FROZEN0_info) {
        SET_INFO(arr,stg_MUT_ARR_PTRS_DIRTY_info);
        recordMutable(arr);
ian@well-typed.com's avatar
ian@well-typed.com committed
220 221
        // must be done after SET_INFO, because it ASSERTs closure_MUTABLE()
        return (arr);
222
  } else {
ian@well-typed.com's avatar
ian@well-typed.com committed
223 224
        SET_INFO(arr,stg_MUT_ARR_PTRS_DIRTY_info);
        return (arr);
225 226 227
  }
}

228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
stg_cloneArrayzh ( gcptr src, W_ offset, W_ n )
{
  cloneArray(stg_MUT_ARR_PTRS_FROZEN_info, src, offset, n)
}

stg_cloneMutableArrayzh ( gcptr src, W_ offset, W_ n )
{
  cloneArray(stg_MUT_ARR_PTRS_DIRTY_info, src, offset, n)
}

// We have to escape the "z" in the name.
stg_freezzeArrayzh ( gcptr src, W_ offset, W_ n )
{
  cloneArray(stg_MUT_ARR_PTRS_FROZEN_info, src, offset, n)
}

stg_thawArrayzh ( gcptr src, W_ offset, W_ n )
{
  cloneArray(stg_MUT_ARR_PTRS_DIRTY_info, src, offset, n)
}

249
// RRN: Uses the ticketed approach; see casMutVar
250
stg_casArrayzh ( gcptr arr, W_ ind, gcptr old, gcptr new )
251
/* MutableArray# s a -> Int# -> a -> a -> State# s -> (# State# s, Int#, Any a #) */
252
{
tibbe's avatar
tibbe committed
253 254
    gcptr h;
    W_ p, len;
255 256

    p = arr + SIZEOF_StgMutArrPtrs + WDS(ind);
257
    (h) = ccall cas(p, old, new);
258 259 260
    
    if (h != old) {
        // Failure, return what was there instead of 'old':
261
        return (1,h);
262 263
    } else {
        // Compare and Swap Succeeded:
rrnewton's avatar
rrnewton committed
264 265 266 267
        SET_HDR(arr, stg_MUT_ARR_PTRS_DIRTY_info, CCCS);
        len = StgMutArrPtrs_ptrs(arr);
        // The write barrier.  We must write a byte into the mark table:
        I8[arr + SIZEOF_StgMutArrPtrs + WDS(len) + (ind >> MUT_ARR_PTRS_CARD_BITS )] = 1;
268
        return (0,new);
269 270 271
    }
}

272
stg_newArrayArrayzh ( W_ n /* words */ )
273
{
tibbe's avatar
tibbe committed
274 275
    W_ words, size, p;
    gcptr arr;
276

277
    MAYBE_GC_N(stg_newArrayArrayzh, n);
278 279 280 281 282 283

    // the mark area contains one byte for each 2^MUT_ARR_PTRS_CARD_BITS words
    // in the array, making sure we round up, and then rounding up to a whole
    // number of words.
    size = n + mutArrPtrsCardWords(n);
    words = BYTES_TO_WDS(SIZEOF_StgMutArrPtrs) + size;
284
    ("ptr" arr) = ccall allocate(MyCapability() "ptr",words);
tibbe's avatar
tibbe committed
285
    TICK_ALLOC_PRIM(SIZEOF_StgMutArrPtrs, WDS(size), 0);
286 287 288 289 290 291 292 293

    SET_HDR(arr, stg_MUT_ARR_PTRS_DIRTY_info, W_[CCCS]);
    StgMutArrPtrs_ptrs(arr) = n;
    StgMutArrPtrs_size(arr) = size;

    // Initialise all elements of the array with a pointer to the new array
    p = arr + SIZEOF_StgMutArrPtrs;
  for:
tibbe's avatar
tibbe committed
294
    if (p < arr + SIZEOF_StgMutArrPtrs + WDS(n)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
295 296 297
        W_[p] = arr;
        p = p + WDS(1);
        goto for;
298 299
    }

300
    return (arr);
301 302
}

pumpkin's avatar
pumpkin committed
303

304 305 306 307
/* -----------------------------------------------------------------------------
   MutVar primitives
   -------------------------------------------------------------------------- */

308
stg_newMutVarzh ( gcptr init )
309 310 311
{
    W_ mv;

312
    ALLOC_PRIM_P (SIZEOF_StgMutVar, stg_newMutVarzh, init);
313 314

    mv = Hp - SIZEOF_StgMutVar + WDS(1);
315
    SET_HDR(mv,stg_MUT_VAR_DIRTY_info,CCCS);
316
    StgMutVar_var(mv) = init;
ian@well-typed.com's avatar
ian@well-typed.com committed
317

318
    return (mv);
319 320
}

321 322 323 324 325
// RRN: To support the "ticketed" approach, we return the NEW rather
// than old value if the CAS is successful.  This is received in an
// opaque form in the Haskell code, preventing the compiler from
// changing its pointer identity.  The ticket can then be safely used
// in future CAS operations.
326
stg_casMutVarzh ( gcptr mv, gcptr old, gcptr new )
327
 /* MutVar# s a -> a -> a -> State# s -> (# State#, Int#, Any a #) */
Simon Marlow's avatar
Simon Marlow committed
328
{
329
    gcptr h;
Simon Marlow's avatar
Simon Marlow committed
330

331 332
    (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var,
                          old, new);
Simon Marlow's avatar
Simon Marlow committed
333
    if (h != old) {
334
        return (1,h);
Simon Marlow's avatar
Simon Marlow committed
335
    } else {
336
        if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
337
           ccall dirty_MUT_VAR(BaseReg "ptr", mv "ptr");
338
        }
339
        return (0,new);
Simon Marlow's avatar
Simon Marlow committed
340 341 342
    }
}

343
stg_atomicModifyMutVarzh ( gcptr mv, gcptr f )
344
{
345
    W_ z, x, y, r, h;
346

ian@well-typed.com's avatar
ian@well-typed.com committed
347
    /* If x is the current contents of the MutVar#, then
348 349 350
       We want to make the new contents point to

         (sel_0 (f x))
ian@well-typed.com's avatar
ian@well-typed.com committed
351

352
       and the return value is
ian@well-typed.com's avatar
ian@well-typed.com committed
353 354

         (sel_1 (f x))
355 356 357 358

        obviously we can share (f x).

         z = [stg_ap_2 f x]  (max (HS + 2) MIN_UPD_SIZE)
ian@well-typed.com's avatar
ian@well-typed.com committed
359
         y = [stg_sel_0 z]   (max (HS + 1) MIN_UPD_SIZE)
360 361 362 363
         r = [stg_sel_1 z]   (max (HS + 1) MIN_UPD_SIZE)
    */

#if MIN_UPD_SIZE > 1
364
#define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
365 366
#define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),WDS(MIN_UPD_SIZE-1))
#else
367
#define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(1))
368 369 370 371
#define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),0)
#endif

#if MIN_UPD_SIZE > 2
372
#define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
373 374
#define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),WDS(MIN_UPD_SIZE-2))
#else
375
#define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(2))
376 377 378 379 380
#define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),0)
#endif

#define SIZE (THUNK_2_SIZE + THUNK_1_SIZE + THUNK_1_SIZE)

381
   HP_CHK_GEN_TICKY(SIZE);
382 383 384 385

   TICK_ALLOC_THUNK_2();
   CCCS_ALLOC(THUNK_2_SIZE);
   z = Hp - THUNK_2_SIZE + WDS(1);
386
   SET_HDR(z, stg_ap_2_upd_info, CCCS);
387
   LDV_RECORD_CREATE(z);
388
   StgThunk_payload(z,0) = f;
389 390 391 392

   TICK_ALLOC_THUNK_1();
   CCCS_ALLOC(THUNK_1_SIZE);
   y = z - THUNK_1_SIZE;
393
   SET_HDR(y, stg_sel_0_upd_info, CCCS);
394
   LDV_RECORD_CREATE(y);
395
   StgThunk_payload(y,0) = z;
396 397 398 399

   TICK_ALLOC_THUNK_1();
   CCCS_ALLOC(THUNK_1_SIZE);
   r = y - THUNK_1_SIZE;
400
   SET_HDR(r, stg_sel_1_upd_info, CCCS);
401
   LDV_RECORD_CREATE(r);
402 403
   StgThunk_payload(r,0) = z;

404 405 406 407
 retry:
   x = StgMutVar_var(mv);
   StgThunk_payload(z,1) = x;
#ifdef THREADED_RTS
408
   (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var, x, y);
409 410 411
   if (h != x) { goto retry; }
#else
   StgMutVar_var(mv) = y;
412
#endif
413

414
   if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
415
     ccall dirty_MUT_VAR(BaseReg "ptr", mv "ptr");
416 417
   }

418
   return (r);
419 420 421 422 423 424 425 426
}

/* -----------------------------------------------------------------------------
   Weak Pointer Primitives
   -------------------------------------------------------------------------- */

STRING(stg_weak_msg,"New weak pointer at %p\n")

427 428 429
stg_mkWeakzh ( gcptr key,
               gcptr value,
               gcptr finalizer /* or stg_NO_FINALIZER_closure */ )
430
{
431
  gcptr w;
432

433
  ALLOC_PRIM (SIZEOF_StgWeak)
434 435

  w = Hp - SIZEOF_StgWeak + WDS(1);
436
  SET_HDR(w, stg_WEAK_info, CCCS);
437

438 439 440 441
  StgWeak_key(w)         = key;
  StgWeak_value(w)       = value;
  StgWeak_finalizer(w)   = finalizer;
  StgWeak_cfinalizers(w) = stg_NO_FINALIZER_closure;
442

443
  ACQUIRE_LOCK(sm_mutex);
444 445
  StgWeak_link(w) = generation_weak_ptr_list(W_[g0]);
  generation_weak_ptr_list(W_[g0]) = w;
446
  RELEASE_LOCK(sm_mutex);
447

448
  IF_DEBUG(weak, ccall debugBelch(stg_weak_msg,w));
449

450
  return (w);
451 452
}

453
stg_mkWeakNoFinalizzerzh ( gcptr key, gcptr value )
454
{
455
  jump stg_mkWeakzh (key, value, stg_NO_FINALIZER_closure);
456 457
}

458 459 460 461 462 463 464
STRING(stg_cfinalizer_msg,"Adding a finalizer to %p\n")

stg_addCFinalizzerToWeakzh ( W_ fptr,   // finalizer
                             W_ ptr,
                             W_ flag,   // has environment (0 or 1)
                             W_ eptr,
                             gcptr w )
465
{
466
  W_ c, info;
467

468
  ALLOC_PRIM (SIZEOF_StgCFinalizerList)
469

470 471
  c = Hp - SIZEOF_StgCFinalizerList + WDS(1);
  SET_HDR(c, stg_C_FINALIZER_LIST_info, CCCS);
472

473 474 475 476
  StgCFinalizerList_fptr(c) = fptr;
  StgCFinalizerList_ptr(c) = ptr;
  StgCFinalizerList_eptr(c) = eptr;
  StgCFinalizerList_flag(c) = flag;
477

478 479 480 481 482 483 484 485
  LOCK_CLOSURE(w, info);

  if (info == stg_DEAD_WEAK_info) {
    // Already dead.
    unlockClosure(w, info);
    return (0);
  }

486 487
  StgCFinalizerList_link(c) = StgWeak_cfinalizers(w);
  StgWeak_cfinalizers(w) = c;
488

489
  unlockClosure(w, info);
490

491
  recordMutable(w);
492

493
  IF_DEBUG(weak, ccall debugBelch(stg_cfinalizer_msg,w));
494

495
  return (1);
496
}
497

498
stg_finalizzeWeakzh ( gcptr w )
499
{
500 501 502
  gcptr f, list;
  W_ info;

503
  LOCK_CLOSURE(w, info);
504 505

  // already dead?
506 507
  if (info == stg_DEAD_WEAK_info) {
      unlockClosure(w, info);
508
      return (0,stg_NO_FINALIZER_closure);
509 510
  }

511 512 513
  f    = StgWeak_finalizer(w);
  list = StgWeak_cfinalizers(w);

514 515 516 517 518 519 520 521
  // kill it
#ifdef PROFILING
  // @LDV profiling
  // A weak pointer is inherently used, so we do not need to call
  // LDV_recordDead_FILL_SLOP_DYNAMIC():
  //    LDV_recordDead_FILL_SLOP_DYNAMIC((StgClosure *)w);
  // or, LDV_recordDead():
  //    LDV_recordDead((StgClosure *)w, sizeofW(StgWeak) - sizeofW(StgProfHeader));
ian@well-typed.com's avatar
ian@well-typed.com committed
522
  // Furthermore, when PROFILING is turned on, dead weak pointers are exactly as
523 524 525 526 527 528 529
  // large as weak pointers, so there is no need to fill the slop, either.
  // See stg_DEAD_WEAK_info in StgMiscClosures.hc.
#endif

  //
  // Todo: maybe use SET_HDR() and remove LDV_recordCreate()?
  //
530
  unlockClosure(w, stg_DEAD_WEAK_info);
531

532
  LDV_RECORD_CREATE(w);
533

534 535
  if (list != stg_NO_FINALIZER_closure) {
    ccall runCFinalizers(list);
536 537
  }

538 539
  /* return the finalizer */
  if (f == stg_NO_FINALIZER_closure) {
540
      return (0,stg_NO_FINALIZER_closure);
541
  } else {
542
      return (1,f);
543 544 545
  }
}

546
stg_deRefWeakzh ( gcptr w )
547
{
548
  W_ code, info;
549
  gcptr val;
550

551 552 553 554 555 556 557
  info = GET_INFO(w);

  if (info == stg_WHITEHOLE_info) {
    // w is locked by another thread. Now it's not immediately clear if w is
    // alive or not. We use lockClosure to wait for the info pointer to become
    // something other than stg_WHITEHOLE_info.

558
    LOCK_CLOSURE(w, info);
559 560 561 562
    unlockClosure(w, info);
  }

  if (info == stg_WEAK_info) {
563 564 565 566 567 568
    code = 1;
    val = StgWeak_value(w);
  } else {
    code = 0;
    val = w;
  }
569
  return (code,val);
570 571 572
}

/* -----------------------------------------------------------------------------
573
   Floating point operations.
574 575
   -------------------------------------------------------------------------- */

576
stg_decodeFloatzuIntzh ( F_ arg )
ian@well-typed.com's avatar
ian@well-typed.com committed
577
{
578
    W_ p;
579
    W_ tmp, mp_tmp1, mp_tmp_w, r1, r2;
580

581
    STK_CHK_GEN_N (WDS(2));
582

583
    reserve 2 = tmp {
ian@well-typed.com's avatar
ian@well-typed.com committed
584

585 586 587 588 589 590 591 592 593
      mp_tmp1  = tmp + WDS(1);
      mp_tmp_w = tmp;

      /* Perform the operation */
      ccall __decodeFloat_Int(mp_tmp1 "ptr", mp_tmp_w "ptr", arg);

      r1 = W_[mp_tmp1];
      r2 = W_[mp_tmp_w];
    }
ian@well-typed.com's avatar
ian@well-typed.com committed
594

595
    /* returns: (Int# (mantissa), Int# (exponent)) */
596
    return (r1, r2);
597 598
}

599
stg_decodeDoublezu2Intzh ( D_ arg )
ian@well-typed.com's avatar
ian@well-typed.com committed
600
{
601 602 603
    W_ p, tmp;
    W_ mp_tmp1, mp_tmp2, mp_result1, mp_result2;
    W_ r1, r2, r3, r4;
604

605
    STK_CHK_GEN_N (WDS(4));
606

607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623
    reserve 4 = tmp {

      mp_tmp1    = tmp + WDS(3);
      mp_tmp2    = tmp + WDS(2);
      mp_result1 = tmp + WDS(1);
      mp_result2 = tmp;
  
      /* Perform the operation */
      ccall __decodeDouble_2Int(mp_tmp1 "ptr", mp_tmp2 "ptr",
                                      mp_result1 "ptr", mp_result2 "ptr",
                                      arg);

      r1 = W_[mp_tmp1];
      r2 = W_[mp_tmp2];
      r3 = W_[mp_result1];
      r4 = W_[mp_result2];
    }
624 625 626

    /* returns:
       (Int# (mant sign), Word# (mant high), Word# (mant low), Int# (expn)) */
627
    return (r1, r2, r3, r4);
628 629
}

630 631 632 633
/* -----------------------------------------------------------------------------
 * Concurrency primitives
 * -------------------------------------------------------------------------- */

634
stg_forkzh ( gcptr closure )
635
{
636
  MAYBE_GC_P(stg_forkzh, closure);
637

638
  gcptr threadid;
639

ian@well-typed.com's avatar
ian@well-typed.com committed
640 641
  ("ptr" threadid) = ccall createIOThread( MyCapability() "ptr",
                                RtsFlags_GcFlags_initialStkSize(RtsFlags),
642
                                closure "ptr");
643 644

  /* start blocked if the current thread is blocked */
645
  StgTSO_flags(threadid) = %lobits16(
ian@well-typed.com's avatar
ian@well-typed.com committed
646
     TO_W_(StgTSO_flags(threadid)) |
647
     TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
648

649
  ccall scheduleThread(MyCapability() "ptr", threadid "ptr");
650

651 652
  // context switch soon, but not immediately: we don't want every
  // forkIO to force a context-switch.
653
  Capability_context_switch(MyCapability()) = 1 :: CInt;
ian@well-typed.com's avatar
ian@well-typed.com committed
654

655
  return (threadid);
656 657
}

658
stg_forkOnzh ( W_ cpu, gcptr closure )
659
{
660
again: MAYBE_GC(again);
661

662
  gcptr threadid;
663

ian@well-typed.com's avatar
ian@well-typed.com committed
664 665
  ("ptr" threadid) = ccall createIOThread( MyCapability() "ptr",
                                RtsFlags_GcFlags_initialStkSize(RtsFlags),
666
                                closure "ptr");
667 668

  /* start blocked if the current thread is blocked */
669
  StgTSO_flags(threadid) = %lobits16(
ian@well-typed.com's avatar
ian@well-typed.com committed
670
     TO_W_(StgTSO_flags(threadid)) |
671
     TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
672

673
  ccall scheduleThreadOn(MyCapability() "ptr", cpu, threadid "ptr");
674

675 676
  // context switch soon, but not immediately: we don't want every
  // forkIO to force a context-switch.
677
  Capability_context_switch(MyCapability()) = 1 :: CInt;
ian@well-typed.com's avatar
ian@well-typed.com committed
678

679
  return (threadid);
680 681
}

682
stg_yieldzh ()
683
{
684 685 686 687 688
  // when we yield to the scheduler, we have to tell it to put the
  // current thread to the back of the queue by setting the
  // context_switch flag.  If we don't do this, it will run the same
  // thread again.
  Capability_context_switch(MyCapability()) = 1 :: CInt;
689
  jump stg_yield_noregs();
690 691
}

692
stg_myThreadIdzh ()
693
{
694
  return (CurrentTSO);
695 696
}

697
stg_labelThreadzh ( gcptr threadid, W_ addr )
698
{
699
#if defined(DEBUG) || defined(TRACING) || defined(DTRACE)
700
  ccall labelThread(MyCapability() "ptr", threadid "ptr", addr "ptr");
701
#endif
702
  return ();
703 704
}

705
stg_isCurrentThreadBoundzh (/* no args */)
706 707
{
  W_ r;
708 709
  (r) = ccall isThreadBound(CurrentTSO);
  return (r);
710 711
}

712
stg_threadStatuszh ( gcptr tso )
713 714 715
{
    W_ why_blocked;
    W_ what_next;
716
    W_ ret, cap, locked;
717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733

    what_next   = TO_W_(StgTSO_what_next(tso));
    why_blocked = TO_W_(StgTSO_why_blocked(tso));
    // Note: these two reads are not atomic, so they might end up
    // being inconsistent.  It doesn't matter, since we
    // only return one or the other.  If we wanted to return the
    // contents of block_info too, then we'd have to do some synchronisation.

    if (what_next == ThreadComplete) {
        ret = 16;  // NB. magic, matches up with GHC.Conc.threadStatus
    } else {
        if (what_next == ThreadKilled) {
            ret = 17;
        } else {
            ret = why_blocked;
        }
    }
734 735 736 737 738 739 740 741 742

    cap = TO_W_(Capability_no(StgTSO_cap(tso)));

    if ((TO_W_(StgTSO_flags(tso)) & TSO_LOCKED) != 0) {
        locked = 1;
    } else {
        locked = 0;
    }

743
    return (ret,cap,locked);
744
}
745 746 747 748 749

/* -----------------------------------------------------------------------------
 * TVar primitives
 * -------------------------------------------------------------------------- */

750 751 752
// Catch retry frame -----------------------------------------------------------

#define CATCH_RETRY_FRAME_FIELDS(w_,p_,info_ptr,        \
Simon Marlow's avatar
Simon Marlow committed
753
                                 p1, p2,                \
754 755 756 757
                                 running_alt_code,      \
                                 first_code,            \
                                 alt_code)              \
  w_ info_ptr,                                          \
Simon Marlow's avatar
Simon Marlow committed
758
  PROF_HDR_FIELDS(w_,p1,p2)                             \
759 760 761
  w_ running_alt_code,                                  \
  p_ first_code,                                        \
  p_ alt_code
762 763


764
INFO_TABLE_RET(stg_catch_retry_frame, CATCH_RETRY_FRAME,
765
               CATCH_RETRY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
766
                                        info_ptr, p1, p2,
767 768 769 770
                                        running_alt_code,
                                        first_code,
                                        alt_code))
    return (P_ ret)
771
{
772 773
    W_ r;
    gcptr trec, outer, arg;
774

775 776 777 778 779 780 781 782 783 784 785 786 787 788 789
    trec = StgTSO_trec(CurrentTSO);
    outer  = StgTRecHeader_enclosing_trec(trec);
    (r) = ccall stmCommitNestedTransaction(MyCapability() "ptr", trec "ptr");
    if (r != 0) {
        // Succeeded (either first branch or second branch)
        StgTSO_trec(CurrentTSO) = outer;
        return (ret);
    } else {
        // Did not commit: re-execute
        P_ new_trec;
        ("ptr" new_trec) = ccall stmStartTransaction(MyCapability() "ptr",
                                                           outer "ptr");
        StgTSO_trec(CurrentTSO) = new_trec;
        if (running_alt_code != 0) {
            jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
790
                (CATCH_RETRY_FRAME_FIELDS(,,info_ptr, p1, p2,
791 792 793 794 795 796
                                          running_alt_code,
                                          first_code,
                                          alt_code))
                (alt_code);
        } else {
            jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
797
                (CATCH_RETRY_FRAME_FIELDS(,,info_ptr, p1, p2,
798 799 800 801 802 803 804
                                          running_alt_code,
                                          first_code,
                                          alt_code))
                (first_code);
        }
    }
}
805

Simon Marlow's avatar
Simon Marlow committed
806
// Atomically frame ------------------------------------------------------------
807

808
// This must match StgAtomicallyFrame in Closures.h
Simon Marlow's avatar
Simon Marlow committed
809
#define ATOMICALLY_FRAME_FIELDS(w_,p_,info_ptr,p1,p2,code,next,result)  \
810
  w_ info_ptr,                                                          \
Simon Marlow's avatar
Simon Marlow committed
811
  PROF_HDR_FIELDS(w_,p1,p2)                                             \
812 813 814 815 816
  p_ code,                                                              \
  p_ next,                                                              \
  p_ result


817
INFO_TABLE_RET(stg_atomically_frame, ATOMICALLY_FRAME,
818 819
               // layout of the frame, and bind the field names
               ATOMICALLY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
820
                                       info_ptr, p1, p2,
821 822 823 824
                                       code,
                                       next_invariant,
                                       frame_result))
    return (P_ result) // value returned to the frame
825
{
826 827
  W_ valid;
  gcptr trec, outer, next_invariant, q;
828

829
  trec   = StgTSO_trec(CurrentTSO);
830
  outer  = StgTRecHeader_enclosing_trec(trec);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
831 832 833

  if (outer == NO_TREC) {
    /* First time back at the atomically frame -- pick up invariants */
834 835 836
    ("ptr" next_invariant) =
        ccall stmGetInvariantsToCheck(MyCapability() "ptr", trec "ptr");
    frame_result = result;
837 838

  } else {
tharris@microsoft.com's avatar
tharris@microsoft.com committed
839 840 841
    /* Second/subsequent time back at the atomically frame -- abort the
     * tx that's checking the invariant and move on to the next one */
    StgTSO_trec(CurrentTSO) = outer;
842 843
    StgInvariantCheckQueue_my_execution(next_invariant) = trec;
    ccall stmAbortTransaction(MyCapability() "ptr", trec "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
844 845
    /* Don't free trec -- it's linked from q and will be stashed in the
     * invariant if we eventually commit. */
846 847
    next_invariant =
       StgInvariantCheckQueue_next_queue_entry(next_invariant);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
848 849 850
    trec = outer;
  }

851
  if (next_invariant != END_INVARIANT_CHECK_QUEUE) {
tharris@microsoft.com's avatar
tharris@microsoft.com committed
852
    /* We can't commit yet: another invariant to check */
853
    ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", trec "ptr");
854
    StgTSO_trec(CurrentTSO) = trec;
855 856
    q = StgInvariantCheckQueue_invariant(next_invariant);
    jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
857 858
        (ATOMICALLY_FRAME_FIELDS(,,info_ptr,p1,p2,
                                 code,next_invariant,frame_result))
859
        (StgAtomicInvariant_code(q));
tharris@microsoft.com's avatar
tharris@microsoft.com committed
860 861 862 863

  } else {

    /* We've got no more invariants to check, try to commit */
864
    (valid) = ccall stmCommitTransaction(MyCapability() "ptr", trec "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
865 866 867
    if (valid != 0) {
      /* Transaction was valid: commit succeeded */
      StgTSO_trec(CurrentTSO) = NO_TREC;
868
      return (frame_result);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
869 870
    } else {
      /* Transaction was not valid: try again */
871
      ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", NO_TREC "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
872
      StgTSO_trec(CurrentTSO) = trec;
873 874 875 876 877
      next_invariant = END_INVARIANT_CHECK_QUEUE;

      jump stg_ap_v_fast
          // push the StgAtomicallyFrame again: the code generator is
          // clever enough to only assign the fields that have changed.
Simon Marlow's avatar
Simon Marlow committed
878 879
          (ATOMICALLY_FRAME_FIELDS(,,info_ptr,p1,p2,
                                   code,next_invariant,frame_result))
880
          (code);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
881
    }
882 883 884
  }
}

885

886
INFO_TABLE_RET(stg_atomically_waiting_frame, ATOMICALLY_FRAME,
887 888
               // layout of the frame, and bind the field names
               ATOMICALLY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
889
                                       info_ptr, p1, p2,