PrimOps.cmm 64.5 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
stg_newArrayzh ( W_ n /* words */, gcptr init )
141
{
142 143
    W_ words, size;
    gcptr p, arr;
144

145
    again: MAYBE_GC(again);
146

147 148 149 150 151
    // 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;
152
    ("ptr" arr) = ccall allocate(MyCapability() "ptr",words);
153 154
    TICK_ALLOC_PRIM(SIZEOF_StgMutArrPtrs, WDS(n), 0);

155
    SET_HDR(arr, stg_MUT_ARR_PTRS_DIRTY_info, CCCS);
156
    StgMutArrPtrs_ptrs(arr) = n;
157
    StgMutArrPtrs_size(arr) = size;
158 159 160 161 162

    // Initialise all elements of the the array with the value in R2
    p = arr + SIZEOF_StgMutArrPtrs;
  for:
    if (p < arr + WDS(words)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
163 164 165
        W_[p] = init;
        p = p + WDS(1);
        goto for;
166
    }
167 168 169
    // Initialise the mark bits with 0
  for2:
    if (p < arr + WDS(size)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
170 171 172
        W_[p] = 0;
        p = p + WDS(1);
        goto for2;
173
    }
174

175
    return (arr);
176 177
}

178
stg_unsafeThawArrayzh ( gcptr arr )
179 180 181
{
  // SUBTLETY TO DO WITH THE OLD GEN MUTABLE LIST
  //
ian@well-typed.com's avatar
ian@well-typed.com committed
182
  // A MUT_ARR_PTRS lives on the mutable list, but a MUT_ARR_PTRS_FROZEN
183 184
  // 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
185
  // the mutable list is not easy).
ian@well-typed.com's avatar
ian@well-typed.com committed
186
  //
187
  // So that we can tell whether a MUT_ARR_PTRS_FROZEN is on the mutable list,
188 189 190
  // 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.
  //
191 192
  // 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
193
  // the closure type is MUT_ARR_PTRS_FROZEN0 if it is on the mutable list,
194 195 196
  // 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.
197
  //
198 199 200
  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
201 202
        // must be done after SET_INFO, because it ASSERTs closure_MUTABLE()
        return (arr);
203
  } else {
ian@well-typed.com's avatar
ian@well-typed.com committed
204 205
        SET_INFO(arr,stg_MUT_ARR_PTRS_DIRTY_info);
        return (arr);
206 207 208
  }
}

209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225
stg_casArrayzh
/* MutableArray# s a -> Int# -> a -> a -> State# s -> (# State# s, Int#, a #) */
{
    W_ arr, p, ind, old, new, h, len;
    arr = R1; // anything else?
    ind = R2;
    old = R3;
    new = R4;

    p = arr + SIZEOF_StgMutArrPtrs + WDS(ind);
    (h) = foreign "C" cas(p, old, new) [];
    
    if (h != old) {
        // Failure, return what was there instead of 'old':
        RET_NP(1,h);
    } else {
        // Compare and Swap Succeeded:
226 227 228 229
	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;
230 231 232 233
        RET_NP(0,h);
    }
}

234
stg_newArrayArrayzh ( W_ n /* words */ )
235
{
236 237
    W_ words, size;
    gcptr p, arr;
238

239
    MAYBE_GC_N(stg_newArrayArrayzh, n);
240 241 242 243 244 245

    // 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;
246
    ("ptr" arr) = ccall allocate(MyCapability() "ptr",words);
247 248 249 250 251 252 253 254 255 256
    TICK_ALLOC_PRIM(SIZEOF_StgMutArrPtrs, WDS(n), 0);

    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:
    if (p < arr + WDS(words)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
257 258 259
        W_[p] = arr;
        p = p + WDS(1);
        goto for;
260 261 262 263
    }
    // Initialise the mark bits with 0
  for2:
    if (p < arr + WDS(size)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
264 265 266
        W_[p] = 0;
        p = p + WDS(1);
        goto for2;
267 268
    }

269
    return (arr);
270 271
}

pumpkin's avatar
pumpkin committed
272

273 274 275 276
/* -----------------------------------------------------------------------------
   MutVar primitives
   -------------------------------------------------------------------------- */

277
stg_newMutVarzh ( gcptr init )
278 279 280
{
    W_ mv;

281
    ALLOC_PRIM_P (SIZEOF_StgMutVar, stg_newMutVarzh, init);
282 283

    mv = Hp - SIZEOF_StgMutVar + WDS(1);
284
    SET_HDR(mv,stg_MUT_VAR_DIRTY_info,CCCS);
285
    StgMutVar_var(mv) = init;
ian@well-typed.com's avatar
ian@well-typed.com committed
286

287
    return (mv);
288 289
}

290
stg_casMutVarzh ( gcptr mv, gcptr old, gcptr new )
Simon Marlow's avatar
Simon Marlow committed
291 292
 /* MutVar# s a -> a -> a -> State# s -> (# State#, Int#, a #) */
{
293
    gcptr h;
Simon Marlow's avatar
Simon Marlow committed
294

295 296
    (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var,
                          old, new);
Simon Marlow's avatar
Simon Marlow committed
297
    if (h != old) {
298
        return (1,h);
Simon Marlow's avatar
Simon Marlow committed
299
    } else {
300
        if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
301
           ccall dirty_MUT_VAR(BaseReg "ptr", mv "ptr");
302
        }
303
        return (0,h);
Simon Marlow's avatar
Simon Marlow committed
304 305 306
    }
}

307
stg_atomicModifyMutVarzh ( gcptr mv, gcptr f )
308
{
309
    W_ z, x, y, r, h;
310

ian@well-typed.com's avatar
ian@well-typed.com committed
311
    /* If x is the current contents of the MutVar#, then
312 313 314
       We want to make the new contents point to

         (sel_0 (f x))
ian@well-typed.com's avatar
ian@well-typed.com committed
315

316
       and the return value is
ian@well-typed.com's avatar
ian@well-typed.com committed
317 318

         (sel_1 (f x))
319 320 321 322

        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
323
         y = [stg_sel_0 z]   (max (HS + 1) MIN_UPD_SIZE)
324 325 326 327
         r = [stg_sel_1 z]   (max (HS + 1) MIN_UPD_SIZE)
    */

#if MIN_UPD_SIZE > 1
328
#define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
329 330
#define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),WDS(MIN_UPD_SIZE-1))
#else
331
#define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(1))
332 333 334 335
#define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),0)
#endif

#if MIN_UPD_SIZE > 2
336
#define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
337 338
#define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),WDS(MIN_UPD_SIZE-2))
#else
339
#define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(2))
340 341 342 343 344
#define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),0)
#endif

#define SIZE (THUNK_2_SIZE + THUNK_1_SIZE + THUNK_1_SIZE)

345
   HP_CHK_GEN_TICKY(SIZE);
346 347 348 349

   TICK_ALLOC_THUNK_2();
   CCCS_ALLOC(THUNK_2_SIZE);
   z = Hp - THUNK_2_SIZE + WDS(1);
350
   SET_HDR(z, stg_ap_2_upd_info, CCCS);
351
   LDV_RECORD_CREATE(z);
352
   StgThunk_payload(z,0) = f;
353 354 355 356

   TICK_ALLOC_THUNK_1();
   CCCS_ALLOC(THUNK_1_SIZE);
   y = z - THUNK_1_SIZE;
357
   SET_HDR(y, stg_sel_0_upd_info, CCCS);
358
   LDV_RECORD_CREATE(y);
359
   StgThunk_payload(y,0) = z;
360 361 362 363

   TICK_ALLOC_THUNK_1();
   CCCS_ALLOC(THUNK_1_SIZE);
   r = y - THUNK_1_SIZE;
364
   SET_HDR(r, stg_sel_1_upd_info, CCCS);
365
   LDV_RECORD_CREATE(r);
366 367
   StgThunk_payload(r,0) = z;

368 369 370 371
 retry:
   x = StgMutVar_var(mv);
   StgThunk_payload(z,1) = x;
#ifdef THREADED_RTS
372
   (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var, x, y);
373 374 375
   if (h != x) { goto retry; }
#else
   StgMutVar_var(mv) = y;
376
#endif
377

378
   if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
379
     ccall dirty_MUT_VAR(BaseReg "ptr", mv "ptr");
380 381
   }

382
   return (r);
383 384 385 386 387 388 389 390
}

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

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

391 392 393
stg_mkWeakzh ( gcptr key,
               gcptr value,
               gcptr finalizer /* or stg_NO_FINALIZER_closure */ )
394
{
395
  gcptr w;
396

397
  ALLOC_PRIM (SIZEOF_StgWeak)
398 399

  w = Hp - SIZEOF_StgWeak + WDS(1);
400
  SET_HDR(w, stg_WEAK_info, CCCS);
401

402 403 404 405
  StgWeak_key(w)         = key;
  StgWeak_value(w)       = value;
  StgWeak_finalizer(w)   = finalizer;
  StgWeak_cfinalizers(w) = stg_NO_FINALIZER_closure;
406

407
  ACQUIRE_LOCK(sm_mutex);
408 409
  StgWeak_link(w) = generation_weak_ptr_list(W_[g0]);
  generation_weak_ptr_list(W_[g0]) = w;
410
  RELEASE_LOCK(sm_mutex);
411

412
  IF_DEBUG(weak, ccall debugBelch(stg_weak_msg,w));
413

414
  return (w);
415 416
}

417
stg_mkWeakNoFinalizzerzh ( gcptr key, gcptr value )
418
{
419
  jump stg_mkWeakzh (key, value, stg_NO_FINALIZER_closure);
420 421
}

422 423 424 425 426 427 428
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 )
429
{
430
  W_ c, info;
431

432
  LOCK_CLOSURE(w, info);
433 434 435 436 437 438 439

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

440
  ALLOC_PRIM (SIZEOF_StgCFinalizerList)
441

442 443
  c = Hp - SIZEOF_StgCFinalizerList + WDS(1);
  SET_HDR(c, stg_C_FINALIZER_LIST_info, CCCS);
444

445 446 447 448
  StgCFinalizerList_fptr(c) = fptr;
  StgCFinalizerList_ptr(c) = ptr;
  StgCFinalizerList_eptr(c) = eptr;
  StgCFinalizerList_flag(c) = flag;
449

450 451
  StgCFinalizerList_link(c) = StgWeak_cfinalizers(w);
  StgWeak_cfinalizers(w) = c;
452

453
  unlockClosure(w, info);
454

455
  recordMutable(w);
456

457
  IF_DEBUG(weak, ccall debugBelch(stg_cfinalizer_msg,w));
458

459
  return (1);
460
}
461

462
stg_finalizzeWeakzh ( gcptr w )
463
{
464 465 466
  gcptr f, list;
  W_ info;

467
  LOCK_CLOSURE(w, info);
468 469

  // already dead?
470 471
  if (info == stg_DEAD_WEAK_info) {
      unlockClosure(w, info);
472
      return (0,stg_NO_FINALIZER_closure);
473 474
  }

475 476 477
  f    = StgWeak_finalizer(w);
  list = StgWeak_cfinalizers(w);

478 479 480 481 482 483 484 485
  // 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
486
  // Furthermore, when PROFILING is turned on, dead weak pointers are exactly as
487 488 489 490 491 492 493
  // 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()?
  //
494
  unlockClosure(w, stg_DEAD_WEAK_info);
495

496
  LDV_RECORD_CREATE(w);
497

498 499
  if (list != stg_NO_FINALIZER_closure) {
    ccall runCFinalizers(list);
500 501
  }

502 503
  /* return the finalizer */
  if (f == stg_NO_FINALIZER_closure) {
504
      return (0,stg_NO_FINALIZER_closure);
505
  } else {
506
      return (1,f);
507 508 509
  }
}

510
stg_deRefWeakzh ( gcptr w )
511
{
512
  W_ code, info;
513
  gcptr val;
514

515 516 517 518 519 520 521
  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.

522
    LOCK_CLOSURE(w, info);
523 524 525 526
    unlockClosure(w, info);
  }

  if (info == stg_WEAK_info) {
527 528 529 530 531 532
    code = 1;
    val = StgWeak_value(w);
  } else {
    code = 0;
    val = w;
  }
533
  return (code,val);
534 535 536
}

/* -----------------------------------------------------------------------------
537
   Floating point operations.
538 539
   -------------------------------------------------------------------------- */

540
stg_decodeFloatzuIntzh ( F_ arg )
ian@well-typed.com's avatar
ian@well-typed.com committed
541
{
542
    W_ p;
543 544 545
    W_ mp_tmp1;
    W_ mp_tmp_w;

546
    STK_CHK_GEN_N (WDS(2));
547 548 549

    mp_tmp1  = Sp - WDS(1);
    mp_tmp_w = Sp - WDS(2);
ian@well-typed.com's avatar
ian@well-typed.com committed
550

551
    /* Perform the operation */
552
    ccall __decodeFloat_Int(mp_tmp1 "ptr", mp_tmp_w "ptr", arg);
ian@well-typed.com's avatar
ian@well-typed.com committed
553

554
    /* returns: (Int# (mantissa), Int# (exponent)) */
555
    return (W_[mp_tmp1], W_[mp_tmp_w]);
556 557
}

558
stg_decodeDoublezu2Intzh ( D_ arg )
ian@well-typed.com's avatar
ian@well-typed.com committed
559
{
560
    W_ p;
561 562 563 564 565
    W_ mp_tmp1;
    W_ mp_tmp2;
    W_ mp_result1;
    W_ mp_result2;

566
    STK_CHK_GEN_N (WDS(4));
567 568 569 570 571

    mp_tmp1    = Sp - WDS(1);
    mp_tmp2    = Sp - WDS(2);
    mp_result1 = Sp - WDS(3);
    mp_result2 = Sp - WDS(4);
572 573

    /* Perform the operation */
574
    ccall __decodeDouble_2Int(mp_tmp1 "ptr", mp_tmp2 "ptr",
575
                                    mp_result1 "ptr", mp_result2 "ptr",
576
                                    arg);
577 578 579

    /* returns:
       (Int# (mant sign), Word# (mant high), Word# (mant low), Int# (expn)) */
580
    return (W_[mp_tmp1], W_[mp_tmp2], W_[mp_result1], W_[mp_result2]);
581 582
}

583 584 585 586
/* -----------------------------------------------------------------------------
 * Concurrency primitives
 * -------------------------------------------------------------------------- */

587
stg_forkzh ( gcptr closure )
588
{
589
  MAYBE_GC_P(stg_forkzh, closure);
590

591
  gcptr threadid;
592

ian@well-typed.com's avatar
ian@well-typed.com committed
593 594
  ("ptr" threadid) = ccall createIOThread( MyCapability() "ptr",
                                RtsFlags_GcFlags_initialStkSize(RtsFlags),
595
                                closure "ptr");
596 597

  /* start blocked if the current thread is blocked */
598
  StgTSO_flags(threadid) = %lobits16(
ian@well-typed.com's avatar
ian@well-typed.com committed
599
     TO_W_(StgTSO_flags(threadid)) |
600
     TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
601

602
  ccall scheduleThread(MyCapability() "ptr", threadid "ptr");
603

604 605
  // context switch soon, but not immediately: we don't want every
  // forkIO to force a context-switch.
606
  Capability_context_switch(MyCapability()) = 1 :: CInt;
ian@well-typed.com's avatar
ian@well-typed.com committed
607

608
  return (threadid);
609 610
}

611
stg_forkOnzh ( W_ cpu, gcptr closure )
612
{
613
again: MAYBE_GC(again);
614

615
  gcptr threadid;
616

ian@well-typed.com's avatar
ian@well-typed.com committed
617 618
  ("ptr" threadid) = ccall createIOThread( MyCapability() "ptr",
                                RtsFlags_GcFlags_initialStkSize(RtsFlags),
619
                                closure "ptr");
620 621

  /* start blocked if the current thread is blocked */
622
  StgTSO_flags(threadid) = %lobits16(
ian@well-typed.com's avatar
ian@well-typed.com committed
623
     TO_W_(StgTSO_flags(threadid)) |
624
     TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
625

626
  ccall scheduleThreadOn(MyCapability() "ptr", cpu, threadid "ptr");
627

628 629
  // context switch soon, but not immediately: we don't want every
  // forkIO to force a context-switch.
630
  Capability_context_switch(MyCapability()) = 1 :: CInt;
ian@well-typed.com's avatar
ian@well-typed.com committed
631

632
  return (threadid);
633 634
}

635
stg_yieldzh ()
636
{
637 638 639 640 641
  // 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;
642
  jump stg_yield_noregs();
643 644
}

645
stg_myThreadIdzh ()
646
{
647
  return (CurrentTSO);
648 649
}

650
stg_labelThreadzh ( gcptr threadid, W_ addr )
651
{
652
#if defined(DEBUG) || defined(TRACING) || defined(DTRACE)
653
  ccall labelThread(MyCapability() "ptr", threadid "ptr", addr "ptr");
654
#endif
655
  return ();
656 657
}

658
stg_isCurrentThreadBoundzh (/* no args */)
659 660
{
  W_ r;
661 662
  (r) = ccall isThreadBound(CurrentTSO);
  return (r);
663 664
}

665
stg_threadStatuszh ( gcptr tso )
666 667 668
{
    W_ why_blocked;
    W_ what_next;
669
    W_ ret, cap, locked;
670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686

    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;
        }
    }
687 688 689 690 691 692 693 694 695

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

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

696
    return (ret,cap,locked);
697
}
698 699 700 701 702

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

703 704 705
// Catch retry frame -----------------------------------------------------------

#define CATCH_RETRY_FRAME_FIELDS(w_,p_,info_ptr,        \
Simon Marlow's avatar
Simon Marlow committed
706
                                 p1, p2,                \
707 708 709 710
                                 running_alt_code,      \
                                 first_code,            \
                                 alt_code)              \
  w_ info_ptr,                                          \
Simon Marlow's avatar
Simon Marlow committed
711
  PROF_HDR_FIELDS(w_,p1,p2)                             \
712 713 714
  w_ running_alt_code,                                  \
  p_ first_code,                                        \
  p_ alt_code
715 716


717
INFO_TABLE_RET(stg_catch_retry_frame, CATCH_RETRY_FRAME,
718
               CATCH_RETRY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
719
                                        info_ptr, p1, p2,
720 721 722 723
                                        running_alt_code,
                                        first_code,
                                        alt_code))
    return (P_ ret)
724
{
725 726
    W_ r;
    gcptr trec, outer, arg;
727

728 729 730 731 732 733 734 735 736 737 738 739 740 741 742
    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
743
                (CATCH_RETRY_FRAME_FIELDS(,,info_ptr, p1, p2,
744 745 746 747 748 749
                                          running_alt_code,
                                          first_code,
                                          alt_code))
                (alt_code);
        } else {
            jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
750
                (CATCH_RETRY_FRAME_FIELDS(,,info_ptr, p1, p2,
751 752 753 754 755 756 757
                                          running_alt_code,
                                          first_code,
                                          alt_code))
                (first_code);
        }
    }
}
758

Simon Marlow's avatar
Simon Marlow committed
759
// Atomically frame ------------------------------------------------------------
760

761
// This must match StgAtomicallyFrame in Closures.h
Simon Marlow's avatar
Simon Marlow committed
762
#define ATOMICALLY_FRAME_FIELDS(w_,p_,info_ptr,p1,p2,code,next,result)  \
763
  w_ info_ptr,                                                          \
Simon Marlow's avatar
Simon Marlow committed
764
  PROF_HDR_FIELDS(w_,p1,p2)                                             \
765 766 767 768 769
  p_ code,                                                              \
  p_ next,                                                              \
  p_ result


770
INFO_TABLE_RET(stg_atomically_frame, ATOMICALLY_FRAME,
771 772
               // layout of the frame, and bind the field names
               ATOMICALLY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
773
                                       info_ptr, p1, p2,
774 775 776 777
                                       code,
                                       next_invariant,
                                       frame_result))
    return (P_ result) // value returned to the frame
778
{
779 780
  W_ valid;
  gcptr trec, outer, next_invariant, q;
781

782
  trec   = StgTSO_trec(CurrentTSO);
783
  outer  = StgTRecHeader_enclosing_trec(trec);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
784 785 786

  if (outer == NO_TREC) {
    /* First time back at the atomically frame -- pick up invariants */
787 788 789
    ("ptr" next_invariant) =
        ccall stmGetInvariantsToCheck(MyCapability() "ptr", trec "ptr");
    frame_result = result;
790 791

  } else {
tharris@microsoft.com's avatar
tharris@microsoft.com committed
792 793 794
    /* 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;
795 796
    StgInvariantCheckQueue_my_execution(next_invariant) = trec;
    ccall stmAbortTransaction(MyCapability() "ptr", trec "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
797 798
    /* Don't free trec -- it's linked from q and will be stashed in the
     * invariant if we eventually commit. */
799 800
    next_invariant =
       StgInvariantCheckQueue_next_queue_entry(next_invariant);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
801 802 803
    trec = outer;
  }

804
  if (next_invariant != END_INVARIANT_CHECK_QUEUE) {
tharris@microsoft.com's avatar
tharris@microsoft.com committed
805
    /* We can't commit yet: another invariant to check */
806
    ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", trec "ptr");
807
    StgTSO_trec(CurrentTSO) = trec;
808 809
    q = StgInvariantCheckQueue_invariant(next_invariant);
    jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
810 811
        (ATOMICALLY_FRAME_FIELDS(,,info_ptr,p1,p2,
                                 code,next_invariant,frame_result))
812
        (StgAtomicInvariant_code(q));
tharris@microsoft.com's avatar
tharris@microsoft.com committed
813 814 815 816

  } else {

    /* We've got no more invariants to check, try to commit */
817
    (valid) = ccall stmCommitTransaction(MyCapability() "ptr", trec "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
818 819 820
    if (valid != 0) {
      /* Transaction was valid: commit succeeded */
      StgTSO_trec(CurrentTSO) = NO_TREC;
821
      return (frame_result);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
822 823
    } else {
      /* Transaction was not valid: try again */
824
      ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", NO_TREC "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
825
      StgTSO_trec(CurrentTSO) = trec;
826 827 828 829 830
      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
831 832
          (ATOMICALLY_FRAME_FIELDS(,,info_ptr,p1,p2,
                                   code,next_invariant,frame_result))
833
          (code);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
834
    }
835 836 837
  }
}

838

839
INFO_TABLE_RET(stg_atomically_waiting_frame, ATOMICALLY_FRAME,
840 841
               // layout of the frame, and bind the field names
               ATOMICALLY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
842
                                       info_ptr, p1, p2,
843 844 845 846
                                       code,
                                       next_invariant,
                                       frame_result))
    return (/* no return values */)
847
{
848
  W_ trec, valid;
849 850

  /* The TSO is currently waiting: should we stop waiting? */
851
  (valid) = ccall stmReWait(MyCapability() "ptr", CurrentTSO "ptr");
Simon Marlow's avatar
Simon Marlow committed
852
  if (valid != 0) {
853 854
      /* Previous attempt is still valid: no point trying again yet */
      jump stg_block_noregs
Simon Marlow's avatar
Simon Marlow committed
855
          (ATOMICALLY_FRAME_FIELDS(,,info_ptr, p1, p2,
856 857
                                   code,next_invariant,frame_result))
          ();
858 859
  } else {
    /* Previous attempt is no longer valid: try again */
860
    ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", NO_TREC "ptr");
861
    StgTSO_trec(CurrentTSO) = trec;
862 863 864

    // change the frame header to stg_atomically_frame_info
    jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
865
        (ATOMICALLY_FRAME_FIELDS(,,stg_atomically_frame_info, p1, p2,
866 867
                                 code,next_invariant,frame_result))
        (code);
868
  }
869
}
870

871
// STM catch frame -------------------------------------------------------------
872 873 874 875 876 877

/* Catch frames are very similar to update frames, but when entering
 * one we just pop the frame off the stack and perform the correct
 * kind of return to the activation record underneath us on the stack.