PrimOps.cmm 65.4 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
}

140 141 142 143 144 145 146 147 148 149 150 151 152
// 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# #) */
{
    gcptr p,h;

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

    return(h);
}

153 154 155 156 157 158 159

stg_fetchAddIntArrayzh( gcptr arr, W_ ind, W_ incr )
/* MutableByteArray# s -> Int# -> Int# -> State# s -> (# State# s, Int# #) */
{
    gcptr p, h;

    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
{
168 169
    W_ words, size;
    gcptr p, 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);
179 180
    TICK_ALLOC_PRIM(SIZEOF_StgMutArrPtrs, WDS(n), 0);

181
    SET_HDR(arr, stg_MUT_ARR_PTRS_DIRTY_info, CCCS);
182
    StgMutArrPtrs_ptrs(arr) = n;
183
    StgMutArrPtrs_size(arr) = size;
184 185 186 187 188

    // 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
189 190 191
        W_[p] = init;
        p = p + WDS(1);
        goto for;
192
    }
193 194 195
    // Initialise the mark bits with 0
  for2:
    if (p < arr + WDS(size)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
196 197 198
        W_[p] = 0;
        p = p + WDS(1);
        goto for2;
199
    }
200

201
    return (arr);
202 203
}

204
stg_unsafeThawArrayzh ( gcptr arr )
205 206 207
{
  // SUBTLETY TO DO WITH THE OLD GEN MUTABLE LIST
  //
ian@well-typed.com's avatar
ian@well-typed.com committed
208
  // A MUT_ARR_PTRS lives on the mutable list, but a MUT_ARR_PTRS_FROZEN
209 210
  // 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
211
  // the mutable list is not easy).
ian@well-typed.com's avatar
ian@well-typed.com committed
212
  //
213
  // So that we can tell whether a MUT_ARR_PTRS_FROZEN is on the mutable list,
214 215 216
  // 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.
  //
217 218
  // 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
219
  // the closure type is MUT_ARR_PTRS_FROZEN0 if it is on the mutable list,
220 221 222
  // 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.
223
  //
224 225 226
  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
227 228
        // must be done after SET_INFO, because it ASSERTs closure_MUTABLE()
        return (arr);
229
  } else {
ian@well-typed.com's avatar
ian@well-typed.com committed
230 231
        SET_INFO(arr,stg_MUT_ARR_PTRS_DIRTY_info);
        return (arr);
232 233 234
  }
}

235
// RRN: Uses the ticketed approach; see casMutVar
236
stg_casArrayzh ( gcptr arr, W_ ind, gcptr old, gcptr new )
237 238
/* MutableArray# s a -> Int# -> a -> a -> State# s -> (# State# s, Int#, a #) */
{
239 240
    gcptr p, h;
    W_ len;
241 242

    p = arr + SIZEOF_StgMutArrPtrs + WDS(ind);
243
    (h) = ccall cas(p, old, new);
244 245 246
    
    if (h != old) {
        // Failure, return what was there instead of 'old':
247
        return (1,h);
248 249
    } else {
        // Compare and Swap Succeeded:
250 251 252 253
	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;
254
        return (0,new);
255 256 257
    }
}

258
stg_newArrayArrayzh ( W_ n /* words */ )
259
{
260 261
    W_ words, size;
    gcptr p, arr;
262

263
    MAYBE_GC_N(stg_newArrayArrayzh, n);
264 265 266 267 268 269

    // 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;
270
    ("ptr" arr) = ccall allocate(MyCapability() "ptr",words);
271 272 273 274 275 276 277 278 279 280
    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
281 282 283
        W_[p] = arr;
        p = p + WDS(1);
        goto for;
284 285 286 287
    }
    // Initialise the mark bits with 0
  for2:
    if (p < arr + WDS(size)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
288 289 290
        W_[p] = 0;
        p = p + WDS(1);
        goto for2;
291 292
    }

293
    return (arr);
294 295
}

pumpkin's avatar
pumpkin committed
296

297 298 299 300
/* -----------------------------------------------------------------------------
   MutVar primitives
   -------------------------------------------------------------------------- */

301
stg_newMutVarzh ( gcptr init )
302 303 304
{
    W_ mv;

305
    ALLOC_PRIM_P (SIZEOF_StgMutVar, stg_newMutVarzh, init);
306 307

    mv = Hp - SIZEOF_StgMutVar + WDS(1);
308
    SET_HDR(mv,stg_MUT_VAR_DIRTY_info,CCCS);
309
    StgMutVar_var(mv) = init;
ian@well-typed.com's avatar
ian@well-typed.com committed
310

311
    return (mv);
312 313
}

314 315 316 317 318
// 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.
319
stg_casMutVarzh ( gcptr mv, gcptr old, gcptr new )
Simon Marlow's avatar
Simon Marlow committed
320 321
 /* MutVar# s a -> a -> a -> State# s -> (# State#, Int#, a #) */
{
322
    gcptr h;
Simon Marlow's avatar
Simon Marlow committed
323

324 325
    (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var,
                          old, new);
Simon Marlow's avatar
Simon Marlow committed
326
    if (h != old) {
327
        return (1,h);
Simon Marlow's avatar
Simon Marlow committed
328
    } else {
329
        if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
330
           ccall dirty_MUT_VAR(BaseReg "ptr", mv "ptr");
331
        }
332
        return (0,new);
Simon Marlow's avatar
Simon Marlow committed
333 334 335
    }
}

336
stg_atomicModifyMutVarzh ( gcptr mv, gcptr f )
337
{
338
    W_ z, x, y, r, h;
339

ian@well-typed.com's avatar
ian@well-typed.com committed
340
    /* If x is the current contents of the MutVar#, then
341 342 343
       We want to make the new contents point to

         (sel_0 (f x))
ian@well-typed.com's avatar
ian@well-typed.com committed
344

345
       and the return value is
ian@well-typed.com's avatar
ian@well-typed.com committed
346 347

         (sel_1 (f x))
348 349 350 351

        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
352
         y = [stg_sel_0 z]   (max (HS + 1) MIN_UPD_SIZE)
353 354 355 356
         r = [stg_sel_1 z]   (max (HS + 1) MIN_UPD_SIZE)
    */

#if MIN_UPD_SIZE > 1
357
#define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
358 359
#define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),WDS(MIN_UPD_SIZE-1))
#else
360
#define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(1))
361 362 363 364
#define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),0)
#endif

#if MIN_UPD_SIZE > 2
365
#define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
366 367
#define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),WDS(MIN_UPD_SIZE-2))
#else
368
#define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(2))
369 370 371 372 373
#define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),0)
#endif

#define SIZE (THUNK_2_SIZE + THUNK_1_SIZE + THUNK_1_SIZE)

374
   HP_CHK_GEN_TICKY(SIZE);
375 376 377 378

   TICK_ALLOC_THUNK_2();
   CCCS_ALLOC(THUNK_2_SIZE);
   z = Hp - THUNK_2_SIZE + WDS(1);
379
   SET_HDR(z, stg_ap_2_upd_info, CCCS);
380
   LDV_RECORD_CREATE(z);
381
   StgThunk_payload(z,0) = f;
382 383 384 385

   TICK_ALLOC_THUNK_1();
   CCCS_ALLOC(THUNK_1_SIZE);
   y = z - THUNK_1_SIZE;
386
   SET_HDR(y, stg_sel_0_upd_info, CCCS);
387
   LDV_RECORD_CREATE(y);
388
   StgThunk_payload(y,0) = z;
389 390 391 392

   TICK_ALLOC_THUNK_1();
   CCCS_ALLOC(THUNK_1_SIZE);
   r = y - THUNK_1_SIZE;
393
   SET_HDR(r, stg_sel_1_upd_info, CCCS);
394
   LDV_RECORD_CREATE(r);
395 396
   StgThunk_payload(r,0) = z;

397 398 399 400
 retry:
   x = StgMutVar_var(mv);
   StgThunk_payload(z,1) = x;
#ifdef THREADED_RTS
401
   (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var, x, y);
402 403 404
   if (h != x) { goto retry; }
#else
   StgMutVar_var(mv) = y;
405
#endif
406

407
   if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
408
     ccall dirty_MUT_VAR(BaseReg "ptr", mv "ptr");
409 410
   }

411
   return (r);
412 413 414 415 416 417 418 419
}

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

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

420 421 422
stg_mkWeakzh ( gcptr key,
               gcptr value,
               gcptr finalizer /* or stg_NO_FINALIZER_closure */ )
423
{
424
  gcptr w;
425

426
  ALLOC_PRIM (SIZEOF_StgWeak)
427 428

  w = Hp - SIZEOF_StgWeak + WDS(1);
429
  SET_HDR(w, stg_WEAK_info, CCCS);
430

431 432 433 434
  StgWeak_key(w)         = key;
  StgWeak_value(w)       = value;
  StgWeak_finalizer(w)   = finalizer;
  StgWeak_cfinalizers(w) = stg_NO_FINALIZER_closure;
435

436
  ACQUIRE_LOCK(sm_mutex);
437 438
  StgWeak_link(w) = generation_weak_ptr_list(W_[g0]);
  generation_weak_ptr_list(W_[g0]) = w;
439
  RELEASE_LOCK(sm_mutex);
440

441
  IF_DEBUG(weak, ccall debugBelch(stg_weak_msg,w));
442

443
  return (w);
444 445
}

446
stg_mkWeakNoFinalizzerzh ( gcptr key, gcptr value )
447
{
448
  jump stg_mkWeakzh (key, value, stg_NO_FINALIZER_closure);
449 450
}

451 452 453 454 455 456 457
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 )
458
{
459
  W_ c, info;
460

461
  LOCK_CLOSURE(w, info);
462 463 464 465 466 467 468

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

469
  ALLOC_PRIM (SIZEOF_StgCFinalizerList)
470

471 472
  c = Hp - SIZEOF_StgCFinalizerList + WDS(1);
  SET_HDR(c, stg_C_FINALIZER_LIST_info, CCCS);
473

474 475 476 477
  StgCFinalizerList_fptr(c) = fptr;
  StgCFinalizerList_ptr(c) = ptr;
  StgCFinalizerList_eptr(c) = eptr;
  StgCFinalizerList_flag(c) = flag;
478

479 480
  StgCFinalizerList_link(c) = StgWeak_cfinalizers(w);
  StgWeak_cfinalizers(w) = c;
481

482
  unlockClosure(w, info);
483

484
  recordMutable(w);
485

486
  IF_DEBUG(weak, ccall debugBelch(stg_cfinalizer_msg,w));
487

488
  return (1);
489
}
490

491
stg_finalizzeWeakzh ( gcptr w )
492
{
493 494 495
  gcptr f, list;
  W_ info;

496
  LOCK_CLOSURE(w, info);
497 498

  // already dead?
499 500
  if (info == stg_DEAD_WEAK_info) {
      unlockClosure(w, info);
501
      return (0,stg_NO_FINALIZER_closure);
502 503
  }

504 505 506
  f    = StgWeak_finalizer(w);
  list = StgWeak_cfinalizers(w);

507 508 509 510 511 512 513 514
  // 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
515
  // Furthermore, when PROFILING is turned on, dead weak pointers are exactly as
516 517 518 519 520 521 522
  // 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()?
  //
523
  unlockClosure(w, stg_DEAD_WEAK_info);
524

525
  LDV_RECORD_CREATE(w);
526

527 528
  if (list != stg_NO_FINALIZER_closure) {
    ccall runCFinalizers(list);
529 530
  }

531 532
  /* return the finalizer */
  if (f == stg_NO_FINALIZER_closure) {
533
      return (0,stg_NO_FINALIZER_closure);
534
  } else {
535
      return (1,f);
536 537 538
  }
}

539
stg_deRefWeakzh ( gcptr w )
540
{
541
  W_ code, info;
542
  gcptr val;
543

544 545 546 547 548 549 550
  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.

551
    LOCK_CLOSURE(w, info);
552 553 554 555
    unlockClosure(w, info);
  }

  if (info == stg_WEAK_info) {
556 557 558 559 560 561
    code = 1;
    val = StgWeak_value(w);
  } else {
    code = 0;
    val = w;
  }
562
  return (code,val);
563 564 565
}

/* -----------------------------------------------------------------------------
566
   Floating point operations.
567 568
   -------------------------------------------------------------------------- */

569
stg_decodeFloatzuIntzh ( F_ arg )
ian@well-typed.com's avatar
ian@well-typed.com committed
570
{
571
    W_ p;
572 573 574
    W_ mp_tmp1;
    W_ mp_tmp_w;

575
    STK_CHK_GEN_N (WDS(2));
576 577 578

    mp_tmp1  = Sp - WDS(1);
    mp_tmp_w = Sp - WDS(2);
ian@well-typed.com's avatar
ian@well-typed.com committed
579

580
    /* Perform the operation */
581
    ccall __decodeFloat_Int(mp_tmp1 "ptr", mp_tmp_w "ptr", arg);
ian@well-typed.com's avatar
ian@well-typed.com committed
582

583
    /* returns: (Int# (mantissa), Int# (exponent)) */
584
    return (W_[mp_tmp1], W_[mp_tmp_w]);
585 586
}

587
stg_decodeDoublezu2Intzh ( D_ arg )
ian@well-typed.com's avatar
ian@well-typed.com committed
588
{
589
    W_ p;
590 591 592 593 594
    W_ mp_tmp1;
    W_ mp_tmp2;
    W_ mp_result1;
    W_ mp_result2;

595
    STK_CHK_GEN_N (WDS(4));
596 597 598 599 600

    mp_tmp1    = Sp - WDS(1);
    mp_tmp2    = Sp - WDS(2);
    mp_result1 = Sp - WDS(3);
    mp_result2 = Sp - WDS(4);
601 602

    /* Perform the operation */
603
    ccall __decodeDouble_2Int(mp_tmp1 "ptr", mp_tmp2 "ptr",
604
                                    mp_result1 "ptr", mp_result2 "ptr",
605
                                    arg);
606 607 608

    /* returns:
       (Int# (mant sign), Word# (mant high), Word# (mant low), Int# (expn)) */
609
    return (W_[mp_tmp1], W_[mp_tmp2], W_[mp_result1], W_[mp_result2]);
610 611
}

612 613 614 615
/* -----------------------------------------------------------------------------
 * Concurrency primitives
 * -------------------------------------------------------------------------- */

616
stg_forkzh ( gcptr closure )
617
{
618
  MAYBE_GC_P(stg_forkzh, closure);
619

620
  gcptr threadid;
621

ian@well-typed.com's avatar
ian@well-typed.com committed
622 623
  ("ptr" threadid) = ccall createIOThread( MyCapability() "ptr",
                                RtsFlags_GcFlags_initialStkSize(RtsFlags),
624
                                closure "ptr");
625 626

  /* start blocked if the current thread is blocked */
627
  StgTSO_flags(threadid) = %lobits16(
ian@well-typed.com's avatar
ian@well-typed.com committed
628
     TO_W_(StgTSO_flags(threadid)) |
629
     TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
630

631
  ccall scheduleThread(MyCapability() "ptr", threadid "ptr");
632

633 634
  // context switch soon, but not immediately: we don't want every
  // forkIO to force a context-switch.
635
  Capability_context_switch(MyCapability()) = 1 :: CInt;
ian@well-typed.com's avatar
ian@well-typed.com committed
636

637
  return (threadid);
638 639
}

640
stg_forkOnzh ( W_ cpu, gcptr closure )
641
{
642
again: MAYBE_GC(again);
643

644
  gcptr threadid;
645

ian@well-typed.com's avatar
ian@well-typed.com committed
646 647
  ("ptr" threadid) = ccall createIOThread( MyCapability() "ptr",
                                RtsFlags_GcFlags_initialStkSize(RtsFlags),
648
                                closure "ptr");
649 650

  /* start blocked if the current thread is blocked */
651
  StgTSO_flags(threadid) = %lobits16(
ian@well-typed.com's avatar
ian@well-typed.com committed
652
     TO_W_(StgTSO_flags(threadid)) |
653
     TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
654

655
  ccall scheduleThreadOn(MyCapability() "ptr", cpu, threadid "ptr");
656

657 658
  // context switch soon, but not immediately: we don't want every
  // forkIO to force a context-switch.
659
  Capability_context_switch(MyCapability()) = 1 :: CInt;
ian@well-typed.com's avatar
ian@well-typed.com committed
660

661
  return (threadid);
662 663
}

664
stg_yieldzh ()
665
{
666 667 668 669 670
  // 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;
671
  jump stg_yield_noregs();
672 673
}

674
stg_myThreadIdzh ()
675
{
676
  return (CurrentTSO);
677 678
}

679
stg_labelThreadzh ( gcptr threadid, W_ addr )
680
{
681
#if defined(DEBUG) || defined(TRACING) || defined(DTRACE)
682
  ccall labelThread(MyCapability() "ptr", threadid "ptr", addr "ptr");
683
#endif
684
  return ();
685 686
}

687
stg_isCurrentThreadBoundzh (/* no args */)
688 689
{
  W_ r;
690 691
  (r) = ccall isThreadBound(CurrentTSO);
  return (r);
692 693
}

694
stg_threadStatuszh ( gcptr tso )
695 696 697
{
    W_ why_blocked;
    W_ what_next;
698
    W_ ret, cap, locked;
699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715

    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;
        }
    }
716 717 718 719 720 721 722 723 724

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

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

725
    return (ret,cap,locked);
726
}
727 728 729 730 731

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

732 733 734
// Catch retry frame -----------------------------------------------------------

#define CATCH_RETRY_FRAME_FIELDS(w_,p_,info_ptr,        \
Simon Marlow's avatar
Simon Marlow committed
735
                                 p1, p2,                \
736 737 738 739
                                 running_alt_code,      \
                                 first_code,            \
                                 alt_code)              \
  w_ info_ptr,                                          \
Simon Marlow's avatar
Simon Marlow committed
740
  PROF_HDR_FIELDS(w_,p1,p2)                             \
741 742 743
  w_ running_alt_code,                                  \
  p_ first_code,                                        \
  p_ alt_code
744 745


746
INFO_TABLE_RET(stg_catch_retry_frame, CATCH_RETRY_FRAME,
747
               CATCH_RETRY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
748
                                        info_ptr, p1, p2,
749 750 751 752
                                        running_alt_code,
                                        first_code,
                                        alt_code))
    return (P_ ret)
753
{
754 755
    W_ r;
    gcptr trec, outer, arg;
756

757 758 759 760 761 762 763 764 765 766 767 768 769 770 771
    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
772
                (CATCH_RETRY_FRAME_FIELDS(,,info_ptr, p1, p2,
773 774 775 776 777 778
                                          running_alt_code,
                                          first_code,
                                          alt_code))
                (alt_code);
        } else {
            jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
779
                (CATCH_RETRY_FRAME_FIELDS(,,info_ptr, p1, p2,
780 781 782 783 784 785 786
                                          running_alt_code,
                                          first_code,
                                          alt_code))
                (first_code);
        }
    }
}
787

Simon Marlow's avatar
Simon Marlow committed
788
// Atomically frame ------------------------------------------------------------
789

790
// This must match StgAtomicallyFrame in Closures.h
Simon Marlow's avatar
Simon Marlow committed
791
#define ATOMICALLY_FRAME_FIELDS(w_,p_,info_ptr,p1,p2,code,next,result)  \
792
  w_ info_ptr,                                                          \
Simon Marlow's avatar
Simon Marlow committed
793
  PROF_HDR_FIELDS(w_,p1,p2)                                             \
794 795 796 797 798
  p_ code,                                                              \
  p_ next,                                                              \
  p_ result


799
INFO_TABLE_RET(stg_atomically_frame, ATOMICALLY_FRAME,
800 801
               // layout of the frame, and bind the field names
               ATOMICALLY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
802
                                       info_ptr, p1, p2,
803 804 805 806
                                       code,
                                       next_invariant,
                                       frame_result))
    return (P_ result) // value returned to the frame
807
{
808 809
  W_ valid;
  gcptr trec, outer, next_invariant, q;
810

811
  trec   = StgTSO_trec(CurrentTSO);
812
  outer  = StgTRecHeader_enclosing_trec(trec);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
813 814 815

  if (outer == NO_TREC) {
    /* First time back at the atomically frame -- pick up invariants */
816 817 818
    ("ptr" next_invariant) =
        ccall stmGetInvariantsToCheck(MyCapability() "ptr", trec "ptr");
    frame_result = result;
819 820

  } else {
tharris@microsoft.com's avatar
tharris@microsoft.com committed
821 822 823
    /* 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;
824 825
    StgInvariantCheckQueue_my_execution(next_invariant) = trec;
    ccall stmAbortTransaction(MyCapability() "ptr", trec "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
826 827
    /* Don't free trec -- it's linked from q and will be stashed in the
     * invariant if we eventually commit. */
828 829
    next_invariant =
       StgInvariantCheckQueue_next_queue_entry(next_invariant);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
830 831 832
    trec = outer;
  }

833
  if (next_invariant != END_INVARIANT_CHECK_QUEUE) {
tharris@microsoft.com's avatar
tharris@microsoft.com committed
834
    /* We can't commit yet: another invariant to check */
835
    ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", trec "ptr");
836
    StgTSO_trec(CurrentTSO) = trec;
837 838
    q = StgInvariantCheckQueue_invariant(next_invariant);
    jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
839 840
        (ATOMICALLY_FRAME_FIELDS(,,info_ptr,p1,p2,
                                 code,next_invariant,frame_result))
841
        (StgAtomicInvariant_code(q));
tharris@microsoft.com's avatar
tharris@microsoft.com committed
842 843 844 845

  } else {

    /* We've got no more invariants to check, try to commit */
846
    (valid) = ccall stmCommitTransaction(MyCapability() "ptr", trec "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
847 848 849
    if (valid != 0) {
      /* Transaction was valid: commit succeeded */
      StgTSO_trec(CurrentTSO) = NO_TREC;
850
      return (frame_result);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
851 852
    } else {
      /* Transaction was not valid: try again */
853
      ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", NO_TREC "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
854
      StgTSO_trec(CurrentTSO) = trec;
855 856 857 858 859
      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
860 861
          (ATOMICALLY_FRAME_FIELDS(,,info_ptr,p1,p2,
                                   code,next_invariant,frame_result))
862
          (code);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
863
    }
864 865 866
  }
}

867

868
INFO_TABLE_RET(stg_atomically_waiting_frame, ATOMICALLY_FRAME,
869 870
               // layout of the frame, and bind the field names
               ATOMICALLY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
871
                                       info_ptr, p1, p2,
872 873 874 875
                                       code,
                                       next_invariant,
                                       frame_result))
    return (/* no return values */)
876
{
877
  W_ trec, valid;
878 879

  /* The TSO is currently waiting: should we stop waiting? */
880
<