PrimOps.cmm 65.2 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
}

140 141 142 143 144 145 146 147 148 149 150 151 152 153
// 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# #) */
{
    W_ len;
    gcptr p,h;

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

    return(h);
}

154
stg_newArrayzh ( W_ n /* words */, gcptr init )
155
{
156 157
    W_ words, size;
    gcptr p, arr;
158

159
    again: MAYBE_GC(again);
160

161 162 163 164 165
    // 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;
166
    ("ptr" arr) = ccall allocate(MyCapability() "ptr",words);
167 168
    TICK_ALLOC_PRIM(SIZEOF_StgMutArrPtrs, WDS(n), 0);

169
    SET_HDR(arr, stg_MUT_ARR_PTRS_DIRTY_info, CCCS);
170
    StgMutArrPtrs_ptrs(arr) = n;
171
    StgMutArrPtrs_size(arr) = size;
172 173 174 175 176

    // 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
177 178 179
        W_[p] = init;
        p = p + WDS(1);
        goto for;
180
    }
181 182 183
    // Initialise the mark bits with 0
  for2:
    if (p < arr + WDS(size)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
184 185 186
        W_[p] = 0;
        p = p + WDS(1);
        goto for2;
187
    }
188

189
    return (arr);
190 191
}

192
stg_unsafeThawArrayzh ( gcptr arr )
193 194 195
{
  // SUBTLETY TO DO WITH THE OLD GEN MUTABLE LIST
  //
ian@well-typed.com's avatar
ian@well-typed.com committed
196
  // A MUT_ARR_PTRS lives on the mutable list, but a MUT_ARR_PTRS_FROZEN
197 198
  // 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
199
  // the mutable list is not easy).
ian@well-typed.com's avatar
ian@well-typed.com committed
200
  //
201
  // So that we can tell whether a MUT_ARR_PTRS_FROZEN is on the mutable list,
202 203 204
  // 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.
  //
205 206
  // 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
207
  // the closure type is MUT_ARR_PTRS_FROZEN0 if it is on the mutable list,
208 209 210
  // 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.
211
  //
212 213 214
  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
215 216
        // must be done after SET_INFO, because it ASSERTs closure_MUTABLE()
        return (arr);
217
  } else {
ian@well-typed.com's avatar
ian@well-typed.com committed
218 219
        SET_INFO(arr,stg_MUT_ARR_PTRS_DIRTY_info);
        return (arr);
220 221 222
  }
}

223
// RRN: Uses the ticketed approach; see casMutVar
224
stg_casArrayzh ( gcptr arr, W_ ind, gcptr old, gcptr new )
225 226
/* MutableArray# s a -> Int# -> a -> a -> State# s -> (# State# s, Int#, a #) */
{
227 228
    gcptr p, h;
    W_ len;
229 230

    p = arr + SIZEOF_StgMutArrPtrs + WDS(ind);
231
    (h) = ccall cas(p, old, new);
232 233 234
    
    if (h != old) {
        // Failure, return what was there instead of 'old':
235
        return (1,h);
236 237
    } else {
        // Compare and Swap Succeeded:
238 239 240 241
	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;
242
        return (0,new);
243 244 245
    }
}

246
stg_newArrayArrayzh ( W_ n /* words */ )
247
{
248 249
    W_ words, size;
    gcptr p, arr;
250

251
    MAYBE_GC_N(stg_newArrayArrayzh, n);
252 253 254 255 256 257

    // 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;
258
    ("ptr" arr) = ccall allocate(MyCapability() "ptr",words);
259 260 261 262 263 264 265 266 267 268
    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
269 270 271
        W_[p] = arr;
        p = p + WDS(1);
        goto for;
272 273 274 275
    }
    // Initialise the mark bits with 0
  for2:
    if (p < arr + WDS(size)) {
ian@well-typed.com's avatar
ian@well-typed.com committed
276 277 278
        W_[p] = 0;
        p = p + WDS(1);
        goto for2;
279 280
    }

281
    return (arr);
282 283
}

pumpkin's avatar
pumpkin committed
284

285 286 287 288
/* -----------------------------------------------------------------------------
   MutVar primitives
   -------------------------------------------------------------------------- */

289
stg_newMutVarzh ( gcptr init )
290 291 292
{
    W_ mv;

293
    ALLOC_PRIM_P (SIZEOF_StgMutVar, stg_newMutVarzh, init);
294 295

    mv = Hp - SIZEOF_StgMutVar + WDS(1);
296
    SET_HDR(mv,stg_MUT_VAR_DIRTY_info,CCCS);
297
    StgMutVar_var(mv) = init;
ian@well-typed.com's avatar
ian@well-typed.com committed
298

299
    return (mv);
300 301
}

302 303 304 305 306
// 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.
307
stg_casMutVarzh ( gcptr mv, gcptr old, gcptr new )
Simon Marlow's avatar
Simon Marlow committed
308 309
 /* MutVar# s a -> a -> a -> State# s -> (# State#, Int#, a #) */
{
310
    gcptr h;
Simon Marlow's avatar
Simon Marlow committed
311

312 313
    (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var,
                          old, new);
Simon Marlow's avatar
Simon Marlow committed
314
    if (h != old) {
315
        return (1,h);
Simon Marlow's avatar
Simon Marlow committed
316
    } else {
317
        if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
318
           ccall dirty_MUT_VAR(BaseReg "ptr", mv "ptr");
319
        }
320
        return (0,new);
Simon Marlow's avatar
Simon Marlow committed
321 322 323
    }
}

324
stg_atomicModifyMutVarzh ( gcptr mv, gcptr f )
325
{
326
    W_ z, x, y, r, h;
327

ian@well-typed.com's avatar
ian@well-typed.com committed
328
    /* If x is the current contents of the MutVar#, then
329 330 331
       We want to make the new contents point to

         (sel_0 (f x))
ian@well-typed.com's avatar
ian@well-typed.com committed
332

333
       and the return value is
ian@well-typed.com's avatar
ian@well-typed.com committed
334 335

         (sel_1 (f x))
336 337 338 339

        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
340
         y = [stg_sel_0 z]   (max (HS + 1) MIN_UPD_SIZE)
341 342 343 344
         r = [stg_sel_1 z]   (max (HS + 1) MIN_UPD_SIZE)
    */

#if MIN_UPD_SIZE > 1
345
#define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
346 347
#define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),WDS(MIN_UPD_SIZE-1))
#else
348
#define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(1))
349 350 351 352
#define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),0)
#endif

#if MIN_UPD_SIZE > 2
353
#define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
354 355
#define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),WDS(MIN_UPD_SIZE-2))
#else
356
#define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(2))
357 358 359 360 361
#define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),0)
#endif

#define SIZE (THUNK_2_SIZE + THUNK_1_SIZE + THUNK_1_SIZE)

362
   HP_CHK_GEN_TICKY(SIZE);
363 364 365 366

   TICK_ALLOC_THUNK_2();
   CCCS_ALLOC(THUNK_2_SIZE);
   z = Hp - THUNK_2_SIZE + WDS(1);
367
   SET_HDR(z, stg_ap_2_upd_info, CCCS);
368
   LDV_RECORD_CREATE(z);
369
   StgThunk_payload(z,0) = f;
370 371 372 373

   TICK_ALLOC_THUNK_1();
   CCCS_ALLOC(THUNK_1_SIZE);
   y = z - THUNK_1_SIZE;
374
   SET_HDR(y, stg_sel_0_upd_info, CCCS);
375
   LDV_RECORD_CREATE(y);
376
   StgThunk_payload(y,0) = z;
377 378 379 380

   TICK_ALLOC_THUNK_1();
   CCCS_ALLOC(THUNK_1_SIZE);
   r = y - THUNK_1_SIZE;
381
   SET_HDR(r, stg_sel_1_upd_info, CCCS);
382
   LDV_RECORD_CREATE(r);
383 384
   StgThunk_payload(r,0) = z;

385 386 387 388
 retry:
   x = StgMutVar_var(mv);
   StgThunk_payload(z,1) = x;
#ifdef THREADED_RTS
389
   (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var, x, y);
390 391 392
   if (h != x) { goto retry; }
#else
   StgMutVar_var(mv) = y;
393
#endif
394

395
   if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
396
     ccall dirty_MUT_VAR(BaseReg "ptr", mv "ptr");
397 398
   }

399
   return (r);
400 401 402 403 404 405 406 407
}

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

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

408 409 410
stg_mkWeakzh ( gcptr key,
               gcptr value,
               gcptr finalizer /* or stg_NO_FINALIZER_closure */ )
411
{
412
  gcptr w;
413

414
  ALLOC_PRIM (SIZEOF_StgWeak)
415 416

  w = Hp - SIZEOF_StgWeak + WDS(1);
417
  SET_HDR(w, stg_WEAK_info, CCCS);
418

419 420 421 422
  StgWeak_key(w)         = key;
  StgWeak_value(w)       = value;
  StgWeak_finalizer(w)   = finalizer;
  StgWeak_cfinalizers(w) = stg_NO_FINALIZER_closure;
423

424
  ACQUIRE_LOCK(sm_mutex);
425 426
  StgWeak_link(w) = generation_weak_ptr_list(W_[g0]);
  generation_weak_ptr_list(W_[g0]) = w;
427
  RELEASE_LOCK(sm_mutex);
428

429
  IF_DEBUG(weak, ccall debugBelch(stg_weak_msg,w));
430

431
  return (w);
432 433
}

434
stg_mkWeakNoFinalizzerzh ( gcptr key, gcptr value )
435
{
436
  jump stg_mkWeakzh (key, value, stg_NO_FINALIZER_closure);
437 438
}

439 440 441 442 443 444 445
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 )
446
{
447
  W_ c, info;
448

449
  LOCK_CLOSURE(w, info);
450 451 452 453 454 455 456

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

457
  ALLOC_PRIM (SIZEOF_StgCFinalizerList)
458

459 460
  c = Hp - SIZEOF_StgCFinalizerList + WDS(1);
  SET_HDR(c, stg_C_FINALIZER_LIST_info, CCCS);
461

462 463 464 465
  StgCFinalizerList_fptr(c) = fptr;
  StgCFinalizerList_ptr(c) = ptr;
  StgCFinalizerList_eptr(c) = eptr;
  StgCFinalizerList_flag(c) = flag;
466

467 468
  StgCFinalizerList_link(c) = StgWeak_cfinalizers(w);
  StgWeak_cfinalizers(w) = c;
469

470
  unlockClosure(w, info);
471

472
  recordMutable(w);
473

474
  IF_DEBUG(weak, ccall debugBelch(stg_cfinalizer_msg,w));
475

476
  return (1);
477
}
478

479
stg_finalizzeWeakzh ( gcptr w )
480
{
481 482 483
  gcptr f, list;
  W_ info;

484
  LOCK_CLOSURE(w, info);
485 486

  // already dead?
487 488
  if (info == stg_DEAD_WEAK_info) {
      unlockClosure(w, info);
489
      return (0,stg_NO_FINALIZER_closure);
490 491
  }

492 493 494
  f    = StgWeak_finalizer(w);
  list = StgWeak_cfinalizers(w);

495 496 497 498 499 500 501 502
  // 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
503
  // Furthermore, when PROFILING is turned on, dead weak pointers are exactly as
504 505 506 507 508 509 510
  // 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()?
  //
511
  unlockClosure(w, stg_DEAD_WEAK_info);
512

513
  LDV_RECORD_CREATE(w);
514

515 516
  if (list != stg_NO_FINALIZER_closure) {
    ccall runCFinalizers(list);
517 518
  }

519 520
  /* return the finalizer */
  if (f == stg_NO_FINALIZER_closure) {
521
      return (0,stg_NO_FINALIZER_closure);
522
  } else {
523
      return (1,f);
524 525 526
  }
}

527
stg_deRefWeakzh ( gcptr w )
528
{
529
  W_ code, info;
530
  gcptr val;
531

532 533 534 535 536 537 538
  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.

539
    LOCK_CLOSURE(w, info);
540 541 542 543
    unlockClosure(w, info);
  }

  if (info == stg_WEAK_info) {
544 545 546 547 548 549
    code = 1;
    val = StgWeak_value(w);
  } else {
    code = 0;
    val = w;
  }
550
  return (code,val);
551 552 553
}

/* -----------------------------------------------------------------------------
554
   Floating point operations.
555 556
   -------------------------------------------------------------------------- */

557
stg_decodeFloatzuIntzh ( F_ arg )
ian@well-typed.com's avatar
ian@well-typed.com committed
558
{
559
    W_ p;
560 561 562
    W_ mp_tmp1;
    W_ mp_tmp_w;

563
    STK_CHK_GEN_N (WDS(2));
564 565 566

    mp_tmp1  = Sp - WDS(1);
    mp_tmp_w = Sp - WDS(2);
ian@well-typed.com's avatar
ian@well-typed.com committed
567

568
    /* Perform the operation */
569
    ccall __decodeFloat_Int(mp_tmp1 "ptr", mp_tmp_w "ptr", arg);
ian@well-typed.com's avatar
ian@well-typed.com committed
570

571
    /* returns: (Int# (mantissa), Int# (exponent)) */
572
    return (W_[mp_tmp1], W_[mp_tmp_w]);
573 574
}

575
stg_decodeDoublezu2Intzh ( D_ arg )
ian@well-typed.com's avatar
ian@well-typed.com committed
576
{
577
    W_ p;
578 579 580 581 582
    W_ mp_tmp1;
    W_ mp_tmp2;
    W_ mp_result1;
    W_ mp_result2;

583
    STK_CHK_GEN_N (WDS(4));
584 585 586 587 588

    mp_tmp1    = Sp - WDS(1);
    mp_tmp2    = Sp - WDS(2);
    mp_result1 = Sp - WDS(3);
    mp_result2 = Sp - WDS(4);
589 590

    /* Perform the operation */
591
    ccall __decodeDouble_2Int(mp_tmp1 "ptr", mp_tmp2 "ptr",
592
                                    mp_result1 "ptr", mp_result2 "ptr",
593
                                    arg);
594 595 596

    /* returns:
       (Int# (mant sign), Word# (mant high), Word# (mant low), Int# (expn)) */
597
    return (W_[mp_tmp1], W_[mp_tmp2], W_[mp_result1], W_[mp_result2]);
598 599
}

600 601 602 603
/* -----------------------------------------------------------------------------
 * Concurrency primitives
 * -------------------------------------------------------------------------- */

604
stg_forkzh ( gcptr closure )
605
{
606
  MAYBE_GC_P(stg_forkzh, closure);
607

608
  gcptr threadid;
609

ian@well-typed.com's avatar
ian@well-typed.com committed
610 611
  ("ptr" threadid) = ccall createIOThread( MyCapability() "ptr",
                                RtsFlags_GcFlags_initialStkSize(RtsFlags),
612
                                closure "ptr");
613 614

  /* start blocked if the current thread is blocked */
615
  StgTSO_flags(threadid) = %lobits16(
ian@well-typed.com's avatar
ian@well-typed.com committed
616
     TO_W_(StgTSO_flags(threadid)) |
617
     TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
618

619
  ccall scheduleThread(MyCapability() "ptr", threadid "ptr");
620

621 622
  // context switch soon, but not immediately: we don't want every
  // forkIO to force a context-switch.
623
  Capability_context_switch(MyCapability()) = 1 :: CInt;
ian@well-typed.com's avatar
ian@well-typed.com committed
624

625
  return (threadid);
626 627
}

628
stg_forkOnzh ( W_ cpu, gcptr closure )
629
{
630
again: MAYBE_GC(again);
631

632
  gcptr threadid;
633

ian@well-typed.com's avatar
ian@well-typed.com committed
634 635
  ("ptr" threadid) = ccall createIOThread( MyCapability() "ptr",
                                RtsFlags_GcFlags_initialStkSize(RtsFlags),
636
                                closure "ptr");
637 638

  /* start blocked if the current thread is blocked */
639
  StgTSO_flags(threadid) = %lobits16(
ian@well-typed.com's avatar
ian@well-typed.com committed
640
     TO_W_(StgTSO_flags(threadid)) |
641
     TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
642

643
  ccall scheduleThreadOn(MyCapability() "ptr", cpu, threadid "ptr");
644

645 646
  // context switch soon, but not immediately: we don't want every
  // forkIO to force a context-switch.
647
  Capability_context_switch(MyCapability()) = 1 :: CInt;
ian@well-typed.com's avatar
ian@well-typed.com committed
648

649
  return (threadid);
650 651
}

652
stg_yieldzh ()
653
{
654 655 656 657 658
  // 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;
659
  jump stg_yield_noregs();
660 661
}

662
stg_myThreadIdzh ()
663
{
664
  return (CurrentTSO);
665 666
}

667
stg_labelThreadzh ( gcptr threadid, W_ addr )
668
{
669
#if defined(DEBUG) || defined(TRACING) || defined(DTRACE)
670
  ccall labelThread(MyCapability() "ptr", threadid "ptr", addr "ptr");
671
#endif
672
  return ();
673 674
}

675
stg_isCurrentThreadBoundzh (/* no args */)
676 677
{
  W_ r;
678 679
  (r) = ccall isThreadBound(CurrentTSO);
  return (r);
680 681
}

682
stg_threadStatuszh ( gcptr tso )
683 684 685
{
    W_ why_blocked;
    W_ what_next;
686
    W_ ret, cap, locked;
687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703

    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;
        }
    }
704 705 706 707 708 709 710 711 712

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

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

713
    return (ret,cap,locked);
714
}
715 716 717 718 719

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

720 721 722
// Catch retry frame -----------------------------------------------------------

#define CATCH_RETRY_FRAME_FIELDS(w_,p_,info_ptr,        \
Simon Marlow's avatar
Simon Marlow committed
723
                                 p1, p2,                \
724 725 726 727
                                 running_alt_code,      \
                                 first_code,            \
                                 alt_code)              \
  w_ info_ptr,                                          \
Simon Marlow's avatar
Simon Marlow committed
728
  PROF_HDR_FIELDS(w_,p1,p2)                             \
729 730 731
  w_ running_alt_code,                                  \
  p_ first_code,                                        \
  p_ alt_code
732 733


734
INFO_TABLE_RET(stg_catch_retry_frame, CATCH_RETRY_FRAME,
735
               CATCH_RETRY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
736
                                        info_ptr, p1, p2,
737 738 739 740
                                        running_alt_code,
                                        first_code,
                                        alt_code))
    return (P_ ret)
741
{
742 743
    W_ r;
    gcptr trec, outer, arg;
744

745 746 747 748 749 750 751 752 753 754 755 756 757 758 759
    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
760
                (CATCH_RETRY_FRAME_FIELDS(,,info_ptr, p1, p2,
761 762 763 764 765 766
                                          running_alt_code,
                                          first_code,
                                          alt_code))
                (alt_code);
        } else {
            jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
767
                (CATCH_RETRY_FRAME_FIELDS(,,info_ptr, p1, p2,
768 769 770 771 772 773 774
                                          running_alt_code,
                                          first_code,
                                          alt_code))
                (first_code);
        }
    }
}
775

Simon Marlow's avatar
Simon Marlow committed
776
// Atomically frame ------------------------------------------------------------
777

778
// This must match StgAtomicallyFrame in Closures.h
Simon Marlow's avatar
Simon Marlow committed
779
#define ATOMICALLY_FRAME_FIELDS(w_,p_,info_ptr,p1,p2,code,next,result)  \
780
  w_ info_ptr,                                                          \
Simon Marlow's avatar
Simon Marlow committed
781
  PROF_HDR_FIELDS(w_,p1,p2)                                             \
782 783 784 785 786
  p_ code,                                                              \
  p_ next,                                                              \
  p_ result


787
INFO_TABLE_RET(stg_atomically_frame, ATOMICALLY_FRAME,
788 789
               // layout of the frame, and bind the field names
               ATOMICALLY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
790
                                       info_ptr, p1, p2,
791 792 793 794
                                       code,
                                       next_invariant,
                                       frame_result))
    return (P_ result) // value returned to the frame
795
{
796 797
  W_ valid;
  gcptr trec, outer, next_invariant, q;
798

799
  trec   = StgTSO_trec(CurrentTSO);
800
  outer  = StgTRecHeader_enclosing_trec(trec);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
801 802 803

  if (outer == NO_TREC) {
    /* First time back at the atomically frame -- pick up invariants */
804 805 806
    ("ptr" next_invariant) =
        ccall stmGetInvariantsToCheck(MyCapability() "ptr", trec "ptr");
    frame_result = result;
807 808

  } else {
tharris@microsoft.com's avatar
tharris@microsoft.com committed
809 810 811
    /* 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;
812 813
    StgInvariantCheckQueue_my_execution(next_invariant) = trec;
    ccall stmAbortTransaction(MyCapability() "ptr", trec "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
814 815
    /* Don't free trec -- it's linked from q and will be stashed in the
     * invariant if we eventually commit. */
816 817
    next_invariant =
       StgInvariantCheckQueue_next_queue_entry(next_invariant);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
818 819 820
    trec = outer;
  }

821
  if (next_invariant != END_INVARIANT_CHECK_QUEUE) {
tharris@microsoft.com's avatar
tharris@microsoft.com committed
822
    /* We can't commit yet: another invariant to check */
823
    ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", trec "ptr");
824
    StgTSO_trec(CurrentTSO) = trec;
825 826
    q = StgInvariantCheckQueue_invariant(next_invariant);
    jump stg_ap_v_fast
Simon Marlow's avatar
Simon Marlow committed
827 828
        (ATOMICALLY_FRAME_FIELDS(,,info_ptr,p1,p2,
                                 code,next_invariant,frame_result))
829
        (StgAtomicInvariant_code(q));
tharris@microsoft.com's avatar
tharris@microsoft.com committed
830 831 832 833

  } else {

    /* We've got no more invariants to check, try to commit */
834
    (valid) = ccall stmCommitTransaction(MyCapability() "ptr", trec "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
835 836 837
    if (valid != 0) {
      /* Transaction was valid: commit succeeded */
      StgTSO_trec(CurrentTSO) = NO_TREC;
838
      return (frame_result);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
839 840
    } else {
      /* Transaction was not valid: try again */
841
      ("ptr" trec) = ccall stmStartTransaction(MyCapability() "ptr", NO_TREC "ptr");
tharris@microsoft.com's avatar
tharris@microsoft.com committed
842
      StgTSO_trec(CurrentTSO) = trec;
843 844 845 846 847
      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
848 849
          (ATOMICALLY_FRAME_FIELDS(,,info_ptr,p1,p2,
                                   code,next_invariant,frame_result))
850
          (code);
tharris@microsoft.com's avatar
tharris@microsoft.com committed
851
    }
852 853 854
  }
}

855

856
INFO_TABLE_RET(stg_atomically_waiting_frame, ATOMICALLY_FRAME,
857 858
               // layout of the frame, and bind the field names
               ATOMICALLY_FRAME_FIELDS(W_,P_,
Simon Marlow's avatar
Simon Marlow committed
859
                                       info_ptr, p1, p2,
860 861 862 863
                                       code,
                                       next_invariant,
                                       frame_result))
    return (/* no return values */)
864
{
865
  W_ trec, valid;
866 867

  /* The TSO is currently waiting: should we stop waiting? */
868
  (valid) = ccall stmReWait(MyCapability() "ptr", CurrentTSO "ptr");
Simon Marlow's avatar
Simon Marlow committed
869
  if (valid != 0) {
870 871
      /* Previous attempt is still valid: no point trying again yet */
      jump stg_block_noregs
Simon Marlow's avatar
Simon Marlow committed
872
          (ATOMICALLY_FRAME_FIELDS(,,info_ptr, p1, p2,
873 874
                                   code,next_invariant,frame_result))
          ();
simonmar's avatar