Compact.c 23 KB
Newer Older
1 2
/* -----------------------------------------------------------------------------
 *
3
 * (c) The GHC Team 2001-2008
4 5 6
 *
 * Compacting garbage collector
 *
7 8 9 10 11
 * Documentation on the architecture of the Garbage Collector can be
 * found in the online commentary:
 * 
 *   http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
 *
12 13
 * ---------------------------------------------------------------------------*/

14
#include "PosixSource.h"
15
#include "Rts.h"
Simon Marlow's avatar
Simon Marlow committed
16

Simon Marlow's avatar
Simon Marlow committed
17
#include "GCThread.h"
Simon Marlow's avatar
Simon Marlow committed
18
#include "Storage.h"
19 20
#include "RtsUtils.h"
#include "BlockAlloc.h"
21 22
#include "GC.h"
#include "Compact.h"
23
#include "Schedule.h"
24
#include "Apply.h"
Simon Marlow's avatar
Simon Marlow committed
25
#include "Trace.h"
Simon Marlow's avatar
Simon Marlow committed
26 27 28
#include "Weak.h"
#include "MarkWeak.h"
#include "Stable.h"
29

30 31
// Turn off inlining when debugging - it obfuscates things
#ifdef DEBUG
sof's avatar
sof committed
32 33
# undef  STATIC_INLINE
# define STATIC_INLINE static
34 35
#endif

36
/* ----------------------------------------------------------------------------
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
   Threading / unthreading pointers.

   The basic idea here is to chain together all the fields pointing at
   a particular object, with the root of the chain in the object's
   info table field.  The original contents of the info pointer goes
   at the end of the chain.

   Adding a new field to the chain is a matter of swapping the
   contents of the field with the contents of the object's info table
   field.

   To unthread the chain, we walk down it updating all the fields on
   the chain with the new location of the object.  We stop when we
   reach the info pointer at the end.

52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70
   The main difficulty here is that we need to be able to identify the
   info pointer at the end of the chain.  We can't use the low bits of
   the pointer for this; they are already being used for
   pointer-tagging.  What's more, we need to retain the
   pointer-tagging tag bits on each pointer during the
   threading/unthreading process.

   Our solution is as follows: 
     - an info pointer (chain length zero) is identified by having tag 0
     - in a threaded chain of length > 0:
        - the pointer-tagging tag bits are attached to the info pointer
        - the first entry in the chain has tag 1
        - second and subsequent entries in the chain have tag 2

   This exploits the fact that the tag on each pointer to a given
   closure is normally the same (if they are not the same, then
   presumably the tag is not essential and it therefore doesn't matter
   if we throw away some of the tags).
   ------------------------------------------------------------------------- */
71

sof's avatar
sof committed
72
STATIC_INLINE void
73
thread (StgClosure **p)
74
{
75 76 77
    StgClosure *q0;
    StgPtr q;
    StgWord iptr;
78 79
    bdescr *bd;

80 81 82
    q0  = *p;
    q   = (StgPtr)UNTAG_CLOSURE(q0);

83 84 85
    // It doesn't look like a closure at the moment, because the info
    // ptr is possibly threaded:
    // ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
Simon Marlow's avatar
Simon Marlow committed
86
    
87
    if (HEAP_ALLOCED(q)) {
88
	bd = Bdescr(q); 
89

90
	if (bd->flags & BF_MARKED)
Simon Marlow's avatar
Simon Marlow committed
91
        {
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108
            iptr = *q;
            switch (GET_CLOSURE_TAG((StgClosure *)iptr))
            {
            case 0: 
                // this is the info pointer; we are creating a new chain.
                // save the original tag at the end of the chain.
                *p = (StgClosure *)((StgWord)iptr + GET_CLOSURE_TAG(q0));
                *q = (StgWord)p + 1;
                break;
            case 1:
            case 2:
                // this is a chain of length 1 or more
                *p = (StgClosure *)iptr;
                *q = (StgWord)p + 2;
                break;
            }
        }
109 110 111
    }
}

112 113 114 115 116 117
static void
thread_root (void *user STG_UNUSED, StgClosure **p)
{
    thread(p);
}

118 119 120 121
// This version of thread() takes a (void *), used to circumvent
// warnings from gcc about pointer punning and strict aliasing.
STATIC_INLINE void thread_ (void *p) { thread((StgClosure **)p); }

sof's avatar
sof committed
122
STATIC_INLINE void
123
unthread( StgPtr p, StgWord free )
124
{
125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148
    StgWord q, r;
    StgPtr q0;

    q = *p;
loop:
    switch (GET_CLOSURE_TAG((StgClosure *)q))
    {
    case 0:
        // nothing to do; the chain is length zero
        return;
    case 1:
        q0 = (StgPtr)(q-1);
        r = *q0;  // r is the info ptr, tagged with the pointer-tag
        *q0 = free;
        *p = (StgWord)UNTAG_CLOSURE((StgClosure *)r);
        return;
    case 2:
        q0 = (StgPtr)(q-2);
        r = *q0;
        *q0 = free;
        q = r;
        goto loop;
    default:
        barf("unthread");
149 150 151
    }
}

152 153 154 155 156
// Traverse a threaded chain and pull out the info pointer at the end.
// The info pointer is also tagged with the appropriate pointer tag
// for this closure, which should be attached to the pointer
// subsequently passed to unthread().
STATIC_INLINE StgWord
157 158
get_threaded_info( StgPtr p )
{
159 160 161
    StgWord q;
    
    q = (W_)GET_INFO(UNTAG_CLOSURE((StgClosure *)p));
162

163 164 165 166 167 168 169 170 171
loop:
    switch (GET_CLOSURE_TAG((StgClosure *)q)) 
    {
    case 0:
        ASSERT(LOOKS_LIKE_INFO_PTR(q));
        return q;
    case 1:
    {
        StgWord r = *(StgPtr)(q-1);
Simon Marlow's avatar
Simon Marlow committed
172
        ASSERT(LOOKS_LIKE_INFO_PTR((StgWord)UNTAG_CLOSURE((StgClosure *)r)));
173 174 175 176 177 178 179
        return r;
    }
    case 2:
        q = *(StgPtr)(q-2);
        goto loop;
    default:
        barf("get_threaded_info");
180 181 182 183 184
    }
}

// A word-aligned memmove will be faster for small objects than libc's or gcc's.
// Remember, the two regions *might* overlap, but: to <= from.
sof's avatar
sof committed
185
STATIC_INLINE void
186
move(StgPtr to, StgPtr from, W_ size)
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
{
    for(; size > 0; --size) {
	*to++ = *from++;
    }
}

static void
thread_static( StgClosure* p )
{
  const StgInfoTable *info;

  // keep going until we've threaded all the objects on the linked
  // list... 
  while (p != END_OF_STATIC_LIST) {

    info = get_itbl(p);
    switch (info->type) {
      
    case IND_STATIC:
206
	thread(&((StgInd *)p)->indirectee);
207
	p = *IND_STATIC_LINK(p);
208
	continue;
209 210
      
    case THUNK_STATIC:
211
	p = *THUNK_STATIC_LINK(p);
212
	continue;
213
    case FUN_STATIC:
214
	p = *FUN_STATIC_LINK(p);
215
	continue;
216
    case CONSTR_STATIC:
217
	p = *STATIC_LINK(info,p);
218
	continue;
219 220 221 222 223 224 225 226
      
    default:
	barf("thread_static: strange closure %d", (int)(info->type));
    }

  }
}

sof's avatar
sof committed
227
STATIC_INLINE void
228
thread_large_bitmap( StgPtr p, StgLargeBitmap *large_bitmap, W_ size )
229
{
230
    W_ i, b;
231 232 233 234 235 236
    StgWord bitmap;

    b = 0;
    bitmap = large_bitmap->bitmap[b];
    for (i = 0; i < size; ) {
	if ((bitmap & 1) == 0) {
237
	    thread((StgClosure **)p);
238 239 240 241 242 243 244 245 246 247 248 249
	}
	i++;
	p++;
	if (i % BITS_IN(W_) == 0) {
	    b++;
	    bitmap = large_bitmap->bitmap[b];
	} else {
	    bitmap = bitmap >> 1;
	}
    }
}

sof's avatar
sof committed
250
STATIC_INLINE StgPtr
251 252 253 254
thread_arg_block (StgFunInfoTable *fun_info, StgClosure **args)
{
    StgPtr p;
    StgWord bitmap;
255
    W_ size;
256 257

    p = (StgPtr)args;
258
    switch (fun_info->f.fun_type) {
259
    case ARG_GEN:
260 261
	bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
	size = BITMAP_SIZE(fun_info->f.b.bitmap);
262 263
	goto small_bitmap;
    case ARG_GEN_BIG:
264 265
	size = GET_FUN_LARGE_BITMAP(fun_info)->size;
	thread_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info), size);
266 267 268
	p += size;
	break;
    default:
269 270
	bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
	size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
271 272 273
    small_bitmap:
	while (size > 0) {
	    if ((bitmap & 1) == 0) {
274
		thread((StgClosure **)p);
275 276 277 278 279 280 281 282 283 284
	    }
	    p++;
	    bitmap = bitmap >> 1;
	    size--;
	}
	break;
    }
    return p;
}

285 286 287
static void
thread_stack(StgPtr p, StgPtr stack_end)
{
288
    const StgRetInfoTable* info;
ken's avatar
ken committed
289
    StgWord bitmap;
290
    W_ size;
291 292 293 294 295
    
    // highly similar to scavenge_stack, but we do pointer threading here.
    
    while (p < stack_end) {

296
	// *p must be the info pointer of an activation
297 298 299
	// record.  All activation records have 'bitmap' style layout
	// info.
	//
300
	info  = get_ret_itbl((StgClosure *)p);
301
	
302
	switch (info->i.type) {
303
	    
304
            // small bitmap (<= 32 entries, or 64 on a 64-bit machine)
305 306 307
        case CATCH_RETRY_FRAME:
        case CATCH_STM_FRAME:
        case ATOMICALLY_FRAME:
308
	case UPDATE_FRAME:
309 310 311
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
        case CATCH_FRAME:
312
	case RET_SMALL:
313 314
	    bitmap = BITMAP_BITS(info->i.layout.bitmap);
	    size   = BITMAP_SIZE(info->i.layout.bitmap);
315
	    p++;
316 317 318
	    // NOTE: the payload starts immediately after the info-ptr, we
	    // don't have an StgHeader in the same sense as a heap closure.
	    while (size > 0) {
319
		if ((bitmap & 1) == 0) {
320
		    thread((StgClosure **)p);
321 322 323
		}
		p++;
		bitmap = bitmap >> 1;
324
		size--;
325 326 327
	    }
	    continue;

328 329 330 331 332 333
	case RET_BCO: {
	    StgBCO *bco;
	    nat size;
	    
	    p++;
	    bco = (StgBCO *)*p;
334
	    thread((StgClosure **)p);
335 336 337 338 339 340 341
	    p++;
	    size = BCO_BITMAP_SIZE(bco);
	    thread_large_bitmap(p, BCO_BITMAP(bco), size);
	    p += size;
	    continue;
	}

ken's avatar
ken committed
342
	    // large bitmap (> 32 entries, or 64 on a 64-bit machine) 
343 344
	case RET_BIG:
	    p++;
345 346
	    size = GET_LARGE_BITMAP(&info->i)->size;
	    thread_large_bitmap(p, GET_LARGE_BITMAP(&info->i), size);
347 348
	    p += size;
	    continue;
349

350 351 352 353 354
	case RET_FUN:
	{
	    StgRetFun *ret_fun = (StgRetFun *)p;
	    StgFunInfoTable *fun_info;
	    
355
	    fun_info = FUN_INFO_PTR_TO_STRUCT((StgInfoTable *)UNTAG_CLOSURE((StgClosure *)
356
                           get_threaded_info((StgPtr)ret_fun->fun)));
357
	         // *before* threading it!
358
	    thread(&ret_fun->fun);
359
	    p = thread_arg_block(fun_info, ret_fun->payload);
360 361 362 363 364
	    continue;
	}

	default:
	    barf("thread_stack: weird activation record found on stack: %d", 
365
		 (int)(info->i.type));
366 367 368 369
	}
    }
}

sof's avatar
sof committed
370
STATIC_INLINE StgPtr
371
thread_PAP_payload (StgClosure *fun, StgClosure **payload, StgWord size)
372 373
{
    StgPtr p;
374
    StgWord bitmap;
375
    StgFunInfoTable *fun_info;
376

377
    fun_info = FUN_INFO_PTR_TO_STRUCT((StgInfoTable *)UNTAG_CLOSURE((StgClosure *)
378
                        get_threaded_info((StgPtr)fun)));
379 380
    ASSERT(fun_info->i.type != PAP);

381
    p = (StgPtr)payload;
382

383
    switch (fun_info->f.fun_type) {
384
    case ARG_GEN:
385
	bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
386 387
	goto small_bitmap;
    case ARG_GEN_BIG:
388
	thread_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info), size);
389 390 391
	p += size;
	break;
    case ARG_BCO:
392
	thread_large_bitmap((StgPtr)payload, BCO_BITMAP(fun), size);
393 394 395
	p += size;
	break;
    default:
396
	bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
397 398 399
    small_bitmap:
	while (size > 0) {
	    if ((bitmap & 1) == 0) {
400
		thread((StgClosure **)p);
401 402 403 404 405 406 407
	    }
	    p++;
	    bitmap = bitmap >> 1;
	    size--;
	}
	break;
    }
408

409 410 411 412 413 414 415 416
    return p;
}

STATIC_INLINE StgPtr
thread_PAP (StgPAP *pap)
{
    StgPtr p;
    p = thread_PAP_payload(pap->fun, pap->payload, pap->n_args);
417
    thread(&pap->fun);
418 419
    return p;
}
420 421 422 423 424 425
    
STATIC_INLINE StgPtr
thread_AP (StgAP *ap)
{
    StgPtr p;
    p = thread_PAP_payload(ap->fun, ap->payload, ap->n_args);
426
    thread(&ap->fun);
427 428
    return p;
}    
429

sof's avatar
sof committed
430
STATIC_INLINE StgPtr
431 432
thread_AP_STACK (StgAP_STACK *ap)
{
433
    thread(&ap->fun);
434 435 436 437 438 439 440
    thread_stack((P_)ap->payload, (P_)ap->payload + ap->size);
    return (P_)ap + sizeofW(StgAP_STACK) + ap->size;
}

static StgPtr
thread_TSO (StgTSO *tso)
{
441
    thread_(&tso->_link);
442
    thread_(&tso->global_link);
443 444 445

    if (   tso->why_blocked == BlockedOnMVar
	|| tso->why_blocked == BlockedOnBlackHole
446
	|| tso->why_blocked == BlockedOnMsgThrowTo
447 448
        || tso->why_blocked == NotBlocked
        ) {
449
	thread_(&tso->block_info.closure);
450
    }
451
    thread_(&tso->blocked_exceptions);
452
    thread_(&tso->bq);
453
    
454
    thread_(&tso->trec);
455

456 457
    thread_(&tso->stackobj);
    return (StgPtr)tso + sizeofW(StgTSO);
458 459 460
}


461 462 463 464 465 466 467 468
static void
update_fwd_large( bdescr *bd )
{
  StgPtr p;
  const StgInfoTable* info;

  for (; bd != NULL; bd = bd->link) {

Simon Marlow's avatar
Simon Marlow committed
469 470 471 472
    // nothing to do in a pinned block; it might not even have an object
    // at the beginning.
    if (bd->flags & BF_PINNED) continue;

473 474 475 476 477 478 479 480 481
    p = bd->start;
    info  = get_itbl((StgClosure *)p);

    switch (info->type) {

    case ARR_WORDS:
      // nothing to follow 
      continue;

482 483
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
484
    case MUT_ARR_PTRS_FROZEN:
485
    case MUT_ARR_PTRS_FROZEN0:
486 487
      // follow everything 
      {
488
          StgMutArrPtrs *a;
489

490 491 492 493 494
          a = (StgMutArrPtrs*)p;
          for (p = (P_)a->payload; p < (P_)&a->payload[a->ptrs]; p++) {
              thread((StgClosure **)p);
          }
          continue;
495 496
      }

497 498 499 500 501 502
    case STACK:
    {
        StgStack *stack = (StgStack*)p;
        thread_stack(stack->sp, stack->stack + stack->stack_size);
        continue;
    }
503 504 505

    case AP_STACK:
	thread_AP_STACK((StgAP_STACK *)p);
506 507 508
	continue;

    case PAP:
509
	thread_PAP((StgPAP *)p);
510 511
	continue;

512 513 514 515 516
    case TREC_CHUNK:
    {
        StgWord i;
        StgTRecChunk *tc = (StgTRecChunk *)p;
	TRecEntry *e = &(tc -> entries[0]);
517
	thread_(&tc->prev_chunk);
518
	for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
519 520 521
	  thread_(&e->tvar);
	  thread(&e->expected_value);
	  thread(&e->new_value);
522 523 524 525
	}
	continue;
    }

526 527 528 529 530 531
    default:
      barf("update_fwd_large: unknown/strange object  %d", (int)(info->type));
    }
  }
}

532 533
// ToDo: too big to inline
static /* STATIC_INLINE */ StgPtr
534 535 536
thread_obj (StgInfoTable *info, StgPtr p)
{
    switch (info->type) {
537 538 539
    case THUNK_0_1:
	return p + sizeofW(StgThunk) + 1;

540 541 542 543 544 545
    case FUN_0_1:
    case CONSTR_0_1:
	return p + sizeofW(StgHeader) + 1;
	
    case FUN_1_0:
    case CONSTR_1_0:
546
	thread(&((StgClosure *)p)->payload[0]);
547 548 549
	return p + sizeofW(StgHeader) + 1;
	
    case THUNK_1_0:
550
	thread(&((StgThunk *)p)->payload[0]);
551
	return p + sizeofW(StgThunk) + 1;
552 553
	
    case THUNK_0_2:
554 555
	return p + sizeofW(StgThunk) + 2;

556 557 558 559 560
    case FUN_0_2:
    case CONSTR_0_2:
	return p + sizeofW(StgHeader) + 2;
	
    case THUNK_1_1:
561
	thread(&((StgThunk *)p)->payload[0]);
562 563
	return p + sizeofW(StgThunk) + 2;

564 565
    case FUN_1_1:
    case CONSTR_1_1:
566
	thread(&((StgClosure *)p)->payload[0]);
567 568 569
	return p + sizeofW(StgHeader) + 2;
	
    case THUNK_2_0:
570 571
	thread(&((StgThunk *)p)->payload[0]);
	thread(&((StgThunk *)p)->payload[1]);
572 573
	return p + sizeofW(StgThunk) + 2;

574 575
    case FUN_2_0:
    case CONSTR_2_0:
576 577
	thread(&((StgClosure *)p)->payload[0]);
	thread(&((StgClosure *)p)->payload[1]);
578 579
	return p + sizeofW(StgHeader) + 2;
	
580 581
    case BCO: {
	StgBCO *bco = (StgBCO *)p;
582 583 584
	thread_(&bco->instrs);
	thread_(&bco->literals);
	thread_(&bco->ptrs);
585 586 587
	return p + bco_sizeW(bco);
    }

588
    case THUNK:
589 590 591 592 593 594
    {
	StgPtr end;
	
	end = (P_)((StgThunk *)p)->payload + 
	    info->layout.payload.ptrs;
	for (p = (P_)((StgThunk *)p)->payload; p < end; p++) {
595
	    thread((StgClosure **)p);
596 597 598 599 600
	}
	return p + info->layout.payload.nptrs;
    }

    case FUN:
601
    case CONSTR:
602 603
    case PRIM:
    case MUT_PRIM:
604 605
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
606
    case TVAR:
607
    case BLACKHOLE:
608
    case BLOCKING_QUEUE:
609 610 611 612 613 614
    {
	StgPtr end;
	
	end = (P_)((StgClosure *)p)->payload + 
	    info->layout.payload.ptrs;
	for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
615
	    thread((StgClosure **)p);
616 617 618 619 620 621 622
	}
	return p + info->layout.payload.nptrs;
    }
    
    case WEAK:
    {
	StgWeak *w = (StgWeak *)p;
623
	thread(&w->cfinalizers);
624 625 626
	thread(&w->key);
	thread(&w->value);
	thread(&w->finalizer);
627
	if (w->link != NULL) {
628
	    thread_(&w->link);
629 630 631 632
	}
	return p + sizeofW(StgWeak);
    }
    
633 634
    case MVAR_CLEAN:
    case MVAR_DIRTY:
635 636
    { 
	StgMVar *mvar = (StgMVar *)p;
637 638 639
	thread_(&mvar->head);
	thread_(&mvar->tail);
	thread(&mvar->value);
640 641 642
	return p + sizeofW(StgMVar);
    }
    
643 644
    case IND:
    case IND_PERM:
645
	thread(&((StgInd *)p)->indirectee);
646
	return p + sizeofW(StgInd);
647 648 649 650

    case THUNK_SELECTOR:
    { 
	StgSelector *s = (StgSelector *)p;
651
	thread(&s->selectee);
652 653 654 655 656 657 658 659
	return p + THUNK_SELECTOR_sizeW();
    }
    
    case AP_STACK:
	return thread_AP_STACK((StgAP_STACK *)p);
	
    case PAP:
	return thread_PAP((StgPAP *)p);
660 661 662

    case AP:
	return thread_AP((StgAP *)p);
663 664 665 666
	
    case ARR_WORDS:
	return p + arr_words_sizeW((StgArrWords *)p);
	
667 668
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
669
    case MUT_ARR_PTRS_FROZEN:
670
    case MUT_ARR_PTRS_FROZEN0:
671 672
	// follow everything 
    {
673 674 675 676
        StgMutArrPtrs *a;

        a = (StgMutArrPtrs *)p;
	for (p = (P_)a->payload; p < (P_)&a->payload[a->ptrs]; p++) {
677
	    thread((StgClosure **)p);
678
	}
679 680

	return (StgPtr)a + mut_arr_ptrs_sizeW(a);
681 682 683 684 685
    }
    
    case TSO:
	return thread_TSO((StgTSO *)p);
    
686 687 688 689 690 691 692
    case STACK:
    {
        StgStack *stack = (StgStack*)p;
        thread_stack(stack->sp, stack->stack + stack->stack_size);
        return p + stack_sizeW(stack);
    }

693 694
    case TREC_CHUNK:
    {
695
        StgWord i;
696 697
        StgTRecChunk *tc = (StgTRecChunk *)p;
	TRecEntry *e = &(tc -> entries[0]);
698
	thread_(&tc->prev_chunk);
699
	for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
700 701 702
	  thread_(&e->tvar);
	  thread(&e->expected_value);
	  thread(&e->new_value);
703 704 705 706
	}
	return p + sizeofW(StgTRecChunk);
    }

707 708
    default:
	barf("update_fwd: unknown/strange object  %d", (int)(info->type));
sof's avatar
sof committed
709
	return NULL;
710 711 712
    }
}

713 714 715 716 717 718 719 720 721 722 723 724 725 726 727
static void
update_fwd( bdescr *blocks )
{
    StgPtr p;
    bdescr *bd;
    StgInfoTable *info;

    bd = blocks;

    // cycle through all the blocks in the step
    for (; bd != NULL; bd = bd->link) {
	p = bd->start;

	// linearly scan the objects in this block
	while (p < bd->free) {
728
	    ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
729
	    info = get_itbl((StgClosure *)p);
730
	    p = thread_obj(info, p);
731 732 733 734 735 736 737 738
	}
    }
} 

static void
update_fwd_compact( bdescr *blocks )
{
    StgPtr p, q, free;
739
#if 0
740
    StgWord m;
741
#endif
742 743 744
    bdescr *bd, *free_bd;
    StgInfoTable *info;
    nat size;
745
    StgWord iptr;
746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786

    bd = blocks;
    free_bd = blocks;
    free = free_bd->start;

    // cycle through all the blocks in the step
    for (; bd != NULL; bd = bd->link) {
	p = bd->start;

	while (p < bd->free ) {

	    while ( p < bd->free && !is_marked(p,bd) ) {
		p++;
	    }
	    if (p >= bd->free) {
		break;
	    }

#if 0
    next:
	m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
	m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));

	while ( p < bd->free ) {

	    if ((m & 1) == 0) {
		m >>= 1;
		p++;
		if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
		    goto next;
		} else {
		    continue;
		}
	    }
#endif

	    // Problem: we need to know the destination for this cell
	    // in order to unthread its info pointer.  But we can't
	    // know the destination without the size, because we may
	    // spill into the next block.  So we have to run down the 
	    // threaded list and get the info ptr first.
787 788 789 790
            //
            // ToDo: one possible avenue of attack is to use the fact
            // that if (p&BLOCK_MASK) >= (free&BLOCK_MASK), then we
            // definitely have enough room.  Also see bug #1147.
791
            iptr = get_threaded_info(p);
792
	    info = INFO_PTR_TO_STRUCT((StgInfoTable *)UNTAG_CLOSURE((StgClosure *)iptr));
793 794 795

	    q = p;

796
	    p = thread_obj(info, p);
797 798 799

	    size = p - q;
	    if (free + size > free_bd->start + BLOCK_SIZE_W) {
800
		// set the next bit in the bitmap to indicate that
801 802 803
		// this object needs to be pushed into the next
		// block.  This saves us having to run down the
		// threaded info pointer list twice during the next pass.
804
		mark(q+1,bd);
805 806
		free_bd = free_bd->link;
		free = free_bd->start;
807
	    } else {
808
		ASSERT(!is_marked(q+1,bd));
809 810
	    }

811
	    unthread(q,(StgWord)free + GET_CLOSURE_TAG((StgClosure *)iptr));
812 813 814 815 816 817 818 819
	    free += size;
#if 0
	    goto next;
#endif
	}
    }
}

820
static W_
821
update_bkwd_compact( generation *gen )
822 823
{
    StgPtr p, free;
824
#if 0
825
    StgWord m;
826
#endif
827 828
    bdescr *bd, *free_bd;
    StgInfoTable *info;
829
    W_ size, free_blocks;
830
    StgWord iptr;
831

832
    bd = free_bd = gen->old_blocks;
833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866
    free = free_bd->start;
    free_blocks = 1;

    // cycle through all the blocks in the step
    for (; bd != NULL; bd = bd->link) {
	p = bd->start;

	while (p < bd->free ) {

	    while ( p < bd->free && !is_marked(p,bd) ) {
		p++;
	    }
	    if (p >= bd->free) {
		break;
	    }

#if 0
    next:
	m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
	m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));

	while ( p < bd->free ) {

	    if ((m & 1) == 0) {
		m >>= 1;
		p++;
		if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
		    goto next;
		} else {
		    continue;
		}
	    }
#endif

867
	    if (is_marked(p+1,bd)) {
868 869 870 871 872 873 874
		// don't forget to update the free ptr in the block desc.
		free_bd->free = free;
		free_bd = free_bd->link;
		free = free_bd->start;
		free_blocks++;
	    }

875 876
            iptr = get_threaded_info(p);
	    unthread(p, (StgWord)free + GET_CLOSURE_TAG((StgClosure *)iptr));
Simon Marlow's avatar
Simon Marlow committed
877
	    ASSERT(LOOKS_LIKE_INFO_PTR((StgWord)((StgClosure *)p)->header.info));
878
	    info = get_itbl((StgClosure *)p);
879
	    size = closure_sizeW_((StgClosure *)p,info);
880

881 882 883
	    if (free != p) {
		move(free,p,size);
	    }
884 885

	    // relocate TSOs
886 887
            if (info->type == STACK) {
                move_STACK((StgStack *)p, (StgStack *)free);
888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905
	    }

	    free += size;
	    p += size;
#if 0
	    goto next;
#endif
	}
    }

    // free the remaining blocks and count what's left.
    free_bd->free = free;
    if (free_bd->link != NULL) {
	freeChain(free_bd->link);
	free_bd->link = NULL;
    }

    return free_blocks;
906 907
}

908
void
909
compact(StgClosure *static_objects)
910
{
911
    W_ n, g, blocks;
912
    generation *gen;
913 914

    // 1. thread the roots
915
    markCapabilities((evac_fn)thread_root, NULL);
916

Simon Marlow's avatar
Simon Marlow committed
917 918
    markScheduler((evac_fn)thread_root, NULL);

919
    // the weak pointer lists...
920
    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
921
        if (generations[g].weak_ptr_list != NULL) {
922
            thread((void *)&generations[g].weak_ptr_list);
923
        }
924
    }
925 926

    if (dead_weak_ptr_list != NULL) {
927
        thread((void *)&dead_weak_ptr_list); // tmp
928 929
    }

930 931
    // mutable lists
    for (g = 1; g < RtsFlags.GcFlags.generations; g++) {
932 933
        bdescr *bd;
        StgPtr p;
934 935 936 937 938 939 940 941
        for (n = 0; n < n_capabilities; n++) {
            for (bd = capabilities[n].mut_lists[g]; 
                 bd != NULL; bd = bd->link) {
                for (p = bd->start; p < bd->free; p++) {
                    thread((StgClosure **)p);
                }
            }
        }
942
    }
943 944

    // the global thread list
945 946
    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
        thread((void *)&generations[g].threads);
947
    }
948

949
    // any threads resurrected during this GC
950
    thread((void *)&resurrected_threads);
951

952
    // the task list
953
    {
954
	Task *task;
955
        InCall *incall;
956
        for (task = all_tasks; task != NULL; task = task->all_next) {
957 958 959 960 961 962
            for (incall = task->incall; incall != NULL; 
                 incall = incall->prev_stack) {
                if (incall->tso) {
                    thread_(&incall->tso);
                }
            }
963 964 965
	}
    }

966
    // the static objects
967
    thread_static(static_objects /* ToDo: ok? */);
968 969

    // the stable pointer table
970
    threadStableTables((evac_fn)thread_root, NULL);
971

972
    // the CAF list (used by GHCi)
973
    markCAFs((evac_fn)thread_root, NULL);
974

975 976
    // 2. update forward ptrs
    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
977 978 979 980
        gen = &generations[g];
        debugTrace(DEBUG_gc, "update_fwd:  %d", g);

        update_fwd(gen->blocks);
Simon Marlow's avatar
Simon Marlow committed
981 982 983 984
        for (n = 0; n < n_capabilities; n++) {
            update_fwd(gc_threads[n]->gens[g].todo_bd);
            update_fwd(gc_threads[n]->gens[g].part_list);
        }
985 986 987 988
        update_fwd_large(gen->scavenged_large_objects);
        if (g == RtsFlags.GcFlags.generations-1 && gen->old_blocks != NULL) {
            debugTrace(DEBUG_gc, "update_fwd:  %d (compact)", g);
            update_fwd_compact(gen->old_blocks);
989 990 991 992
	}
    }

    // 3. update backward ptrs
993 994 995
    gen = oldest_gen;
    if (gen->old_blocks != NULL) {
	blocks = update_bkwd_compact(gen);
Simon Marlow's avatar
Simon Marlow committed
996
	debugTrace(DEBUG_gc, 
997 998 999
		   "update_bkwd: %d (compact, old: %d blocks, now %d blocks)",
		   gen->no, gen->n_old_blocks, blocks);
	gen->n_old_blocks = blocks;
1000 1001
    }
}