RetainerProfile.c 72.4 KB
Newer Older
1 2 3 4 5 6 7 8 9
/* -----------------------------------------------------------------------------
 *
 * (c) The GHC Team, 2001
 * Author: Sungwoo Park
 *
 * Retainer profiling.
 *
 * ---------------------------------------------------------------------------*/

Ben Gamari's avatar
Ben Gamari committed
10
#if defined(PROFILING)
11

12
// Turn off inlining when debugging - it obfuscates things
Ben Gamari's avatar
Ben Gamari committed
13
#if defined(DEBUG)
14 15 16 17 18
#define INLINE
#else
#define INLINE inline
#endif

Simon Marlow's avatar
Simon Marlow committed
19
#include "PosixSource.h"
20
#include "Rts.h"
Simon Marlow's avatar
Simon Marlow committed
21

22 23 24 25 26 27
#include "RtsUtils.h"
#include "RetainerProfile.h"
#include "RetainerSet.h"
#include "Schedule.h"
#include "Printer.h"
#include "Weak.h"
Simon Marlow's avatar
Simon Marlow committed
28
#include "sm/Sanity.h"
29 30 31
#include "Profiling.h"
#include "Stats.h"
#include "ProfHeap.h"
32
#include "Apply.h"
33
#include "Stable.h" /* markStableTables */
Simon Marlow's avatar
Simon Marlow committed
34
#include "sm/Storage.h" // for END_OF_STATIC_LIST
35

36 37 38 39 40 41 42 43 44 45 46 47
/* Note [What is a retainer?]
   ~~~~~~~~~~~~~~~~~~~~~~~~~~
The definition of what sorts of things are counted as retainers is a bit hard to
pin down. Intuitively, we want to identify closures which will help the user
identify memory leaks due to thunks. In practice we also end up lumping mutable
objects in this group for reasons that have been lost to time.

The definition of retainer is implemented in isRetainer(), defined later in this
file.
*/


48 49 50 51 52 53 54 55 56 57 58 59 60
/*
  Note: what to change in order to plug-in a new retainer profiling scheme?
    (1) type retainer in ../includes/StgRetainerProf.h
    (2) retainer function R(), i.e., getRetainerFrom()
    (3) the two hashing functions, hashKeySingleton() and hashKeyAddElement(),
        in RetainerSet.h, if needed.
    (4) printRetainer() and printRetainerSetShort() in RetainerSet.c.
 */

/* -----------------------------------------------------------------------------
 * Declarations...
 * -------------------------------------------------------------------------- */

61
static uint32_t retainerGeneration;  // generation
62

63 64 65
static uint32_t numObjectVisited;    // total number of objects visited
static uint32_t timesAnyObjectVisited;  // number of times any objects are
                                        // visited
66 67 68 69 70 71 72 73 74

/*
  The rs field in the profile header of any object points to its retainer
  set in an indirect way: if flip is 0, it points to the retainer set;
  if flip is 1, it points to the next byte after the retainer set (even
  for NULL pointers). Therefore, with flip 1, (rs ^ 1) is the actual
  pointer. See retainerSetOf().
 */

75
StgWord flip = 0;     // flip bit
76 77 78 79 80
                      // must be 0 if DEBUG_RETAINER is on (for static closures)

#define setRetainerSetToNull(c)   \
  (c)->header.prof.hp.rs = (RetainerSet *)((StgWord)NULL | flip)

81
static void retainStack(StgClosure *, retainer, StgPtr, StgPtr);
82
static void retainClosure(StgClosure *, StgClosure *, retainer);
Ben Gamari's avatar
Ben Gamari committed
83
#if defined(DEBUG_RETAINER)
84 85 86
static void belongToHeap(StgPtr p);
#endif

Ben Gamari's avatar
Ben Gamari committed
87
#if defined(DEBUG_RETAINER)
88 89 90 91 92 93 94
/*
  cStackSize records how many times retainStack() has been invoked recursively,
  that is, the number of activation records for retainStack() on the C stack.
  maxCStackSize records its max value.
  Invariants:
    cStackSize <= maxCStackSize
 */
95
static uint32_t cStackSize, maxCStackSize;
96

97
static uint32_t sumOfNewCost;        // sum of the cost of each object, computed
98
                                // when the object is first visited
99
static uint32_t sumOfNewCostExtra;   // for those objects not visited during
100
                                // retainer profiling, e.g., MUT_VAR
101
static uint32_t costArray[N_CLOSURE_TYPES];
102

103
uint32_t sumOfCostLinear;            // sum of the costs of all object, computed
104 105
                                // when linearly traversing the heap after
                                // retainer profiling
106
uint32_t costArrayLinear[N_CLOSURE_TYPES];
107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
#endif

/* -----------------------------------------------------------------------------
 * Retainer stack - header
 *   Note:
 *     Although the retainer stack implementation could be separated *
 *     from the retainer profiling engine, there does not seem to be
 *     any advantage in doing that; retainer stack is an integral part
 *     of retainer profiling engine and cannot be use elsewhere at
 *     all.
 * -------------------------------------------------------------------------- */

typedef enum {
    posTypeStep,
    posTypePtrs,
    posTypeSRT,
123
    posTypeLargeSRT,
124 125 126 127 128 129 130 131
} nextPosType;

typedef union {
    // fixed layout or layout specified by a field in the closure
    StgWord step;

    // layout.payload
    struct {
132 133 134
        // See StgClosureInfo in InfoTables.h
        StgHalfWord pos;
        StgHalfWord ptrs;
135
        StgPtr payload;
136 137 138 139
    } ptrs;

    // SRT
    struct {
140 141
        StgClosure **srt;
        StgWord    srt_bitmap;
142
    } srt;
143 144 145

    // Large SRT
    struct {
146 147
        StgLargeSRT *srt;
        StgWord offset;
148
    } large_srt;
149

150 151 152 153 154 155 156 157 158
} nextPos;

typedef struct {
    nextPosType type;
    nextPos next;
} stackPos;

typedef struct {
    StgClosure *c;
159
    retainer c_child_r;
160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179
    stackPos info;
} stackElement;

/*
  Invariants:
    firstStack points to the first block group.
    currentStack points to the block group currently being used.
    currentStack->free == stackLimit.
    stackTop points to the topmost byte in the stack of currentStack.
    Unless the whole stack is empty, stackTop must point to the topmost
    object (or byte) in the whole stack. Thus, it is only when the whole stack
    is empty that stackTop == stackLimit (not during the execution of push()
    and pop()).
    stackBottom == currentStack->start.
    stackLimit == currentStack->start + BLOCK_SIZE_W * currentStack->blocks.
  Note:
    When a current stack becomes empty, stackTop is set to point to
    the topmost element on the previous block group so as to satisfy
    the invariants described above.
 */
sof's avatar
sof committed
180
static bdescr *firstStack = NULL;
181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203
static bdescr *currentStack;
static stackElement *stackBottom, *stackTop, *stackLimit;

/*
  currentStackBoundary is used to mark the current stack chunk.
  If stackTop == currentStackBoundary, it means that the current stack chunk
  is empty. It is the responsibility of the user to keep currentStackBoundary
  valid all the time if it is to be employed.
 */
static stackElement *currentStackBoundary;

/*
  stackSize records the current size of the stack.
  maxStackSize records its high water mark.
  Invariants:
    stackSize <= maxStackSize
  Note:
    stackSize is just an estimate measure of the depth of the graph. The reason
    is that some heap objects have only a single child and may not result
    in a new element being pushed onto the stack. Therefore, at the end of
    retainer profiling, maxStackSize + maxCStackSize is some value no greater
    than the actual depth of the graph.
 */
Ben Gamari's avatar
Ben Gamari committed
204
#if defined(DEBUG_RETAINER)
205 206 207 208 209 210 211 212 213 214 215
static int stackSize, maxStackSize;
#endif

// number of blocks allocated for one stack
#define BLOCKS_IN_STACK 1

/* -----------------------------------------------------------------------------
 * Add a new block group to the stack.
 * Invariants:
 *  currentStack->link == s.
 * -------------------------------------------------------------------------- */
216
static INLINE void
217 218 219 220 221 222 223 224 225 226 227 228 229 230
newStackBlock( bdescr *bd )
{
    currentStack = bd;
    stackTop     = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
    stackBottom  = (stackElement *)bd->start;
    stackLimit   = (stackElement *)stackTop;
    bd->free     = (StgPtr)stackLimit;
}

/* -----------------------------------------------------------------------------
 * Return to the previous block group.
 * Invariants:
 *   s->link == currentStack.
 * -------------------------------------------------------------------------- */
231
static INLINE void
232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247
returnToOldStack( bdescr *bd )
{
    currentStack = bd;
    stackTop = (stackElement *)bd->free;
    stackBottom = (stackElement *)bd->start;
    stackLimit = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
    bd->free = (StgPtr)stackLimit;
}

/* -----------------------------------------------------------------------------
 *  Initializes the traverse stack.
 * -------------------------------------------------------------------------- */
static void
initializeTraverseStack( void )
{
    if (firstStack != NULL) {
248
        freeChain(firstStack);
249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270
    }

    firstStack = allocGroup(BLOCKS_IN_STACK);
    firstStack->link = NULL;
    firstStack->u.back = NULL;

    newStackBlock(firstStack);
}

/* -----------------------------------------------------------------------------
 * Frees all the block groups in the traverse stack.
 * Invariants:
 *   firstStack != NULL
 * -------------------------------------------------------------------------- */
static void
closeTraverseStack( void )
{
    freeChain(firstStack);
    firstStack = NULL;
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
271
 * Returns true if the whole stack is empty.
272
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
273
static INLINE bool
274 275 276 277 278
isEmptyRetainerStack( void )
{
    return (firstStack == currentStack) && stackTop == stackLimit;
}

sof's avatar
sof committed
279 280 281
/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
282
W_
283
retainerStackBlocks( void )
sof's avatar
sof committed
284 285
{
    bdescr* bd;
286
    W_ res = 0;
sof's avatar
sof committed
287

288
    for (bd = firstStack; bd != NULL; bd = bd->link)
sof's avatar
sof committed
289 290 291 292 293
      res += bd->blocks;

    return res;
}

294
/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
295
 * Returns true if stackTop is at the stack boundary of the current stack,
296 297
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
298
static INLINE bool
299 300 301 302 303 304 305 306 307 308
isOnBoundary( void )
{
    return stackTop == currentStackBoundary;
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
309
static INLINE void
310
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
311 312 313 314 315 316 317 318 319 320
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
321
static INLINE StgClosure *
322 323 324
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
325
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
326
    } else {
327
        return NULL;
328 329 330 331 332 333
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
334
static INLINE void
335
init_srt_fun( stackPos *info, const StgFunInfoTable *infoTable )
336
{
337
    if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
338 339 340
        info->type = posTypeLargeSRT;
        info->next.large_srt.srt = (StgLargeSRT *)GET_FUN_SRT(infoTable);
        info->next.large_srt.offset = 0;
341
    } else {
342 343 344
        info->type = posTypeSRT;
        info->next.srt.srt = (StgClosure **)GET_FUN_SRT(infoTable);
        info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
345
    }
346 347
}

348
static INLINE void
349
init_srt_thunk( stackPos *info, const StgThunkInfoTable *infoTable )
350
{
351
    if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
352 353 354
        info->type = posTypeLargeSRT;
        info->next.large_srt.srt = (StgLargeSRT *)GET_SRT(infoTable);
        info->next.large_srt.offset = 0;
355
    } else {
356 357 358
        info->type = posTypeSRT;
        info->next.srt.srt = (StgClosure **)GET_SRT(infoTable);
        info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
359
    }
360 361 362 363 364
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
365
static INLINE StgClosure *
366 367 368
find_srt( stackPos *info )
{
    StgClosure *c;
369
    StgWord bitmap;
370

371
    if (info->type == posTypeSRT) {
372 373 374 375
        // Small SRT bitmap
        bitmap = info->next.srt.srt_bitmap;
        while (bitmap != 0) {
            if ((bitmap & 1) != 0) {
376
#if defined(COMPILING_WINDOWS_DLL)
377 378 379 380
                if ((unsigned long)(*(info->next.srt.srt)) & 0x1)
                    c = (* (StgClosure **)((unsigned long)*(info->next.srt.srt)) & ~0x1);
                else
                    c = *(info->next.srt.srt);
381
#else
382
                c = *(info->next.srt.srt);
383
#endif
384 385 386 387 388 389 390 391 392 393
                bitmap = bitmap >> 1;
                info->next.srt.srt++;
                info->next.srt.srt_bitmap = bitmap;
                return c;
            }
            bitmap = bitmap >> 1;
            info->next.srt.srt++;
        }
        // bitmap is now zero...
        return NULL;
394 395
    }
    else {
396
        // Large SRT bitmap
397
        uint32_t i = info->next.large_srt.offset;
398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
        StgWord bitmap;

        // Follow the pattern from GC.c:scavenge_large_srt_bitmap().
        bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
        bitmap = bitmap >> (i % BITS_IN(StgWord));
        while (i < info->next.large_srt.srt->l.size) {
            if ((bitmap & 1) != 0) {
                c = ((StgClosure **)info->next.large_srt.srt->srt)[i];
                i++;
                info->next.large_srt.offset = i;
                return c;
            }
            i++;
            if (i % BITS_IN(W_) == 0) {
                bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
            } else {
                bitmap = bitmap >> 1;
            }
        }
        // reached the end of this bitmap.
        info->next.large_srt.offset = i;
        return NULL;
420 421 422 423 424 425 426
    }
}

/* -----------------------------------------------------------------------------
 *  push() pushes a stackElement representing the next child of *c
 *  onto the traverse stack. If *c has no child, *first_child is set
 *  to NULL and nothing is pushed onto the stack. If *c has only one
427
 *  child, *c_child is set to that child and nothing is pushed onto
428 429 430 431 432 433
 *  the stack.  If *c has more than two children, *first_child is set
 *  to the first child and a stackElement representing the second
 *  child is pushed onto the stack.

 *  Invariants:
 *     *c_child_r is the most recent retainer of *c's children.
434
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
435 436 437
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
438
static INLINE void
439
push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
440 441 442 443
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
444
#if defined(DEBUG_RETAINER)
445
    // debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
446 447 448
#endif

    ASSERT(get_itbl(c)->type != TSO);
449
    ASSERT(get_itbl(c)->type != AP_STACK);
450 451 452 453 454 455 456 457 458 459

    //
    // fill in se
    //

    se.c = c;
    se.c_child_r = c_child_r;

    // fill in se.info
    switch (get_itbl(c)->type) {
460
        // no child, no SRT
461 462 463
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
gcampax's avatar
gcampax committed
464
    case COMPACT_NFDATA:
465 466
        *first_child = NULL;
        return;
467

468
        // one child (fixed), no SRT
469 470
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
471 472
        *first_child = ((StgMutVar *)c)->var;
        return;
473
    case THUNK_SELECTOR:
474 475
        *first_child = ((StgSelector *)c)->selectee;
        return;
476
    case BLACKHOLE:
477 478
        *first_child = ((StgInd *)c)->indirectee;
        return;
479 480
    case CONSTR_1_0:
    case CONSTR_1_1:
481 482
        *first_child = c->payload[0];
        return;
483

484 485 486
        // For CONSTR_2_0 and MVAR, we use se.info.step to record the position
        // of the next child. We do not write a separate initialization code.
        // Also we do not have to initialize info.type;
487

488 489
        // two children (fixed), no SRT
        // need to push a stackElement, but nothing to store in se.info
490
    case CONSTR_2_0:
491 492 493 494
        *first_child = c->payload[0];         // return the first pointer
        // se.info.type = posTypeStep;
        // se.info.next.step = 2;            // 2 = second
        break;
495

496 497
        // three children (fixed), no SRT
        // need to push a stackElement
498 499
    case MVAR_CLEAN:
    case MVAR_DIRTY:
500 501 502 503 504 505 506 507
        // head must be TSO and the head of a linked list of TSOs.
        // Shoule it be a child? Seems to be yes.
        *first_child = (StgClosure *)((StgMVar *)c)->head;
        // se.info.type = posTypeStep;
        se.info.next.step = 2;            // 2 = second
        break;

        // three children (fixed), no SRT
508
    case WEAK:
509 510 511 512
        *first_child = ((StgWeak *)c)->key;
        // se.info.type = posTypeStep;
        se.info.next.step = 2;
        break;
513

514
        // layout.payload.ptrs, no SRT
515
    case TVAR:
516
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
517
    case CONSTR_NOCAF:
518
    case PRIM:
519
    case MUT_PRIM:
520
    case BCO:
521 522 523 524 525 526 527 528
        init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs,
                  (StgPtr)c->payload);
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;   // no child
        break;

        // StgMutArrPtr.ptrs, no SRT
529 530
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
531
    case MUT_ARR_PTRS_FROZEN:
532
    case MUT_ARR_PTRS_FROZEN0:
533 534 535 536 537 538 539 540
        init_ptrs(&se.info, ((StgMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;

        // StgMutArrPtr.ptrs, no SRT
541 542 543 544
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case SMALL_MUT_ARR_PTRS_FROZEN:
    case SMALL_MUT_ARR_PTRS_FROZEN0:
545 546 547 548 549 550
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;
551

552 553 554
    // layout.payload.ptrs, SRT
    case FUN:           // *c is a heap object.
    case FUN_2_0:
555 556 557 558 559 560
        init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs, (StgPtr)c->payload);
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            // no child from ptrs, so check SRT
            goto fun_srt_only;
        break;
561

562 563
    case THUNK:
    case THUNK_2_0:
564 565 566 567 568 569 570 571 572
        init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs,
                  (StgPtr)((StgThunk *)c)->payload);
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            // no child from ptrs, so check SRT
            goto thunk_srt_only;
        break;

        // 1 fixed child, SRT
573 574
    case FUN_1_0:
    case FUN_1_1:
575 576 577 578
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;
579

580 581
    case THUNK_1_0:
    case THUNK_1_1:
582 583 584 585
        *first_child = ((StgThunk *)c)->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        break;
586 587

    case FUN_STATIC:      // *c is a heap object.
588
        ASSERT(get_itbl(c)->srt_bitmap != 0);
589 590
    case FUN_0_1:
    case FUN_0_2:
591 592
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
593 594 595 596
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;
597 598 599

    // SRT only
    case THUNK_STATIC:
600
        ASSERT(get_itbl(c)->srt_bitmap != 0);
601 602
    case THUNK_0_1:
    case THUNK_0_2:
603 604
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
605 606 607 608 609
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

610
    case TREC_CHUNK:
611 612 613
        *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
        se.info.next.step = 0;  // entry no.
        break;
614

615
        // cannot appear
616
    case PAP:
617 618
    case AP:
    case AP_STACK:
619
    case TSO:
620
    case STACK:
621
    case IND_STATIC:
622
        // stack objects
623 624
    case UPDATE_FRAME:
    case CATCH_FRAME:
625
    case UNDERFLOW_FRAME:
626 627 628 629
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
630
        // invalid objects
631 632 633
    case IND:
    case INVALID_OBJECT:
    default:
634
        barf("Invalid object *c in push(): %d", get_itbl(c)->type);
635
        return;
636 637 638
    }

    if (stackTop - 1 < stackBottom) {
Ben Gamari's avatar
Ben Gamari committed
639
#if defined(DEBUG_RETAINER)
640
        // debugBelch("push() to the next stack.\n");
641
#endif
642 643 644 645 646 647 648 649 650 651 652 653 654
        // currentStack->free is updated when the active stack is switched
        // to the next stack.
        currentStack->free = (StgPtr)stackTop;

        if (currentStack->link == NULL) {
            nbd = allocGroup(BLOCKS_IN_STACK);
            nbd->link = NULL;
            nbd->u.back = currentStack;
            currentStack->link = nbd;
        } else
            nbd = currentStack->link;

        newStackBlock(nbd);
655 656 657 658 659 660 661 662
    }

    // adjust stackTop (acutal push)
    stackTop--;
    // If the size of stackElement was huge, we would better replace the
    // following statement by either a memcpy() call or a switch statement
    // on the type of the element. Currently, the size of stackElement is
    // small enough (5 words) that this direct assignment seems to be enough.
663 664 665 666 667 668

    // ToDo: The line below leads to the warning:
    //    warning: 'se.info.type' may be used uninitialized in this function
    // This is caused by the fact that there are execution paths through the
    // large switch statement above where some cases do not initialize this
    // field. Is this really harmless? Can we avoid the warning?
669 670
    *stackTop = se;

Ben Gamari's avatar
Ben Gamari committed
671
#if defined(DEBUG_RETAINER)
672 673 674
    stackSize++;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
    // ASSERT(stackSize >= 0);
675
    // debugBelch("stackSize = %d\n", stackSize);
676 677 678 679 680 681 682 683 684 685 686 687 688
#endif
}

/* -----------------------------------------------------------------------------
 *  popOff() and popOffReal(): Pop a stackElement off the traverse stack.
 *  Invariants:
 *    stackTop cannot be equal to stackLimit unless the whole stack is
 *    empty, in which case popOff() is not allowed.
 *  Note:
 *    You can think of popOffReal() as a part of popOff() which is
 *    executed at the end of popOff() in necessary. Since popOff() is
 *    likely to be executed quite often while popOffReal() is not, we
 *    separate popOffReal() from popOff(), which is declared as an
689
 *    INLINE function (for the sake of execution speed).  popOffReal()
690 691 692 693 694 695 696
 *    is called only within popOff() and nowhere else.
 * -------------------------------------------------------------------------- */
static void
popOffReal(void)
{
    bdescr *pbd;    // Previous Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
697
#if defined(DEBUG_RETAINER)
698
    // debugBelch("pop() to the previous stack.\n");
699 700 701 702 703 704
#endif

    ASSERT(stackTop + 1 == stackLimit);
    ASSERT(stackBottom == (stackElement *)currentStack->start);

    if (firstStack == currentStack) {
705 706 707
        // The stack is completely empty.
        stackTop++;
        ASSERT(stackTop == stackLimit);
Ben Gamari's avatar
Ben Gamari committed
708
#if defined(DEBUG_RETAINER)
709 710 711 712 713 714
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
        /*
          ASSERT(stackSize >= 0);
          debugBelch("stackSize = %d\n", stackSize);
        */
715
#endif
716
        return;
717 718 719 720 721 722 723 724 725 726 727 728
    }

    // currentStack->free is updated when the active stack is switched back
    // to the previous stack.
    currentStack->free = (StgPtr)stackLimit;

    // find the previous block descriptor
    pbd = currentStack->u.back;
    ASSERT(pbd != NULL);

    returnToOldStack(pbd);

Ben Gamari's avatar
Ben Gamari committed
729
#if defined(DEBUG_RETAINER)
730 731 732 733
    stackSize--;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
    /*
      ASSERT(stackSize >= 0);
734
      debugBelch("stackSize = %d\n", stackSize);
735 736 737 738
    */
#endif
}

739
static INLINE void
740
popOff(void) {
Ben Gamari's avatar
Ben Gamari committed
741
#if defined(DEBUG_RETAINER)
742
    // debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
743 744 745 746 747 748 749
#endif

    ASSERT(stackTop != stackLimit);
    ASSERT(!isEmptyRetainerStack());

    // <= (instead of <) is wrong!
    if (stackTop + 1 < stackLimit) {
750
        stackTop++;
Ben Gamari's avatar
Ben Gamari committed
751
#if defined(DEBUG_RETAINER)
752 753 754 755 756 757
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
        /*
          ASSERT(stackSize >= 0);
          debugBelch("stackSize = %d\n", stackSize);
        */
758
#endif
759
        return;
760 761 762 763 764 765 766 767 768 769 770 771 772 773 774
    }

    popOffReal();
}

/* -----------------------------------------------------------------------------
 *  Finds the next object to be considered for retainer profiling and store
 *  its pointer to *c.
 *  Test if the topmost stack element indicates that more objects are left,
 *  and if so, retrieve the first object and store its pointer to *c. Also,
 *  set *cp and *r appropriately, both of which are stored in the stack element.
 *  The topmost stack element then is overwritten so as for it to now denote
 *  the next object.
 *  If the topmost stack element indicates no more objects are left, pop
 *  off the stack element until either an object can be retrieved or
Ben Gamari's avatar
Ben Gamari committed
775
 *  the current stack chunk becomes empty, indicated by true returned by
776 777 778 779 780
 *  isOnBoundary(), in which case *c is set to NULL.
 *  Note:
 *    It is okay to call this function even when the current stack chunk
 *    is empty.
 * -------------------------------------------------------------------------- */
781
static INLINE void
782
pop( StgClosure **c, StgClosure **cp, retainer *r )
783 784 785
{
    stackElement *se;

Ben Gamari's avatar
Ben Gamari committed
786
#if defined(DEBUG_RETAINER)
787
    // debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
788 789 790
#endif

    do {
791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809
        if (isOnBoundary()) {     // if the current stack chunk is depleted
            *c = NULL;
            return;
        }

        se = stackTop;

        switch (get_itbl(se->c)->type) {
            // two children (fixed), no SRT
            // nothing in se.info
        case CONSTR_2_0:
            *c = se->c->payload[1];
            *cp = se->c;
            *r = se->c_child_r;
            popOff();
            return;

            // three children (fixed), no SRT
            // need to push a stackElement
810 811
        case MVAR_CLEAN:
        case MVAR_DIRTY:
812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844
            if (se->info.next.step == 2) {
                *c = (StgClosure *)((StgMVar *)se->c)->tail;
                se->info.next.step++;             // move to the next step
                // no popOff
            } else {
                *c = ((StgMVar *)se->c)->value;
                popOff();
            }
            *cp = se->c;
            *r = se->c_child_r;
            return;

            // three children (fixed), no SRT
        case WEAK:
            if (se->info.next.step == 2) {
                *c = ((StgWeak *)se->c)->value;
                se->info.next.step++;
                // no popOff
            } else {
                *c = ((StgWeak *)se->c)->finalizer;
                popOff();
            }
            *cp = se->c;
            *r = se->c_child_r;
            return;

        case TREC_CHUNK: {
            // These are pretty complicated: we have N entries, each
            // of which contains 3 fields that we want to follow.  So
            // we divide the step counter: the 2 low bits indicate
            // which field, and the rest of the bits indicate the
            // entry number (starting from zero).
            TRecEntry *entry;
845 846
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864
            if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
                *c = NULL;
                popOff();
                return;
            }
            entry = &((StgTRecChunk *)se->c)->entries[entry_no];
            if (field_no == 0) {
                *c = (StgClosure *)entry->tvar;
            } else if (field_no == 1) {
                *c = entry->expected_value;
            } else {
                *c = entry->new_value;
            }
            *cp = se->c;
            *r = se->c_child_r;
            se->info.next.step++;
            return;
        }
865

866 867
        case TVAR:
        case CONSTR:
868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946
        case PRIM:
        case MUT_PRIM:
        case BCO:
            // StgMutArrPtr.ptrs, no SRT
        case MUT_ARR_PTRS_CLEAN:
        case MUT_ARR_PTRS_DIRTY:
        case MUT_ARR_PTRS_FROZEN:
        case MUT_ARR_PTRS_FROZEN0:
            *c = find_ptrs(&se->info);
            if (*c == NULL) {
                popOff();
                break;
            }
            *cp = se->c;
            *r = se->c_child_r;
            return;

            // layout.payload.ptrs, SRT
        case FUN:         // always a heap object
        case FUN_2_0:
            if (se->info.type == posTypePtrs) {
                *c = find_ptrs(&se->info);
                if (*c != NULL) {
                    *cp = se->c;
                    *r = se->c_child_r;
                    return;
                }
                init_srt_fun(&se->info, get_fun_itbl(se->c));
            }
            goto do_srt;

        case THUNK:
        case THUNK_2_0:
            if (se->info.type == posTypePtrs) {
                *c = find_ptrs(&se->info);
                if (*c != NULL) {
                    *cp = se->c;
                    *r = se->c_child_r;
                    return;
                }
                init_srt_thunk(&se->info, get_thunk_itbl(se->c));
            }
            goto do_srt;

            // SRT
        do_srt:
        case THUNK_STATIC:
        case FUN_STATIC:
        case FUN_0_1:
        case FUN_0_2:
        case THUNK_0_1:
        case THUNK_0_2:
        case FUN_1_0:
        case FUN_1_1:
        case THUNK_1_0:
        case THUNK_1_1:
            *c = find_srt(&se->info);
            if (*c != NULL) {
                *cp = se->c;
                *r = se->c_child_r;
                return;
            }
            popOff();
            break;

            // no child (fixed), no SRT
        case CONSTR_0_1:
        case CONSTR_0_2:
        case ARR_WORDS:
            // one child (fixed), no SRT
        case MUT_VAR_CLEAN:
        case MUT_VAR_DIRTY:
        case THUNK_SELECTOR:
        case CONSTR_1_1:
            // cannot appear
        case PAP:
        case AP:
        case AP_STACK:
        case TSO:
947 948
        case STACK:
        case IND_STATIC:
Simon Marlow's avatar
Simon Marlow committed
949
        case CONSTR_NOCAF:
950
            // stack objects
951
        case UPDATE_FRAME:
952
        case CATCH_FRAME:
953 954
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
955 956 957 958 959 960 961 962 963 964
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default:
            barf("Invalid object *c in pop()");
            return;
        }
Ben Gamari's avatar
Ben Gamari committed
965
    } while (true);
966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984
}

/* -----------------------------------------------------------------------------
 * RETAINER PROFILING ENGINE
 * -------------------------------------------------------------------------- */

void
initRetainerProfiling( void )
{
    initializeAllRetainerSet();
    retainerGeneration = 0;
}

/* -----------------------------------------------------------------------------
 *  This function must be called before f-closing prof_file.
 * -------------------------------------------------------------------------- */
void
endRetainerProfiling( void )
{
Ben Gamari's avatar
Ben Gamari committed
985
#if defined(SECOND_APPROACH)
986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003
    outputAllRetainerSet(prof_file);
#endif
}

/* -----------------------------------------------------------------------------
 *  Returns the actual pointer to the retainer set of the closure *c.
 *  It may adjust RSET(c) subject to flip.
 *  Side effects:
 *    RSET(c) is initialized to NULL if its current value does not
 *    conform to flip.
 *  Note:
 *    Even though this function has side effects, they CAN be ignored because
 *    subsequent calls to retainerSetOf() always result in the same return value
 *    and retainerSetOf() is the only way to retrieve retainerSet of a given
 *    closure.
 *    We have to perform an XOR (^) operation each time a closure is examined.
 *    The reason is that we do not know when a closure is visited last.
 * -------------------------------------------------------------------------- */
1004
static INLINE void
1005 1006 1007
maybeInitRetainerSet( StgClosure *c )
{
    if (!isRetainerSetFieldValid(c)) {
1008
        setRetainerSetToNull(c);
1009 1010 1011 1012
    }
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
1013
 * Returns true if *c is a retainer.
1014
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
1015
static INLINE bool
1016 1017 1018
isRetainer( StgClosure *c )
{
    switch (get_itbl(c)->type) {
1019 1020 1021 1022
        //
        //  True case
        //
        // TSOs MUST be retainers: they constitute the set of roots.
1023
    case TSO:
1024
    case STACK:
1025

1026
        // mutable objects
1027
    case MUT_PRIM:
1028 1029
    case MVAR_CLEAN:
    case MVAR_DIRTY:
1030
    case TVAR:
1031 1032
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
1033 1034
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
1035 1036 1037
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case BLOCKING_QUEUE:
1038

1039
        // thunks are retainers.
1040 1041 1042 1043 1044 1045 1046
    case THUNK:
    case THUNK_1_0:
    case THUNK_0_1:
    case THUNK_2_0:
    case THUNK_1_1:
    case THUNK_0_2:
    case THUNK_SELECTOR:
1047 1048
    case AP:
    case AP_STACK:
1049

1050
        // Static thunks, or CAFS, are obviously retainers.
1051 1052
    case THUNK_STATIC:

1053 1054
        // WEAK objects are roots; there is separate code in which traversing
        // begins from WEAK objects.
1055
    case WEAK:
Ben Gamari's avatar
Ben Gamari committed
1056
        return true;
1057

1058 1059 1060
        //
        // False case
        //
1061

1062
        // constructors
1063
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
1064
    case CONSTR_NOCAF:
1065 1066 1067 1068 1069
    case CONSTR_1_0:
    case CONSTR_0_1:
    case CONSTR_2_0:
    case CONSTR_1_1:
    case CONSTR_0_2:
1070
        // functions
1071 1072 1073 1074 1075 1076
    case FUN:
    case FUN_1_0:
    case FUN_0_1:
    case FUN_2_0:
    case FUN_1_1:
    case FUN_0_2:
1077
        // partial applications
1078
    case PAP:
1079
        // indirection
Ian Lynagh's avatar
Ian Lynagh committed
1080 1081 1082 1083
    // IND_STATIC used to be an error, but at the moment it can happen
    // as isAlive doesn't look through IND_STATIC as it ignores static
    // closures. See trac #3956 for a program that hit this error.
    case IND_STATIC:
1084
    case BLACKHOLE:
1085
    case WHITEHOLE:
1086
        // static objects
1087
    case FUN_STATIC:
1088
        // misc
1089
    case PRIM:
1090 1091
    case BCO:
    case ARR_WORDS:
1092
    case COMPACT_NFDATA:
1093
        // STM
1094
    case TREC_CHUNK:
1095 1096 1097
        // immutable arrays
    case MUT_ARR_PTRS_FROZEN:
    case MUT_ARR_PTRS_FROZEN0:
1098 1099
    case SMALL_MUT_ARR_PTRS_FROZEN:
    case SMALL_MUT_ARR_PTRS_FROZEN0:
Ben Gamari's avatar
Ben Gamari committed
1100
        return false;
1101

1102 1103 1104 1105 1106
        //
        // Error case
        //
        // Stack objects are invalid because they are never treated as
        // legal objects during retainer profiling.
1107 1108
    case UPDATE_FRAME:
    case CATCH_FRAME:
1109 1110
    case CATCH_RETRY_FRAME:
    case CATCH_STM_FRAME:
1111
    case UNDERFLOW_FRAME:
1112
    case ATOMICALLY_FRAME:
1113 1114 1115 1116
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
1117
    case RET_FUN:
1118
        // other cases
1119 1120 1121
    case IND:
    case INVALID_OBJECT:
    default:
1122
        barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
Ben Gamari's avatar
Ben Gamari committed
1123
        return false;
1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135
    }
}

/* -----------------------------------------------------------------------------
 *  Returns the retainer function value for the closure *c, i.e., R(*c).
 *  This function does NOT return the retainer(s) of *c.
 *  Invariants:
 *    *c must be a retainer.
 *  Note:
 *    Depending on the definition of this function, the maintenance of retainer
 *    sets can be made easier. If most retainer sets are likely to be created
 *    again across garbage collections, refreshAllRetainerSet() in
1136
 *    RetainerSet.c can simply do nothing.
1137 1138 1139 1140
 *    If this is not the case, we can free all the retainer sets and
 *    re-initialize the hash table.
 *    See refreshAllRetainerSet() in RetainerSet.c.
 * -------------------------------------------------------------------------- */
1141
static INLINE retainer
1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155
getRetainerFrom( StgClosure *c )
{
    ASSERT(isRetainer(c));

    return c->header.prof.ccs;
}

/* -----------------------------------------------------------------------------
 *  Associates the retainer set *s with the closure *c, that is, *s becomes
 *  the retainer set of *c.
 *  Invariants:
 *    c != NULL
 *    s != NULL
 * -------------------------------------------------------------------------- */
1156
static INLINE void
1157
associate( StgClosure *c, RetainerSet *s )
1158 1159 1160 1161 1162 1163
{
    // StgWord has the same size as pointers, so the following type
    // casting is okay.
    RSET(c) = (RetainerSet *)((StgWord)s | flip);
}

1164 1165 1166 1167 1168
/* -----------------------------------------------------------------------------
   Call retainClosure for each of the closures covered by a large bitmap.
   -------------------------------------------------------------------------- */

static void
1169
retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, uint32_t size,
1170
                     StgClosure *c, retainer c_child_r)
1171
{
1172
    uint32_t i, b;
1173
    StgWord bitmap;
1174

1175 1176 1177
    b = 0;
    bitmap = large_bitmap->bitmap[b];
    for (i = 0; i < size; ) {
1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188
        if ((bitmap & 1) == 0) {
            retainClosure((StgClosure *)*p, c, c_child_r);
        }
        i++;
        p++;
        if (i % BITS_IN(W_) == 0) {
            b++;
            bitmap = large_bitmap->bitmap[b];
        } else {
            bitmap = bitmap >> 1;
        }
1189 1190 1191
    }
}

1192
static INLINE StgPtr
1193
retain_small_bitmap (StgPtr p, uint32_t size, StgWord bitmap,
1194
                     StgClosure *c, retainer c_child_r)
1195 1196
{
    while (size > 0) {
1197 1198 1199 1200 1201 1202
        if ((bitmap & 1) == 0) {
            retainClosure((StgClosure *)*p, c, c_child_r);
        }
        p++;
        bitmap = bitmap >> 1;
        size--;
1203 1204 1205 1206
    }
    return p;
}

1207 1208 1209 1210 1211 1212 1213
/* -----------------------------------------------------------------------------
 * Call retainClosure for each of the closures in an SRT.
 * ------------------------------------------------------------------------- */

static void
retain_large_srt_bitmap (StgLargeSRT *srt, StgClosure *c, retainer c_child_r)
{
1214
    uint32_t i, b, size;
1215 1216
    StgWord bitmap;
    StgClosure **p;
1217

1218 1219 1220 1221 1222
    b = 0;
    p = (StgClosure **)srt->srt;
    size   = srt->l.size;
    bitmap = srt->l.bitmap[b];
    for (i = 0; i < size; ) {
1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233
        if ((bitmap & 1) != 0) {
            retainClosure((StgClosure *)*p, c, c_child_r);
        }
        i++;
        p++;
        if (i % BITS_IN(W_) == 0) {
            b++;
            bitmap = srt->l.bitmap[b];
        } else {
            bitmap = bitmap >> 1;
        }
1234 1235 1236 1237
    }
}

static INLINE void
1238 1239
retainSRT (StgClosure **srt, uint32_t srt_bitmap, StgClosure *c,
            retainer c_child_r)
1240
{
1241
  uint32_t bitmap;
1242 1243 1244 1245 1246
  StgClosure **p;

  bitmap = srt_bitmap;
  p = srt;

1247
  if (bitmap == (StgHalfWord)(-1)) {
1248 1249 1250 1251 1252 1253
      retain_large_srt_bitmap( (StgLargeSRT *)srt, c, c_child_r );
      return;
  }

  while (bitmap != 0) {
      if ((bitmap & 1) != 0) {
1254
#if defined(COMPILING_WINDOWS_DLL)
1255 1256 1257 1258 1259 1260
          if ( (unsigned long)(*srt) & 0x1 ) {
              retainClosure(* (StgClosure**) ((unsigned long) (*srt) & ~0x1),
                            c, c_child_r);
          } else {
              retainClosure(*srt,c,c_child_r);
          }
1261
#else
1262
          retainClosure(*srt,c,c_child_r);
1263 1264 1265 1266 1267 1268 1269
#endif
      }
      p++;
      bitmap = bitmap >> 1;
  }
}

1270 1271 1272 1273 1274 1275
/* -----------------------------------------------------------------------------
 *  Process all the objects in the stack chunk from stackStart to stackEnd
 *  with *c and *c_child_r being their parent and their most recent retainer,
 *  respectively. Treat stackOptionalFun as another child of *c if it is
 *  not NULL.
 *  Invariants:
1276
 *    *c is one of the following: TSO, AP_STACK.
1277 1278 1279 1280 1281
 *    If *c is TSO, c == c_child_r.
 *    stackStart < stackEnd.
 *    RSET(c) and RSET(c_child_r) are valid, i.e., their
 *    interpretation conforms to the current value of flip (even when they
 *    are interpreted to be NULL).
1282
 *    If *c is TSO, its state is not ThreadComplete,or ThreadKilled,
1283
 *    which means that its stack is ready to process.
1284 1285 1286 1287 1288
 *  Note:
 *    This code was almost plagiarzied from GC.c! For each pointer,
 *    retainClosure() is invoked instead of evacuate().
 * -------------------------------------------------------------------------- */
static void
1289
retainStack( StgClosure *c, retainer c_child_r,
1290
             StgPtr stackStart, StgPtr stackEnd )
1291 1292
{
    stackElement *oldStackBoundary;
1293
    StgPtr p;
1294
    const StgRetInfoTable *info;
1295
    StgWord bitmap;
1296
    uint32_t size;
1297

Ben Gamari's avatar
Ben Gamari committed
1298
#if defined(DEBUG_RETAINER)
1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311
    cStackSize++;
    if (cStackSize > maxCStackSize) maxCStackSize = cStackSize;
#endif

    /*
      Each invocation of retainStack() creates a new virtual
      stack. Since all such stacks share a single common stack, we
      record the current currentStackBoundary, which will be restored
      at the exit.
    */
    oldStackBoundary = currentStackBoundary;
    currentStackBoundary = stackTop;

Ben Gamari's avatar
Ben Gamari committed
1312
#if defined(DEBUG_RETAINER)
1313
    // debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n", oldStackBoundary, currentStackBoundary);
1314 1315
#endif

1316 1317
    ASSERT(get_itbl(c)->type == STACK);

1318 1319
    p = stackStart;
    while (p < stackEnd) {
1320
        info = get_ret_itbl((StgClosure *)p);
1321

1322
        switch(info->i.type) {
1323

1324 1325 1326 1327
        case UPDATE_FRAME:
            retainClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
            p += sizeofW(StgUpdateFrame);
            continue;
1328

1329 1330
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366
        case CATCH_FRAME:
        case CATCH_STM_FRAME:
        case CATCH_RETRY_FRAME:
        case ATOMICALLY_FRAME:
        case RET_SMALL:
            bitmap = BITMAP_BITS(info->i.layout.bitmap);
            size   = BITMAP_SIZE(info->i.layout.bitmap);
            p++;
            p = retain_small_bitmap(p, size, bitmap, c, c_child_r);

        follow_srt:
            retainSRT((StgClosure **)GET_SRT(info), info->i.srt_bitmap, c, c_child_r);
            continue;

        case RET_BCO: {
            StgBCO *bco;

            p++;
            retainClosure((StgClosure *)*p, c, c_child_r);
            bco = (StgBCO *)*p;
            p++;
            size = BCO_BITMAP_SIZE(bco);
            retain_large_bitmap(p, BCO_BITMAP(bco), size, c, c_child_r);
            p += size;
            continue;
        }

            // large bitmap (> 32 entries, or > 64 on a 64-bit machine)
        case RET_BIG:
            size = GET_LARGE_BITMAP(&info->i)->size;
            p++;
            retain_large_bitmap(p, GET_LARGE_BITMAP(&info->i),
                                size, c, c_child_r);
            p += size;
            // and don't forget to follow the SRT
            goto follow_srt;
1367

1368
        case RET_FUN: {
1369
            StgRetFun *ret_fun = (StgRetFun *)p;
1370
            const StgFunInfoTable *fun_info;
1371