RetainerProfile.c 71.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11
/* -----------------------------------------------------------------------------
 *
 * (c) The GHC Team, 2001
 * Author: Sungwoo Park
 *
 * Retainer profiling.
 *
 * ---------------------------------------------------------------------------*/

#ifdef PROFILING

12 13 14 15 16 17 18
// Turn off inlining when debugging - it obfuscates things
#ifdef DEBUG
#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 48

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

49
static uint32_t retainerGeneration;  // generation
50

51 52 53
static uint32_t numObjectVisited;    // total number of objects visited
static uint32_t timesAnyObjectVisited;  // number of times any objects are
                                        // visited
54 55 56 57 58 59 60 61 62

/*
  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().
 */

63
StgWord flip = 0;     // flip bit
64 65 66 67 68
                      // must be 0 if DEBUG_RETAINER is on (for static closures)

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

69
static void retainStack(StgClosure *, retainer, StgPtr, StgPtr);
70
static void retainClosure(StgClosure *, StgClosure *, retainer);
71 72 73 74 75 76 77 78 79 80 81 82
#ifdef DEBUG_RETAINER
static void belongToHeap(StgPtr p);
#endif

#ifdef DEBUG_RETAINER
/*
  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
 */
83
static uint32_t cStackSize, maxCStackSize;
84

85
static uint32_t sumOfNewCost;        // sum of the cost of each object, computed
86
                                // when the object is first visited
87
static uint32_t sumOfNewCostExtra;   // for those objects not visited during
88
                                // retainer profiling, e.g., MUT_VAR
89
static uint32_t costArray[N_CLOSURE_TYPES];
90

91
uint32_t sumOfCostLinear;            // sum of the costs of all object, computed
92 93
                                // when linearly traversing the heap after
                                // retainer profiling
94
uint32_t costArrayLinear[N_CLOSURE_TYPES];
95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110
#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,
111
    posTypeLargeSRT,
112 113 114 115 116 117 118 119 120 121
} nextPosType;

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

    // layout.payload
    struct {
    // See StgClosureInfo in InfoTables.h
#if SIZEOF_VOID_P == 8
122 123
        StgWord32 pos;
        StgWord32 ptrs;
124
#else
125 126
        StgWord16 pos;
        StgWord16 ptrs;
127
#endif
128
        StgPtr payload;
129 130 131 132
    } ptrs;

    // SRT
    struct {
133 134
        StgClosure **srt;
        StgWord    srt_bitmap;
135
    } srt;
136 137 138

    // Large SRT
    struct {
139 140
        StgLargeSRT *srt;
        StgWord offset;
141
    } large_srt;
142

143 144 145 146 147 148 149 150 151
} nextPos;

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

typedef struct {
    StgClosure *c;
152
    retainer c_child_r;
153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172
    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
173
static bdescr *firstStack = NULL;
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208
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.
 */
#ifdef DEBUG_RETAINER
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.
 * -------------------------------------------------------------------------- */
209
static INLINE void
210 211 212 213 214 215 216 217 218 219 220 221 222 223
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.
 * -------------------------------------------------------------------------- */
224
static INLINE void
225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240
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) {
241
        freeChain(firstStack);
242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265
    }

    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;
}

/* -----------------------------------------------------------------------------
 * Returns rtsTrue if the whole stack is empty.
 * -------------------------------------------------------------------------- */
266
static INLINE rtsBool
267 268 269 270 271
isEmptyRetainerStack( void )
{
    return (firstStack == currentStack) && stackTop == stackLimit;
}

sof's avatar
sof committed
272 273 274
/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
275
#ifdef DEBUG
276
W_
277
retainerStackBlocks( void )
sof's avatar
sof committed
278 279
{
    bdescr* bd;
280
    W_ res = 0;
sof's avatar
sof committed
281

282
    for (bd = firstStack; bd != NULL; bd = bd->link)
sof's avatar
sof committed
283 284 285 286
      res += bd->blocks;

    return res;
}
287
#endif
sof's avatar
sof committed
288

289 290 291 292
/* -----------------------------------------------------------------------------
 * Returns rtsTrue if stackTop is at the stack boundary of the current stack,
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
293
static INLINE rtsBool
294 295 296 297 298 299 300 301 302 303
isOnBoundary( void )
{
    return stackTop == currentStackBoundary;
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
304
static INLINE void
305
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
306 307 308 309 310 311 312 313 314 315
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
316
static INLINE StgClosure *
317 318 319
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
320
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
321
    } else {
322
        return NULL;
323 324 325 326 327 328
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
329
static INLINE void
330
init_srt_fun( stackPos *info, StgFunInfoTable *infoTable )
331
{
332
    if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
333 334 335
        info->type = posTypeLargeSRT;
        info->next.large_srt.srt = (StgLargeSRT *)GET_FUN_SRT(infoTable);
        info->next.large_srt.offset = 0;
336
    } else {
337 338 339
        info->type = posTypeSRT;
        info->next.srt.srt = (StgClosure **)GET_FUN_SRT(infoTable);
        info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
340
    }
341 342
}

343
static INLINE void
344 345
init_srt_thunk( stackPos *info, StgThunkInfoTable *infoTable )
{
346
    if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
347 348 349
        info->type = posTypeLargeSRT;
        info->next.large_srt.srt = (StgLargeSRT *)GET_SRT(infoTable);
        info->next.large_srt.offset = 0;
350
    } else {
351 352 353
        info->type = posTypeSRT;
        info->next.srt.srt = (StgClosure **)GET_SRT(infoTable);
        info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
354
    }
355 356 357 358 359
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
360
static INLINE StgClosure *
361 362 363
find_srt( stackPos *info )
{
    StgClosure *c;
364
    StgWord bitmap;
365

366
    if (info->type == posTypeSRT) {
367 368 369 370
        // Small SRT bitmap
        bitmap = info->next.srt.srt_bitmap;
        while (bitmap != 0) {
            if ((bitmap & 1) != 0) {
371
#if defined(COMPILING_WINDOWS_DLL)
372 373 374 375
                if ((unsigned long)(*(info->next.srt.srt)) & 0x1)
                    c = (* (StgClosure **)((unsigned long)*(info->next.srt.srt)) & ~0x1);
                else
                    c = *(info->next.srt.srt);
376
#else
377
                c = *(info->next.srt.srt);
378
#endif
379 380 381 382 383 384 385 386 387 388
                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;
389 390
    }
    else {
391
        // Large SRT bitmap
392
        uint32_t i = info->next.large_srt.offset;
393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414
        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;
415 416 417 418 419 420 421 422 423 424 425 426 427 428
    }
}

/* -----------------------------------------------------------------------------
 *  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
 *  child, *c_chlid is set to that child and nothing is pushed onto
 *  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.
429
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
430 431 432
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
433
static INLINE void
434
push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
435 436 437 438 439
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

#ifdef DEBUG_RETAINER
440
    // debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
441 442 443
#endif

    ASSERT(get_itbl(c)->type != TSO);
444
    ASSERT(get_itbl(c)->type != AP_STACK);
445 446 447 448 449 450 451 452 453 454

    //
    // fill in se
    //

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

    // fill in se.info
    switch (get_itbl(c)->type) {
455
        // no child, no SRT
456 457 458
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
459 460
        *first_child = NULL;
        return;
461

462
        // one child (fixed), no SRT
463 464
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
465 466
        *first_child = ((StgMutVar *)c)->var;
        return;
467
    case THUNK_SELECTOR:
468 469
        *first_child = ((StgSelector *)c)->selectee;
        return;
470
    case BLACKHOLE:
471 472
        *first_child = ((StgInd *)c)->indirectee;
        return;
473 474
    case CONSTR_1_0:
    case CONSTR_1_1:
475 476
        *first_child = c->payload[0];
        return;
477

478 479 480
        // 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;
481

482 483
        // two children (fixed), no SRT
        // need to push a stackElement, but nothing to store in se.info
484
    case CONSTR_2_0:
485 486 487 488
        *first_child = c->payload[0];         // return the first pointer
        // se.info.type = posTypeStep;
        // se.info.next.step = 2;            // 2 = second
        break;
489

490 491
        // three children (fixed), no SRT
        // need to push a stackElement
492 493
    case MVAR_CLEAN:
    case MVAR_DIRTY:
494 495 496 497 498 499 500 501
        // 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
502
    case WEAK:
503 504 505 506
        *first_child = ((StgWeak *)c)->key;
        // se.info.type = posTypeStep;
        se.info.next.step = 2;
        break;
507

508
        // layout.payload.ptrs, no SRT
509
    case TVAR:
510
    case CONSTR:
511
    case PRIM:
512
    case MUT_PRIM:
513 514
    case BCO:
    case CONSTR_STATIC:
515 516 517 518 519 520 521 522
        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
523 524
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
525
    case MUT_ARR_PTRS_FROZEN:
526
    case MUT_ARR_PTRS_FROZEN0:
527 528 529 530 531 532 533 534
        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
535 536 537 538
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case SMALL_MUT_ARR_PTRS_FROZEN:
    case SMALL_MUT_ARR_PTRS_FROZEN0:
539 540 541 542 543 544
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;
545

546 547 548
    // layout.payload.ptrs, SRT
    case FUN:           // *c is a heap object.
    case FUN_2_0:
549 550 551 552 553 554
        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;
555

556 557
    case THUNK:
    case THUNK_2_0:
558 559 560 561 562 563 564 565 566
        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
567 568
    case FUN_1_0:
    case FUN_1_1:
569 570 571 572
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;
573

574 575
    case THUNK_1_0:
    case THUNK_1_1:
576 577 578 579
        *first_child = ((StgThunk *)c)->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        break;
580 581

    case FUN_STATIC:      // *c is a heap object.
582
        ASSERT(get_itbl(c)->srt_bitmap != 0);
583 584
    case FUN_0_1:
    case FUN_0_2:
585 586
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
587 588 589 590
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;
591 592 593

    // SRT only
    case THUNK_STATIC:
594
        ASSERT(get_itbl(c)->srt_bitmap != 0);
595 596
    case THUNK_0_1:
    case THUNK_0_2:
597 598
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
599 600 601 602 603
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

604
    case TREC_CHUNK:
605 606 607
        *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
        se.info.next.step = 0;  // entry no.
        break;
608

609
        // cannot appear
610
    case PAP:
611 612
    case AP:
    case AP_STACK:
613
    case TSO:
614
    case STACK:
615 616
    case IND_STATIC:
    case CONSTR_NOCAF_STATIC:
617
        // stack objects
618 619
    case UPDATE_FRAME:
    case CATCH_FRAME:
620
    case UNDERFLOW_FRAME:
621 622 623 624
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
625
        // invalid objects
626 627 628
    case IND:
    case INVALID_OBJECT:
    default:
629 630
        barf("Invalid object *c in push()");
        return;
631 632 633 634
    }

    if (stackTop - 1 < stackBottom) {
#ifdef DEBUG_RETAINER
635
        // debugBelch("push() to the next stack.\n");
636
#endif
637 638 639 640 641 642 643 644 645 646 647 648 649
        // 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);
650 651 652 653 654 655 656 657
    }

    // 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.
658 659 660 661 662 663

    // 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?
664 665 666 667 668 669
    *stackTop = se;

#ifdef DEBUG_RETAINER
    stackSize++;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
    // ASSERT(stackSize >= 0);
670
    // debugBelch("stackSize = %d\n", stackSize);
671 672 673 674 675 676 677 678 679 680 681 682 683
#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
684
 *    INLINE function (for the sake of execution speed).  popOffReal()
685 686 687 688 689 690 691 692
 *    is called only within popOff() and nowhere else.
 * -------------------------------------------------------------------------- */
static void
popOffReal(void)
{
    bdescr *pbd;    // Previous Block Descriptor

#ifdef DEBUG_RETAINER
693
    // debugBelch("pop() to the previous stack.\n");
694 695 696 697 698 699
#endif

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

    if (firstStack == currentStack) {
700 701 702
        // The stack is completely empty.
        stackTop++;
        ASSERT(stackTop == stackLimit);
703
#ifdef DEBUG_RETAINER
704 705 706 707 708 709
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
        /*
          ASSERT(stackSize >= 0);
          debugBelch("stackSize = %d\n", stackSize);
        */
710
#endif
711
        return;
712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728
    }

    // 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);

#ifdef DEBUG_RETAINER
    stackSize--;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
    /*
      ASSERT(stackSize >= 0);
729
      debugBelch("stackSize = %d\n", stackSize);
730 731 732 733
    */
#endif
}

734
static INLINE void
735 736
popOff(void) {
#ifdef DEBUG_RETAINER
737
    // debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
738 739 740 741 742 743 744
#endif

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

    // <= (instead of <) is wrong!
    if (stackTop + 1 < stackLimit) {
745
        stackTop++;
746
#ifdef DEBUG_RETAINER
747 748 749 750 751 752
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
        /*
          ASSERT(stackSize >= 0);
          debugBelch("stackSize = %d\n", stackSize);
        */
753
#endif
754
        return;
755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775
    }

    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
 *  the current stack chunk becomes empty, indicated by rtsTrue returned by
 *  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.
 * -------------------------------------------------------------------------- */
776
static INLINE void
777
pop( StgClosure **c, StgClosure **cp, retainer *r )
778 779 780 781
{
    stackElement *se;

#ifdef DEBUG_RETAINER
782
    // debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
783 784 785
#endif

    do {
786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804
        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
805 806
        case MVAR_CLEAN:
        case MVAR_DIRTY:
807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839
            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;
840 841
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859
            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;
        }
860

861 862
        case TVAR:
        case CONSTR:
863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942
        case PRIM:
        case MUT_PRIM:
        case BCO:
        case CONSTR_STATIC:
            // 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:
943 944
        case STACK:
        case IND_STATIC:
945 946
        case CONSTR_NOCAF_STATIC:
            // stack objects
947
        case UPDATE_FRAME:
948
        case CATCH_FRAME:
949 950
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
951 952 953 954 955 956 957 958 959 960
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default:
            barf("Invalid object *c in pop()");
            return;
        }
961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999
    } while (rtsTrue);
}

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

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

/* -----------------------------------------------------------------------------
 *  This function must be called before f-closing prof_file.
 * -------------------------------------------------------------------------- */
void
endRetainerProfiling( void )
{
#ifdef SECOND_APPROACH
    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.
 * -------------------------------------------------------------------------- */
1000
static INLINE void
1001 1002 1003
maybeInitRetainerSet( StgClosure *c )
{
    if (!isRetainerSetFieldValid(c)) {
1004
        setRetainerSetToNull(c);
1005 1006 1007 1008 1009 1010
    }
}

/* -----------------------------------------------------------------------------
 * Returns rtsTrue if *c is a retainer.
 * -------------------------------------------------------------------------- */
1011
static INLINE rtsBool
1012 1013 1014
isRetainer( StgClosure *c )
{
    switch (get_itbl(c)->type) {
1015 1016 1017 1018
        //
        //  True case
        //
        // TSOs MUST be retainers: they constitute the set of roots.
1019
    case TSO:
1020
    case STACK:
1021

1022
        // mutable objects
1023
    case MUT_PRIM:
1024 1025
    case MVAR_CLEAN:
    case MVAR_DIRTY:
1026
    case TVAR:
1027 1028
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
1029 1030
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
1031

1032
        // thunks are retainers.
1033 1034 1035 1036 1037 1038 1039
    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:
1040 1041
    case AP:
    case AP_STACK:
1042

1043
        // Static thunks, or CAFS, are obviously retainers.
1044 1045
    case THUNK_STATIC:

1046 1047
        // WEAK objects are roots; there is separate code in which traversing
        // begins from WEAK objects.
1048
    case WEAK:
1049
        return rtsTrue;
1050

1051 1052 1053
        //
        // False case
        //
1054

1055
        // constructors
1056 1057 1058 1059 1060 1061
    case CONSTR:
    case CONSTR_1_0:
    case CONSTR_0_1:
    case CONSTR_2_0:
    case CONSTR_1_1:
    case CONSTR_0_2:
1062
        // functions
1063 1064 1065 1066 1067 1068
    case FUN:
    case FUN_1_0:
    case FUN_0_1:
    case FUN_2_0:
    case FUN_1_1:
    case FUN_0_2:
1069
        // partial applications
1070
    case PAP:
1071
        // indirection
Ian Lynagh's avatar
Ian Lynagh committed
1072 1073 1074 1075
    // 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:
1076
    case BLACKHOLE:
1077
        // static objects
1078 1079
    case CONSTR_STATIC:
    case FUN_STATIC:
1080
        // misc
1081
    case PRIM:
1082 1083
    case BCO:
    case ARR_WORDS:
1084
        // STM
1085
    case TREC_CHUNK:
1086 1087 1088
        // immutable arrays
    case MUT_ARR_PTRS_FROZEN:
    case MUT_ARR_PTRS_FROZEN0:
1089
        return rtsFalse;
1090

1091 1092 1093 1094 1095
        //
        // Error case
        //
        // CONSTR_NOCAF_STATIC
        // cannot be *c, *cp, *r in the retainer profiling loop.
1096
    case CONSTR_NOCAF_STATIC:
1097 1098
        // Stack objects are invalid because they are never treated as
        // legal objects during retainer profiling.
1099 1100
    case UPDATE_FRAME:
    case CATCH_FRAME:
1101
    case UNDERFLOW_FRAME:
1102 1103 1104 1105
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
1106
        // other cases
1107 1108 1109
    case IND:
    case INVALID_OBJECT:
    default:
1110 1111
        barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
        return rtsFalse;
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123
    }
}

/* -----------------------------------------------------------------------------
 *  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
1124
 *    RetainerSet.c can simply do nothing.
1125 1126 1127 1128
 *    If this is not the case, we can free all the retainer sets and
 *    re-initialize the hash table.
 *    See refreshAllRetainerSet() in RetainerSet.c.
 * -------------------------------------------------------------------------- */
1129
static INLINE retainer
1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152
getRetainerFrom( StgClosure *c )
{
    ASSERT(isRetainer(c));

#if defined(RETAINER_SCHEME_INFO)
    // Retainer scheme 1: retainer = info table
    return get_itbl(c);
#elif defined(RETAINER_SCHEME_CCS)
    // Retainer scheme 2: retainer = cost centre stack
    return c->header.prof.ccs;
#elif defined(RETAINER_SCHEME_CC)
    // Retainer scheme 3: retainer = cost centre
    return c->header.prof.ccs->cc;
#endif
}

/* -----------------------------------------------------------------------------
 *  Associates the retainer set *s with the closure *c, that is, *s becomes
 *  the retainer set of *c.
 *  Invariants:
 *    c != NULL
 *    s != NULL
 * -------------------------------------------------------------------------- */
1153
static INLINE void
1154
associate( StgClosure *c, RetainerSet *s )
1155 1156 1157 1158 1159 1160
{
    // StgWord has the same size as pointers, so the following type
    // casting is okay.
    RSET(c) = (RetainerSet *)((StgWord)s | flip);
}

1161 1162 1163 1164 1165
/* -----------------------------------------------------------------------------
   Call retainClosure for each of the closures covered by a large bitmap.
   -------------------------------------------------------------------------- */

static void
1166
retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, uint32_t size,
1167
                     StgClosure *c, retainer c_child_r)
1168
{
1169
    uint32_t i, b;
1170
    StgWord bitmap;
1171

1172 1173 1174
    b = 0;
    bitmap = large_bitmap->bitmap[b];
    for (i = 0; i < size; ) {
1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185
        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;
        }
1186 1187 1188
    }
}

1189
static INLINE StgPtr
1190
retain_small_bitmap (StgPtr p, uint32_t size, StgWord bitmap,
1191
                     StgClosure *c, retainer c_child_r)
1192 1193
{
    while (size > 0) {
1194 1195 1196 1197 1198 1199
        if ((bitmap & 1) == 0) {
            retainClosure((StgClosure *)*p, c, c_child_r);
        }
        p++;
        bitmap = bitmap >> 1;
        size--;
1200 1201 1202 1203
    }
    return p;
}

1204 1205 1206 1207 1208 1209 1210
/* -----------------------------------------------------------------------------
 * Call retainClosure for each of the closures in an SRT.
 * ------------------------------------------------------------------------- */

static void
retain_large_srt_bitmap (StgLargeSRT *srt, StgClosure *c, retainer c_child_r)
{
1211
    uint32_t i, b, size;
1212 1213
    StgWord bitmap;
    StgClosure **p;
1214

1215 1216 1217 1218 1219
    b = 0;
    p = (StgClosure **)srt->srt;
    size   = srt->l.size;
    bitmap = srt->l.bitmap[b];
    for (i = 0; i < size; ) {
1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230
        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;
        }
1231 1232 1233 1234
    }
}

static INLINE void
1235 1236
retainSRT (StgClosure **srt, uint32_t srt_bitmap, StgClosure *c,
            retainer c_child_r)
1237
{
1238
  uint32_t bitmap;
1239 1240 1241 1242 1243
  StgClosure **p;

  bitmap = srt_bitmap;
  p = srt;

1244
  if (bitmap == (StgHalfWord)(-1)) {
1245 1246 1247 1248 1249 1250
      retain_large_srt_bitmap( (StgLargeSRT *)srt, c, c_child_r );
      return;
  }

  while (bitmap != 0) {
      if ((bitmap & 1) != 0) {
1251
#if defined(COMPILING_WINDOWS_DLL)
1252 1253 1254 1255 1256 1257
          if ( (unsigned long)(*srt) & 0x1 ) {
              retainClosure(* (StgClosure**) ((unsigned long) (*srt) & ~0x1),
                            c, c_child_r);
          } else {
              retainClosure(*srt,c,c_child_r);
          }
1258
#else
1259
          retainClosure(*srt,c,c_child_r);
1260 1261 1262 1263 1264 1265 1266
#endif
      }
      p++;
      bitmap = bitmap >> 1;
  }
}

1267 1268 1269 1270 1271 1272
/* -----------------------------------------------------------------------------
 *  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:
1273
 *    *c is one of the following: TSO, AP_STACK.
1274 1275 1276 1277 1278
 *    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).
1279
 *    If *c is TSO, its state is not ThreadComplete,or ThreadKilled,
1280
 *    which means that its stack is ready to process.
1281 1282 1283 1284 1285
 *  Note:
 *    This code was almost plagiarzied from GC.c! For each pointer,
 *    retainClosure() is invoked instead of evacuate().
 * -------------------------------------------------------------------------- */
static void
1286
retainStack( StgClosure *c, retainer c_child_r,
1287
             StgPtr stackStart, StgPtr stackEnd )
1288 1289
{
    stackElement *oldStackBoundary;
1290 1291
    StgPtr p;
    StgRetInfoTable *info;
1292
    StgWord bitmap;
1293
    uint32_t size;
1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309

#ifdef DEBUG_RETAINER
    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;

#ifdef DEBUG_RETAINER
1310
    // debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n", oldStackBoundary, currentStackBoundary);
1311 1312
#endif

1313 1314
    ASSERT(get_itbl(c)->type == STACK);

1315 1316
    p = stackStart;
    while (p < stackEnd) {
1317
        info = get_ret_itbl((StgClosure *)p);
1318

1319
        switch(info->i.type) {
1320

1321 1322 1323 1324
        case UPDATE_FRAME:
            retainClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
            p += sizeofW(StgUpdateFrame);
            continue;
1325

1326 1327
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
1328 1329 1330 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
        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;
1364

1365
        case RET_FUN: {
1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397
            StgRetFun *ret_fun = (StgRetFun *)p;
            StgFunInfoTable *fun_info;

            retainClosure(ret_fun->fun, c, c_child_r);
            fun_info = get_fun_itbl(UNTAG_CLOSURE(ret_fun->fun));

            p = (P_)&ret_fun->payload;
            switch (fun_info->f.fun_type) {
            case ARG_GEN:
                bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
                size = BITMAP_SIZE(fun_info->f.b.bitmap);
                p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
                break;
            case ARG_GEN_BIG:
                size = GET_FUN_LARGE_BITMAP(fun_info)->size;
                retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
                                    size, c, c_child_r);
                p += size;
                break;
            default:
                bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
                size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
                p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
                break;
            }
            goto follow_srt;
        }

        default:
            barf("Invalid object found in retainStack(): %d",
                 (int)(info->i.type));
        }
1398 1399 1400 1401 1402
    }

    // restore currentStackBoundary
    currentStackBoundary = oldStackBoundary;
#ifdef DEBUG_RETAINER
1403
    // debugBelch("retainStack() finished: currentStackBoundary = 0x%x\n", currentStackBoundary);
1404 1405 1406 1407 1408 1409 1410
#endif

#ifdef DEBUG_RETAINER
    cStackSize--;
#endif
}

1411 1412 1413 1414
/* ----------------------------------------------------------------------------
 * Call retainClosure for each of the children of a PAP/AP
 * ------------------------------------------------------------------------- */

1415
static INLINE StgPtr
Simon Marlow's avatar
Simon Marlow committed
1416
retain_PAP_payload (StgClosure *pap,    /* NOT tagged */
1417
                    retainer c_child_r, /* NOT tagged */
Simon Marlow's avatar
Simon Marlow committed
1418
                    StgClosure *fun,    /* tagged */
1419
                    StgClosure** payload, StgWord n_args)
1420 1421
{
    StgPtr p;
1422
    StgWord bitmap;
1423 1424
    StgFunInfoTable *fun_info;

1425
    retainClosure(fun, pap, c_child_r);
Simon Marlow's avatar
Simon Marlow committed
1426
    fun = UNTAG_CLOSURE(fun);
1427
    fun_info = get_fun_itbl(fun);
1428 1429
    ASSERT(fun_info->i.type != PAP);

1430
    p = (StgPtr)payload;
1431

1432
    switch (fun_info->f.fun_type) {
1433
    case ARG_GEN:
1434 1435 1436 1437
        bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
        p = retain_small_bitmap(p, n_args, bitmap,
                                pap, c_child_r);
        break;
1438
    case ARG_GEN_BIG:
1439 1440 1441 1442
        retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
                            n_args, pap, c_child_r);
        p += n_args;
        break;
1443
    case ARG_BCO:
1444 1445 1446 1447
        retain_large_bitmap((StgPtr)payload, BCO_BITMAP(fun),
                            n_args, pap, c_child_r);
        p += n_args;
        break;
1448
    default:
1449 1450 1451
        bitmap = BITMAP_BITS(stg