RetainerProfile.c 69.2 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"
David Feuer's avatar
David Feuer committed
33 34
#include "StablePtr.h" /* markStablePtrTable */
#include "StableName.h" /* rememberOldStableNameAddresses */
Simon Marlow's avatar
Simon Marlow committed
35
#include "sm/Storage.h" // for END_OF_STATIC_LIST
36

37 38 39 40 41 42 43 44 45 46 47 48
/* 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.
*/


49 50 51 52 53 54 55 56 57 58 59 60 61
/*
  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...
 * -------------------------------------------------------------------------- */

62
static uint32_t retainerGeneration;  // generation
63

64 65 66
static uint32_t numObjectVisited;    // total number of objects visited
static uint32_t timesAnyObjectVisited;  // number of times any objects are
                                        // visited
67 68 69 70 71 72 73 74 75

/*
  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().
 */

76
StgWord flip = 0;     // flip bit
77 78 79 80 81
                      // must be 0 if DEBUG_RETAINER is on (for static closures)

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

82
static void retainStack(StgClosure *, retainer, StgPtr, StgPtr);
83
static void retainClosure(StgClosure *, StgClosure *, retainer);
Ben Gamari's avatar
Ben Gamari committed
84
#if defined(DEBUG_RETAINER)
85 86 87
static void belongToHeap(StgPtr p);
#endif

Ben Gamari's avatar
Ben Gamari committed
88
#if defined(DEBUG_RETAINER)
89 90 91 92 93 94 95
/*
  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
 */
96
static uint32_t cStackSize, maxCStackSize;
97

98
static uint32_t sumOfNewCost;        // sum of the cost of each object, computed
99
                                // when the object is first visited
100
static uint32_t sumOfNewCostExtra;   // for those objects not visited during
101
                                // retainer profiling, e.g., MUT_VAR
102
static uint32_t costArray[N_CLOSURE_TYPES];
103

104
uint32_t sumOfCostLinear;            // sum of the costs of all object, computed
105 106
                                // when linearly traversing the heap after
                                // retainer profiling
107
uint32_t costArrayLinear[N_CLOSURE_TYPES];
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131
#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,
} 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
        StgClosure *srt;
141 142 143 144 145 146 147 148 149 150
    } srt;
} nextPos;

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

typedef struct {
    StgClosure *c;
151
    retainer c_child_r;
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
    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
172
static bdescr *firstStack = NULL;
173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195
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
196
#if defined(DEBUG_RETAINER)
197 198 199 200 201 202 203 204 205 206 207
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.
 * -------------------------------------------------------------------------- */
208
static INLINE void
209 210 211 212 213 214 215 216 217 218 219 220 221 222
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.
 * -------------------------------------------------------------------------- */
223
static INLINE void
224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239
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) {
240
        freeChain(firstStack);
241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262
    }

    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
263
 * Returns true if the whole stack is empty.
264
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
265
static INLINE bool
266 267 268 269 270
isEmptyRetainerStack( void )
{
    return (firstStack == currentStack) && stackTop == stackLimit;
}

sof's avatar
sof committed
271 272 273
/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
274
W_
275
retainerStackBlocks( void )
sof's avatar
sof committed
276 277
{
    bdescr* bd;
278
    W_ res = 0;
sof's avatar
sof committed
279

280
    for (bd = firstStack; bd != NULL; bd = bd->link)
sof's avatar
sof committed
281 282 283 284 285
      res += bd->blocks;

    return res;
}

286
/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
287
 * Returns true if stackTop is at the stack boundary of the current stack,
288 289
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
290
static INLINE bool
291 292 293 294 295 296 297 298 299 300
isOnBoundary( void )
{
    return stackTop == currentStackBoundary;
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
301
static INLINE void
302
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
303 304 305 306 307 308 309 310 311 312
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
313
static INLINE StgClosure *
314 315 316
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
317
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
318
    } else {
319
        return NULL;
320 321 322 323 324 325
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
326
static INLINE void
327
init_srt_fun( stackPos *info, const StgFunInfoTable *infoTable )
328
{
329 330 331
    info->type = posTypeSRT;
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_FUN_SRT(infoTable);
332
    } else {
333
        info->next.srt.srt = NULL;
334
    }
335 336
}

337
static INLINE void
338
init_srt_thunk( stackPos *info, const StgThunkInfoTable *infoTable )
339
{
340 341
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_SRT(infoTable);
342
    } else {
343
        info->next.srt.srt = NULL;
344
    }
345 346 347 348 349
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
350
static INLINE StgClosure *
351 352 353
find_srt( stackPos *info )
{
    StgClosure *c;
354
    if (info->type == posTypeSRT) {
355 356 357
        c = info->next.srt.srt;
        info->next.srt.srt = NULL;
        return c;
358 359 360 361 362 363 364
    }
}

/* -----------------------------------------------------------------------------
 *  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
365
 *  child, *c_child is set to that child and nothing is pushed onto
366 367 368 369 370 371
 *  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.
372
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
373 374 375
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
376
static INLINE void
377
push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
378 379 380 381
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
382
#if defined(DEBUG_RETAINER)
383
    // debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
384 385 386
#endif

    ASSERT(get_itbl(c)->type != TSO);
387
    ASSERT(get_itbl(c)->type != AP_STACK);
388 389 390 391 392 393 394 395 396 397

    //
    // fill in se
    //

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

    // fill in se.info
    switch (get_itbl(c)->type) {
398
        // no child, no SRT
399 400 401
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
gcampax's avatar
gcampax committed
402
    case COMPACT_NFDATA:
403 404
        *first_child = NULL;
        return;
405

406
        // one child (fixed), no SRT
407 408
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
409 410
        *first_child = ((StgMutVar *)c)->var;
        return;
411
    case THUNK_SELECTOR:
412 413
        *first_child = ((StgSelector *)c)->selectee;
        return;
414
    case BLACKHOLE:
415 416
        *first_child = ((StgInd *)c)->indirectee;
        return;
417 418
    case CONSTR_1_0:
    case CONSTR_1_1:
419 420
        *first_child = c->payload[0];
        return;
421

422 423 424
        // 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;
425

426 427
        // two children (fixed), no SRT
        // need to push a stackElement, but nothing to store in se.info
428
    case CONSTR_2_0:
429 430 431 432
        *first_child = c->payload[0];         // return the first pointer
        // se.info.type = posTypeStep;
        // se.info.next.step = 2;            // 2 = second
        break;
433

434 435
        // three children (fixed), no SRT
        // need to push a stackElement
436 437
    case MVAR_CLEAN:
    case MVAR_DIRTY:
438 439 440 441 442 443 444 445
        // 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
446
    case WEAK:
447 448 449 450
        *first_child = ((StgWeak *)c)->key;
        // se.info.type = posTypeStep;
        se.info.next.step = 2;
        break;
451

452
        // layout.payload.ptrs, no SRT
453
    case TVAR:
454
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
455
    case CONSTR_NOCAF:
456
    case PRIM:
457
    case MUT_PRIM:
458
    case BCO:
459 460 461 462 463 464 465 466
        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
467 468
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
469 470
    case MUT_ARR_PTRS_FROZEN_CLEAN:
    case MUT_ARR_PTRS_FROZEN_DIRTY:
471 472 473 474 475 476 477 478
        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
479 480
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
481 482
    case SMALL_MUT_ARR_PTRS_FROZEN_CLEAN:
    case SMALL_MUT_ARR_PTRS_FROZEN_DIRTY:
483 484 485 486 487 488
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;
489

490
    // layout.payload.ptrs, SRT
491 492 493
    case FUN_STATIC:
        ASSERT(get_itbl(c)->srt != 0);
        /* fallthrough */
494 495
    case FUN:           // *c is a heap object.
    case FUN_2_0:
496 497 498 499 500 501
        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;
502

503 504
    case THUNK:
    case THUNK_2_0:
505 506 507 508 509 510 511 512 513
        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
514 515
    case FUN_1_0:
    case FUN_1_1:
516 517 518 519
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;
520

521 522
    case THUNK_1_0:
    case THUNK_1_1:
523 524 525 526
        *first_child = ((StgThunk *)c)->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        break;
527

528
    case FUN_0_1:      // *c is a heap object.
529
    case FUN_0_2:
530 531
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
532 533 534 535
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;
536 537 538

    // SRT only
    case THUNK_STATIC:
539
        ASSERT(get_itbl(c)->srt != 0);
540 541
    case THUNK_0_1:
    case THUNK_0_2:
542 543
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
544 545 546 547 548
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

549
    case TREC_CHUNK:
550 551 552
        *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
        se.info.next.step = 0;  // entry no.
        break;
553

554
        // cannot appear
555
    case PAP:
556 557
    case AP:
    case AP_STACK:
558
    case TSO:
559
    case STACK:
560
    case IND_STATIC:
561
        // stack objects
562 563
    case UPDATE_FRAME:
    case CATCH_FRAME:
564
    case UNDERFLOW_FRAME:
565 566 567 568
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
569
        // invalid objects
570 571 572
    case IND:
    case INVALID_OBJECT:
    default:
573
        barf("Invalid object *c in push(): %d", get_itbl(c)->type);
574
        return;
575 576 577
    }

    if (stackTop - 1 < stackBottom) {
Ben Gamari's avatar
Ben Gamari committed
578
#if defined(DEBUG_RETAINER)
579
        // debugBelch("push() to the next stack.\n");
580
#endif
581 582 583 584 585 586 587 588 589 590 591 592 593
        // 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);
594 595 596 597 598 599 600 601
    }

    // 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.
602 603 604 605 606 607

    // 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?
608 609
    *stackTop = se;

Ben Gamari's avatar
Ben Gamari committed
610
#if defined(DEBUG_RETAINER)
611 612 613
    stackSize++;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
    // ASSERT(stackSize >= 0);
614
    // debugBelch("stackSize = %d\n", stackSize);
615 616 617 618 619 620 621 622 623 624 625 626 627
#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
628
 *    INLINE function (for the sake of execution speed).  popOffReal()
629 630 631 632 633 634 635
 *    is called only within popOff() and nowhere else.
 * -------------------------------------------------------------------------- */
static void
popOffReal(void)
{
    bdescr *pbd;    // Previous Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
636
#if defined(DEBUG_RETAINER)
637
    // debugBelch("pop() to the previous stack.\n");
638 639 640 641 642 643
#endif

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

    if (firstStack == currentStack) {
644 645 646
        // The stack is completely empty.
        stackTop++;
        ASSERT(stackTop == stackLimit);
Ben Gamari's avatar
Ben Gamari committed
647
#if defined(DEBUG_RETAINER)
648 649 650 651 652 653
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
        /*
          ASSERT(stackSize >= 0);
          debugBelch("stackSize = %d\n", stackSize);
        */
654
#endif
655
        return;
656 657 658 659 660 661 662 663 664 665 666 667
    }

    // 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
668
#if defined(DEBUG_RETAINER)
669 670 671 672
    stackSize--;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
    /*
      ASSERT(stackSize >= 0);
673
      debugBelch("stackSize = %d\n", stackSize);
674 675 676 677
    */
#endif
}

678
static INLINE void
679
popOff(void) {
Ben Gamari's avatar
Ben Gamari committed
680
#if defined(DEBUG_RETAINER)
681
    // debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
682 683 684 685 686 687 688
#endif

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

    // <= (instead of <) is wrong!
    if (stackTop + 1 < stackLimit) {
689
        stackTop++;
Ben Gamari's avatar
Ben Gamari committed
690
#if defined(DEBUG_RETAINER)
691 692 693 694 695 696
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
        /*
          ASSERT(stackSize >= 0);
          debugBelch("stackSize = %d\n", stackSize);
        */
697
#endif
698
        return;
699 700 701 702 703 704 705 706 707 708 709 710 711 712 713
    }

    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
714
 *  the current stack chunk becomes empty, indicated by true returned by
715 716 717 718 719
 *  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.
 * -------------------------------------------------------------------------- */
720
static INLINE void
721
pop( StgClosure **c, StgClosure **cp, retainer *r )
722 723 724
{
    stackElement *se;

Ben Gamari's avatar
Ben Gamari committed
725
#if defined(DEBUG_RETAINER)
726
    // debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
727 728 729
#endif

    do {
730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748
        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
749 750
        case MVAR_CLEAN:
        case MVAR_DIRTY:
751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783
            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;
784 785
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803
            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;
        }
804

805 806
        case TVAR:
        case CONSTR:
807 808 809 810 811 812
        case PRIM:
        case MUT_PRIM:
        case BCO:
            // StgMutArrPtr.ptrs, no SRT
        case MUT_ARR_PTRS_CLEAN:
        case MUT_ARR_PTRS_DIRTY:
813 814
        case MUT_ARR_PTRS_FROZEN_CLEAN:
        case MUT_ARR_PTRS_FROZEN_DIRTY:
815 816 817 818
        case SMALL_MUT_ARR_PTRS_CLEAN:
        case SMALL_MUT_ARR_PTRS_DIRTY:
        case SMALL_MUT_ARR_PTRS_FROZEN_CLEAN:
        case SMALL_MUT_ARR_PTRS_FROZEN_DIRTY:
819 820 821 822 823 824 825 826 827 828 829
            *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
830
        case FUN_STATIC:
831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889
        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_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:
890 891
        case STACK:
        case IND_STATIC:
Simon Marlow's avatar
Simon Marlow committed
892
        case CONSTR_NOCAF:
893
            // stack objects
894
        case UPDATE_FRAME:
895
        case CATCH_FRAME:
896 897
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
898 899 900 901 902 903 904
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default:
905
            barf("Invalid object *c in pop(): %d", get_itbl(se->c)->type);
906 907
            return;
        }
Ben Gamari's avatar
Ben Gamari committed
908
    } while (true);
909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927
}

/* -----------------------------------------------------------------------------
 * 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
928
#if defined(SECOND_APPROACH)
929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946
    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.
 * -------------------------------------------------------------------------- */
947
static INLINE void
948 949 950
maybeInitRetainerSet( StgClosure *c )
{
    if (!isRetainerSetFieldValid(c)) {
951
        setRetainerSetToNull(c);
952 953 954 955
    }
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
956
 * Returns true if *c is a retainer.
957
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
958
static INLINE bool
959 960 961
isRetainer( StgClosure *c )
{
    switch (get_itbl(c)->type) {
962 963 964 965
        //
        //  True case
        //
        // TSOs MUST be retainers: they constitute the set of roots.
966
    case TSO:
967
    case STACK:
968

969
        // mutable objects
970
    case MUT_PRIM:
971 972
    case MVAR_CLEAN:
    case MVAR_DIRTY:
973
    case TVAR:
974 975
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
976 977
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
978 979 980
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case BLOCKING_QUEUE:
981

982
        // thunks are retainers.
983 984 985 986 987 988 989
    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:
990 991
    case AP:
    case AP_STACK:
992

993
        // Static thunks, or CAFS, are obviously retainers.
994 995
    case THUNK_STATIC:

996 997
        // WEAK objects are roots; there is separate code in which traversing
        // begins from WEAK objects.
998
    case WEAK:
Ben Gamari's avatar
Ben Gamari committed
999
        return true;
1000

1001 1002 1003
        //
        // False case
        //
1004

1005
        // constructors
1006
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
1007
    case CONSTR_NOCAF:
1008 1009 1010 1011 1012
    case CONSTR_1_0:
    case CONSTR_0_1:
    case CONSTR_2_0:
    case CONSTR_1_1:
    case CONSTR_0_2:
1013
        // functions
1014 1015 1016 1017 1018 1019
    case FUN:
    case FUN_1_0:
    case FUN_0_1:
    case FUN_2_0:
    case FUN_1_1:
    case FUN_0_2:
1020
        // partial applications
1021
    case PAP:
1022
        // indirection
Ian Lynagh's avatar
Ian Lynagh committed
1023 1024 1025 1026
    // 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:
1027
    case BLACKHOLE:
1028
    case WHITEHOLE:
1029
        // static objects
1030
    case FUN_STATIC:
1031
        // misc
1032
    case PRIM:
1033 1034
    case BCO:
    case ARR_WORDS:
1035
    case COMPACT_NFDATA:
1036
        // STM
1037
    case TREC_CHUNK:
1038
        // immutable arrays
1039 1040 1041 1042
    case MUT_ARR_PTRS_FROZEN_CLEAN:
    case MUT_ARR_PTRS_FROZEN_DIRTY:
    case SMALL_MUT_ARR_PTRS_FROZEN_CLEAN:
    case SMALL_MUT_ARR_PTRS_FROZEN_DIRTY:
Ben Gamari's avatar
Ben Gamari committed
1043
        return false;
1044

1045 1046 1047 1048 1049
        //
        // Error case
        //
        // Stack objects are invalid because they are never treated as
        // legal objects during retainer profiling.
1050 1051
    case UPDATE_FRAME:
    case CATCH_FRAME:
1052 1053
    case CATCH_RETRY_FRAME:
    case CATCH_STM_FRAME:
1054
    case UNDERFLOW_FRAME:
1055
    case ATOMICALLY_FRAME:
1056 1057 1058 1059
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
1060
    case RET_FUN:
1061
        // other cases
1062 1063 1064
    case IND:
    case INVALID_OBJECT:
    default:
1065
        barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
Ben Gamari's avatar
Ben Gamari committed
1066
        return false;
1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078
    }
}

/* -----------------------------------------------------------------------------
 *  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
1079
 *    RetainerSet.c can simply do nothing.
1080 1081 1082 1083
 *    If this is not the case, we can free all the retainer sets and
 *    re-initialize the hash table.
 *    See refreshAllRetainerSet() in RetainerSet.c.
 * -------------------------------------------------------------------------- */
1084
static INLINE retainer
1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098
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
 * -------------------------------------------------------------------------- */
1099
static INLINE void
1100
associate( StgClosure *c, RetainerSet *s )
1101 1102 1103 1104 1105 1106
{
    // StgWord has the same size as pointers, so the following type
    // casting is okay.
    RSET(c) = (RetainerSet *)((StgWord)s | flip);
}

1107 1108 1109 1110 1111
/* -----------------------------------------------------------------------------
   Call retainClosure for each of the closures covered by a large bitmap.
   -------------------------------------------------------------------------- */

static void
1112
retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, uint32_t size,
1113
                     StgClosure *c, retainer c_child_r)
1114
{
1115
    uint32_t i, b;
1116
    StgWord bitmap;
1117

1118 1119 1120
    b = 0;
    bitmap = large_bitmap->bitmap[b];
    for (i = 0; i < size; ) {
1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131
        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;
        }
1132 1133 1134
    }
}

1135
static INLINE StgPtr
1136
retain_small_bitmap (StgPtr p, uint32_t size, StgWord bitmap,
1137
                     StgClosure *c, retainer c_child_r)
1138 1139
{
    while (size > 0) {
1140 1141 1142 1143 1144 1145
        if ((bitmap & 1) == 0) {
            retainClosure((StgClosure *)*p, c, c_child_r);
        }
        p++;
        bitmap = bitmap >> 1;
        size--;
1146 1147 1148 1149
    }
    return p;
}

1150 1151 1152 1153 1154 1155
/* -----------------------------------------------------------------------------
 *  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:
1156
 *    *c is one of the following: TSO, AP_STACK.
1157 1158 1159 1160 1161
 *    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).
1162
 *    If *c is TSO, its state is not ThreadComplete,or ThreadKilled,
1163
 *    which means that its stack is ready to process.
1164 1165 1166 1167 1168
 *  Note:
 *    This code was almost plagiarzied from GC.c! For each pointer,
 *    retainClosure() is invoked instead of evacuate().
 * -------------------------------------------------------------------------- */
static void
1169
retainStack( StgClosure *c, retainer c_child_r,
1170
             StgPtr stackStart, StgPtr stackEnd )
1171 1172
{
    stackElement *oldStackBoundary;
1173
    StgPtr p;
1174
    const StgRetInfoTable *info;
1175
    StgWord bitmap;
1176
    uint32_t size;
1177

Ben Gamari's avatar
Ben Gamari committed
1178
#if defined(DEBUG_RETAINER)
1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191
    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
1192
#if defined(DEBUG_RETAINER)
1193
    // debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n", oldStackBoundary, currentStackBoundary);
1194 1195
#endif

1196 1197
    ASSERT(get_itbl(c)->type == STACK);

1198 1199
    p = stackStart;
    while (p < stackEnd) {
1200
        info = get_ret_itbl((StgClosure *)p);
1201

1202
        switch(info->i.type) {
1203

1204 1205 1206 1207
        case UPDATE_FRAME:
            retainClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
            p += sizeofW(StgUpdateFrame);
            continue;
1208

1209 1210
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221
        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:
1222 1223 1224
            if (info->i.srt) {
                retainClosure(GET_SRT(info),c,c_child_r);
            }
1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248
            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;
1249

1250
        case RET_FUN: {
1251
            StgRetFun *ret_fun = (StgRetFun *)p;
1252
            const StgFunInfoTable *fun_info;
1253 1254

            retainClosure(ret_fun->fun, c, c_child_r);
1255
            fun_info = get_fun_itbl(UNTAG_CONST_CLOSURE(ret_fun->fun));
1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282

            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));
        }
1283 1284 1285 1286
    }

    // restore currentStackBoundary
    currentStackBoundary = oldStackBoundary;
Ben Gamari's avatar
Ben Gamari committed
1287
#if defined(DEBUG_RETAINER)
1288
    // debugBelch("retainStack() finished: currentStackBoundary = 0x%x\n", currentStackBoundary);
1289 1290
#endif

Ben Gamari's avatar
Ben Gamari committed
1291
#if defined(DEBUG_RETAINER)
1292 1293 1294 1295
    cStackSize--;
#endif
}

1296 1297 1298 1299
/* ----------------------------------------------------------------------------
 * Call retainClosure for each of the children of a PAP/AP
 * ------------------------------------------------------------------------- */

1300
static INLINE StgPtr
Simon Marlow's avatar
Simon Marlow committed
1301
retain_PAP_payload (StgClosure *pap,    /* NOT tagged */
1302
                    retainer c_child_r, /* NOT tagged */
Simon Marlow's avatar
Simon Marlow committed
1303
                    StgClosure *fun,    /* tagged */
1304
                    StgClosure** payload, StgWord n_args)
1305 1306
{
    StgPtr p;
1307
    StgWord bitmap;
1308
    const StgFunInfoTable *fun_info;
1309

1310
    retainClosure(fun, pap, c_child_r);
Simon Marlow's avatar
Simon Marlow committed
1311
    fun = UNTAG_CLOSURE(fun);
1312
    fun_info = get_fun_itbl(fun);
1313 1314
    ASSERT(fun_info->i.type != PAP);

1315
    p = (StgPtr)payload;
1316

1317
    switch (fun_info->f.fun_type) {
1318
    case ARG_GEN:
1319 1320 1321 1322
        bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
        p = retain_small_bitmap(p, n_args, bitmap,
                                pap, c_child_r);
        break;
1323
    case ARG_GEN_BIG:
1324 1325 1326 1327
        retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
                            n_args, pap, c_child_r);
        p += n_args;
        break;
1328
    case ARG_BCO:
1329 1330 1331 1332
        retain_large_bitmap((StgPtr)payload, BCO_BITMAP(fun),
                            n_args, pap, c_child_r);
        p += n_args;
        break;
1333
    default:
1334 1335 1336
        bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
        p = retain_small_bitmap(p, n_args, bitmap, pap, c_child_r);
        break;
1337 1338 1339 1340
    }
    return p;
}

1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354
/* -----------------------------------------------------------------------------
 *  Compute the retainer set of *c0 and all its desecents by traversing.
 *  *cp0 is the parent of *c0, and *r0 is the most recent retainer of *c0.
 *  Invariants:
 *    c0 = cp0 = r0 holds only for root objects.
 *    RSET(cp0) and RSET(r0) are valid, i.e., their
 *    interpretation conforms to the current value of flip (even when they
 *    are interpreted to be NULL).
 *    However, RSET(c0) may be corrupt, i.e., it may not conform to
 *    the current value of flip. If it does not, during the execution
 *    of this function,