RetainerProfile.c 71.4 KB
Newer Older
1 2 3 4 5 6 7 8 9
/* -----------------------------------------------------------------------------
 *
 * (c) The GHC Team, 2001
 * Author: Sungwoo Park
 *
 * Retainer profiling.
 *
 * ---------------------------------------------------------------------------*/

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

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

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

22 23 24 25 26 27
#include "RtsUtils.h"
#include "RetainerProfile.h"
#include "RetainerSet.h"
#include "Schedule.h"
#include "Printer.h"
#include "Weak.h"
Simon Marlow's avatar
Simon Marlow committed
28
#include "sm/Sanity.h"
29 30 31
#include "Profiling.h"
#include "Stats.h"
#include "ProfHeap.h"
32
#include "Apply.h"
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
/* Note [What is a retainer?]
   ~~~~~~~~~~~~~~~~~~~~~~~~~~
39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
Retainer profiling is a profiling technique that gives information why
objects can't be freed and lists the consumers that hold pointers to
the heap objects. It does not list all the objects that keeps references
to the other, because then we would keep too much information that will
make the report unusable, for example the cons element of the list would keep
all the tail cells. As a result we are keeping only the objects of the
certain types, see 'isRetainer()' function for more discussion.

More formal definition of the retainer can be given the following way.

An object p is a retainer object of the object l, if all requirements
hold:

  1. p can be a retainer (see `isRetainer()`)
  2. l is reachable from p
  3. There are no other retainers on the path from p to l.

Exact algorithm and additional information can be found the historical
document 'docs/storage-mgt/rp.tex'. Details that are related to the
RTS implementation may be out of date, but the general
information about the retainers is still applicable.
60 61 62
*/


63 64 65 66 67 68 69 70 71
/*
  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.
 */

72 73
// TODO: Change references to c_child_r in comments to 'data'.

74 75 76 77
/* -----------------------------------------------------------------------------
 * Declarations...
 * -------------------------------------------------------------------------- */

78
static uint32_t retainerGeneration;  // generation
79

80 81 82
static uint32_t numObjectVisited;    // total number of objects visited
static uint32_t timesAnyObjectVisited;  // number of times any objects are
                                        // visited
83 84 85 86 87 88 89 90 91

/*
  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().
 */

92
StgWord flip = 0;     // flip bit
93 94 95 96 97
                      // must be 0 if DEBUG_RETAINER is on (for static closures)

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

Ben Gamari's avatar
Ben Gamari committed
98
#if defined(DEBUG_RETAINER)
99
static uint32_t sumOfNewCost;        // sum of the cost of each object, computed
100
                                // when the object is first visited
101
static uint32_t sumOfNewCostExtra;   // for those objects not visited during
102
                                // retainer profiling, e.g., MUT_VAR
103
static uint32_t costArray[N_CLOSURE_TYPES];
104

105
uint32_t sumOfCostLinear;            // sum of the costs of all object, computed
106 107
                                // when linearly traversing the heap after
                                // retainer profiling
108
uint32_t costArrayLinear[N_CLOSURE_TYPES];
109 110 111 112 113 114 115 116 117 118 119 120 121
#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 {
122 123
    // Object with fixed layout. Keeps an information about that
    // element was processed. (stackPos.next.step)
124
    posTypeStep,
125 126
    // Description of the pointers-first heap object. Keeps information
    // about layout. (stackPos.next.ptrs)
127
    posTypePtrs,
128
    // Keeps SRT bitmap (stackPos.next.srt)
129
    posTypeSRT,
130 131 132
    // Keeps a new object that was not inspected yet. Keeps a parent
    // element (stackPos.next.parent)
    posTypeFresh
133 134 135 136 137 138 139 140
} nextPosType;

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

    // layout.payload
    struct {
141 142 143
        // See StgClosureInfo in InfoTables.h
        StgHalfWord pos;
        StgHalfWord ptrs;
144
        StgPtr payload;
145 146 147 148
    } ptrs;

    // SRT
    struct {
149
        StgClosure *srt;
150 151 152
    } srt;
} nextPos;

153 154
// Tagged stack element, that keeps information how to process
// the next element in the traverse stack.
155 156 157 158 159
typedef struct {
    nextPosType type;
    nextPos next;
} stackPos;

160 161 162 163 164 165 166
typedef union {
     /**
      * Most recent retainer for the corresponding closure on the stack.
      */
    retainer c_child_r;
} stackData;

167 168
// Element in the traverse stack, keeps the element, information
// how to continue processing the element, and it's retainer set.
169 170
typedef struct {
    stackPos info;
171 172 173
    StgClosure *c;
    StgClosure *cp; // parent of 'c'
    stackData data;
174 175
} stackElement;

176
typedef struct {
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193
/*
  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.
 */
194 195 196
    bdescr *firstStack;
    bdescr *currentStack;
    stackElement *stackBottom, *stackTop, *stackLimit;
197 198 199 200 201 202 203

/*
  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.
 */
204
    stackElement *currentStackBoundary;
205

206
#if defined(DEBUG_RETAINER)
207 208 209 210 211 212 213 214 215
/*
  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
216
    retainer profiling, maxStackSize is some value no greater
217 218
    than the actual depth of the graph.
 */
219
    int stackSize, maxStackSize;
220
#endif
221 222 223 224 225
} traverseState;

traverseState g_retainerTraverseState;


226
static void retainStack(traverseState *, StgClosure *, stackData, StgPtr, StgPtr);
227
static void retainClosure(traverseState *, StgClosure *, StgClosure *, retainer);
228
static void retainPushClosure(traverseState *, StgClosure *, StgClosure *, stackData);
229 230 231 232 233 234 235
static void retainActualPush(traverseState *, stackElement *);

#if defined(DEBUG_RETAINER)
static void belongToHeap(StgPtr p);
static uint32_t checkHeapSanityForRetainerProfiling( void );
#endif

236 237 238 239 240 241 242 243 244

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

/* -----------------------------------------------------------------------------
 * Add a new block group to the stack.
 * Invariants:
 *  currentStack->link == s.
 * -------------------------------------------------------------------------- */
245
static INLINE void
246
newStackBlock( traverseState *ts, bdescr *bd )
247
{
248 249 250 251 252
    ts->currentStack = bd;
    ts->stackTop     = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
    ts->stackBottom  = (stackElement *)bd->start;
    ts->stackLimit   = (stackElement *)ts->stackTop;
    bd->free     = (StgPtr)ts->stackLimit;
253 254 255 256 257 258 259
}

/* -----------------------------------------------------------------------------
 * Return to the previous block group.
 * Invariants:
 *   s->link == currentStack.
 * -------------------------------------------------------------------------- */
260
static INLINE void
261
returnToOldStack( traverseState *ts, bdescr *bd )
262
{
263 264 265 266 267
    ts->currentStack = bd;
    ts->stackTop = (stackElement *)bd->free;
    ts->stackBottom = (stackElement *)bd->start;
    ts->stackLimit = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
    bd->free = (StgPtr)ts->stackLimit;
268 269 270 271 272 273
}

/* -----------------------------------------------------------------------------
 *  Initializes the traverse stack.
 * -------------------------------------------------------------------------- */
static void
274
initializeTraverseStack( traverseState *ts )
275
{
276 277
    if (ts->firstStack != NULL) {
        freeChain(ts->firstStack);
278 279
    }

280 281 282
    ts->firstStack = allocGroup(BLOCKS_IN_STACK);
    ts->firstStack->link = NULL;
    ts->firstStack->u.back = NULL;
283

284
    newStackBlock(ts, ts->firstStack);
285 286 287 288 289 290 291 292
}

/* -----------------------------------------------------------------------------
 * Frees all the block groups in the traverse stack.
 * Invariants:
 *   firstStack != NULL
 * -------------------------------------------------------------------------- */
static void
293
closeTraverseStack( traverseState *ts )
294
{
295 296
    freeChain(ts->firstStack);
    ts->firstStack = NULL;
297 298 299
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
300
 * Returns true if the whole stack is empty.
301
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
302
static INLINE bool
303
isEmptyRetainerStack( traverseState *ts )
304
{
305
    return (ts->firstStack == ts->currentStack) && ts->stackTop == ts->stackLimit;
306 307
}

sof's avatar
sof committed
308 309 310
/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
311
W_
312
retainerStackBlocks( void )
sof's avatar
sof committed
313 314
{
    bdescr* bd;
315
    W_ res = 0;
316
    traverseState *ts = &g_retainerTraverseState;
sof's avatar
sof committed
317

318
    for (bd = ts->firstStack; bd != NULL; bd = bd->link)
sof's avatar
sof committed
319 320 321 322 323
      res += bd->blocks;

    return res;
}

324
/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
325
 * Returns true if stackTop is at the stack boundary of the current stack,
326 327
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
328
static INLINE bool
329
isOnBoundary( traverseState *ts )
330
{
331
    return ts->stackTop == ts->currentStackBoundary;
332 333 334 335 336 337 338
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
339
static INLINE void
340
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
341 342 343 344 345 346 347 348 349 350
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
351
static INLINE StgClosure *
352 353 354
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
355
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
356
    } else {
357
        return NULL;
358 359 360 361 362 363
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
364
static INLINE void
365
init_srt_fun( stackPos *info, const StgFunInfoTable *infoTable )
366
{
367 368 369
    info->type = posTypeSRT;
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_FUN_SRT(infoTable);
370
    } else {
371
        info->next.srt.srt = NULL;
372
    }
373 374
}

375
static INLINE void
376
init_srt_thunk( stackPos *info, const StgThunkInfoTable *infoTable )
377
{
Ömer Sinan Ağacan's avatar
Ömer Sinan Ağacan committed
378
    info->type = posTypeSRT;
379 380
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_SRT(infoTable);
381
    } else {
382
        info->next.srt.srt = NULL;
383
    }
384 385 386 387 388
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
389
static INLINE StgClosure *
390 391 392
find_srt( stackPos *info )
{
    StgClosure *c;
393
    if (info->type == posTypeSRT) {
394 395 396
        c = info->next.srt.srt;
        info->next.srt.srt = NULL;
        return c;
397 398 399
    }
}

400 401 402 403
/* -----------------------------------------------------------------------------
 * Pushes an element onto traverse stack
 * -------------------------------------------------------------------------- */
static void
404
retainActualPush(traverseState *ts, stackElement *se) {
405
    bdescr *nbd;      // Next Block Descriptor
406
    if (ts->stackTop - 1 < ts->stackBottom) {
407 408 409 410 411
#if defined(DEBUG_RETAINER)
        // debugBelch("push() to the next stack.\n");
#endif
        // currentStack->free is updated when the active stack is switched
        // to the next stack.
412
        ts->currentStack->free = (StgPtr)ts->stackTop;
413

414
        if (ts->currentStack->link == NULL) {
415 416
            nbd = allocGroup(BLOCKS_IN_STACK);
            nbd->link = NULL;
417 418
            nbd->u.back = ts->currentStack;
            ts->currentStack->link = nbd;
419
        } else
420
            nbd = ts->currentStack->link;
421

422
        newStackBlock(ts, nbd);
423 424 425
    }

    // adjust stackTop (acutal push)
426
    ts->stackTop--;
427 428 429 430
    // 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.
431
    *ts->stackTop = *se;
432 433

#if defined(DEBUG_RETAINER)
434 435 436 437
    ts->stackSize++;
    if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
    ASSERT(ts->stackSize >= 0);
    debugBelch("stackSize = %d\n", ts->stackSize);
438
#endif
439

440 441 442 443 444 445
}

/* Push an object onto traverse stack. This method can be used anytime
 * instead of calling retainClosure(), it exists in order to use an
 * explicit stack instead of direct recursion.
 *
446
 *  *cp - object's parent
447 448 449 450
 *  *c - closure
 *  c_child_r - closure retainer.
 */
static INLINE void
451
retainPushClosure( traverseState *ts, StgClosure *c, StgClosure *cp, stackData data) {
452 453 454
    stackElement se;

    se.c = c;
455 456
    se.cp = cp;
    se.data = data;
457 458
    se.info.type = posTypeFresh;

459
    retainActualPush(ts, &se);
460 461
};

462 463 464 465
/* -----------------------------------------------------------------------------
 *  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
466
 *  child, *c_child is set to that child and nothing is pushed onto
467 468 469 470 471 472
 *  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.
473
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
474 475 476
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
477
static INLINE void
478
push( traverseState *ts, StgClosure *c, stackData data, StgClosure **first_child )
479 480 481 482
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
483
#if defined(DEBUG_RETAINER)
484
    debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
485 486 487
#endif

    ASSERT(get_itbl(c)->type != TSO);
488
    ASSERT(get_itbl(c)->type != AP_STACK);
489 490 491 492 493 494

    //
    // fill in se
    //

    se.c = c;
495 496
    se.data = data;
    // Note: se.cp ommitted on purpose, only retainPushClosure uses that.
497 498 499

    // fill in se.info
    switch (get_itbl(c)->type) {
500
        // no child, no SRT
501 502 503
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
gcampax's avatar
gcampax committed
504
    case COMPACT_NFDATA:
505 506
        *first_child = NULL;
        return;
507

508
        // one child (fixed), no SRT
509 510
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
511 512
        *first_child = ((StgMutVar *)c)->var;
        return;
513
    case THUNK_SELECTOR:
514 515
        *first_child = ((StgSelector *)c)->selectee;
        return;
516
    case BLACKHOLE:
517 518
        *first_child = ((StgInd *)c)->indirectee;
        return;
519 520
    case CONSTR_1_0:
    case CONSTR_1_1:
521 522
        *first_child = c->payload[0];
        return;
523

524 525 526
        // 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;
527

528 529
        // two children (fixed), no SRT
        // need to push a stackElement, but nothing to store in se.info
530
    case CONSTR_2_0:
531
        *first_child = c->payload[0];         // return the first pointer
532 533
        se.info.type = posTypeStep;
        se.info.next.step = 2;            // 2 = second
534
        break;
535

536 537
        // three children (fixed), no SRT
        // need to push a stackElement
538 539
    case MVAR_CLEAN:
    case MVAR_DIRTY:
540 541 542
        // 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;
543
        se.info.type = posTypeStep;
544 545 546 547
        se.info.next.step = 2;            // 2 = second
        break;

        // three children (fixed), no SRT
548
    case WEAK:
549
        *first_child = ((StgWeak *)c)->key;
550
        se.info.type = posTypeStep;
551 552
        se.info.next.step = 2;
        break;
553

554
        // layout.payload.ptrs, no SRT
555
    case TVAR:
556
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
557
    case CONSTR_NOCAF:
558
    case PRIM:
559
    case MUT_PRIM:
560
    case BCO:
561 562 563 564 565 566 567 568
        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
569 570
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
571 572
    case MUT_ARR_PTRS_FROZEN_CLEAN:
    case MUT_ARR_PTRS_FROZEN_DIRTY:
573 574 575 576 577 578 579 580
        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
581 582
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
583 584
    case SMALL_MUT_ARR_PTRS_FROZEN_CLEAN:
    case SMALL_MUT_ARR_PTRS_FROZEN_DIRTY:
585 586 587 588 589 590
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;
591

592
    // layout.payload.ptrs, SRT
593
    case FUN_STATIC:
594 595
    case FUN:           // *c is a heap object.
    case FUN_2_0:
596 597 598 599 600 601
        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;
602

603 604
    case THUNK:
    case THUNK_2_0:
605 606 607 608 609 610 611 612 613
        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
614 615
    case FUN_1_0:
    case FUN_1_1:
616 617 618 619
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;
620

621 622
    case THUNK_1_0:
    case THUNK_1_1:
623 624 625 626
        *first_child = ((StgThunk *)c)->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        break;
627

628
    case FUN_0_1:      // *c is a heap object.
629
    case FUN_0_2:
630 631
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
632 633 634 635
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;
636 637 638

    // SRT only
    case THUNK_STATIC:
639
        ASSERT(get_itbl(c)->srt != 0);
640 641
    case THUNK_0_1:
    case THUNK_0_2:
642 643
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
644 645 646 647 648
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

649
    case TREC_CHUNK:
650
        *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
651
        se.info.type = posTypeStep;
652 653
        se.info.next.step = 0;  // entry no.
        break;
654

655
        // cannot appear
656
    case PAP:
657 658
    case AP:
    case AP_STACK:
659
    case TSO:
660
    case STACK:
661
    case IND_STATIC:
662
        // stack objects
663 664
    case UPDATE_FRAME:
    case CATCH_FRAME:
665
    case UNDERFLOW_FRAME:
666 667 668 669
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
670
        // invalid objects
671 672 673
    case IND:
    case INVALID_OBJECT:
    default:
674
        barf("Invalid object *c in push(): %d", get_itbl(c)->type);
675
        return;
676 677
    }

678 679 680 681
    // se.cp has to be initialized when type==posTypeFresh. We don't do that
    // here though. So type must be !=posTypeFresh.
    ASSERT(se.info.type != posTypeFresh);

682
    retainActualPush(ts, &se);
683 684 685 686 687 688 689 690 691 692 693 694
}

/* -----------------------------------------------------------------------------
 *  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
695
 *    INLINE function (for the sake of execution speed).  popOffReal()
696 697 698
 *    is called only within popOff() and nowhere else.
 * -------------------------------------------------------------------------- */
static void
699
popOffReal(traverseState *ts)
700 701 702
{
    bdescr *pbd;    // Previous Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
703
#if defined(DEBUG_RETAINER)
704
    debugBelch("pop() to the previous stack.\n");
705 706
#endif

707 708
    ASSERT(ts->stackTop + 1 == ts->stackLimit);
    ASSERT(ts->stackBottom == (stackElement *)ts->currentStack->start);
709

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

    // currentStack->free is updated when the active stack is switched back
    // to the previous stack.
725
    ts->currentStack->free = (StgPtr)ts->stackLimit;
726 727

    // find the previous block descriptor
728
    pbd = ts->currentStack->u.back;
729 730
    ASSERT(pbd != NULL);

731
    returnToOldStack(ts, pbd);
732

Ben Gamari's avatar
Ben Gamari committed
733
#if defined(DEBUG_RETAINER)
734 735 736 737
    ts->stackSize--;
    if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
    ASSERT(ts->stackSize >= 0);
    debugBelch("stackSize = %d\n", ts->stackSize);
738 739 740
#endif
}

741
static INLINE void
742
popOff(traverseState *ts) {
Ben Gamari's avatar
Ben Gamari committed
743
#if defined(DEBUG_RETAINER)
744
    debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
745 746
#endif

747 748
    ASSERT(ts->stackTop != ts->stackLimit);
    ASSERT(!isEmptyRetainerStack(ts));
749 750

    // <= (instead of <) is wrong!
751 752
    if (ts->stackTop + 1 < ts->stackLimit) {
        ts->stackTop++;
Ben Gamari's avatar
Ben Gamari committed
753
#if defined(DEBUG_RETAINER)
754 755 756 757
        ts->stackSize--;
        if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
        ASSERT(ts->stackSize >= 0);
        debugBelch("stackSize = %d\n", ts->stackSize);
758
#endif
759
        return;
760 761
    }

762
    popOffReal(ts);
763 764 765 766 767
}

/* -----------------------------------------------------------------------------
 *  Finds the next object to be considered for retainer profiling and store
 *  its pointer to *c.
768 769 770
 *  If the unprocessed object was stored in the stack (posTypeFresh), the
 *  this object is returned as-is. Otherwise Test if the topmost stack
 *  element indicates that more objects are left,
771
 *  and if so, retrieve the first object and store its pointer to *c. Also,
772 773 774
 *  set *cp and *data 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.
775 776
 *  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
777
 *  the current stack chunk becomes empty, indicated by true returned by
778 779 780 781 782
 *  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.
 * -------------------------------------------------------------------------- */
783
static INLINE void
784
pop( traverseState *ts, StgClosure **c, StgClosure **cp, stackData *data )
785 786 787
{
    stackElement *se;

Ben Gamari's avatar
Ben Gamari committed
788
#if defined(DEBUG_RETAINER)
789
    debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
790 791 792
#endif

    do {
793
        if (isOnBoundary(ts)) {     // if the current stack chunk is depleted
794 795 796 797
            *c = NULL;
            return;
        }

798
        se = ts->stackTop;
799

800 801
        // If this is a top-level element, you should pop that out.
        if (se->info.type == posTypeFresh) {
802
            *cp = se->cp;
803
            *c = se->c;
804
            *data = se->data;
805
            popOff(ts);
806 807 808
            return;
        }

809 810 811 812 813 814
        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;
815
            *data = se->data;
816
            popOff(ts);
817 818 819 820
            return;

            // three children (fixed), no SRT
            // need to push a stackElement
821 822
        case MVAR_CLEAN:
        case MVAR_DIRTY:
823 824 825 826 827 828
            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;
829
                popOff(ts);
830 831
            }
            *cp = se->c;
832
            *data = se->data;
833 834 835 836 837 838 839 840 841 842
            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;
843
                popOff(ts);
844 845
            }
            *cp = se->c;
846
            *data = se->data;
847 848 849 850 851 852 853 854 855
            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;
856 857
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
858 859
            if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
                *c = NULL;
860
                popOff(ts);
861
                break;
862 863 864 865 866 867 868 869 870 871
            }
            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;
872
            *data = se->data;
873 874 875
            se->info.next.step++;
            return;
        }
876

877 878
        case TVAR:
        case CONSTR:
879 880 881 882 883 884
        case PRIM:
        case MUT_PRIM:
        case BCO:
            // StgMutArrPtr.ptrs, no SRT
        case MUT_ARR_PTRS_CLEAN:
        case MUT_ARR_PTRS_DIRTY:
885 886
        case MUT_ARR_PTRS_FROZEN_CLEAN:
        case MUT_ARR_PTRS_FROZEN_DIRTY:
887 888 889 890
        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:
891 892
            *c = find_ptrs(&se->info);
            if (*c == NULL) {
893
                popOff(ts);
894 895 896
                break;
            }
            *cp = se->c;
897
            *data = se->data;
898 899 900 901
            return;

            // layout.payload.ptrs, SRT
        case FUN:         // always a heap object
902
        case FUN_STATIC:
903 904 905 906 907
        case FUN_2_0:
            if (se->info.type == posTypePtrs) {
                *c = find_ptrs(&se->info);
                if (*c != NULL) {
                    *cp = se->c;
908
                    *data = se->data;
909 910 911 912 913 914 915 916 917 918 919 920
                    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;
921
                    *data = se->data;
922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941
                    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;
942
                *data = se->data;
943 944
                return;
            }
945
            popOff(ts);
946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961
            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:
962 963
        case STACK:
        case IND_STATIC:
Simon Marlow's avatar
Simon Marlow committed
964
        case CONSTR_NOCAF:
965
            // stack objects
966
        case UPDATE_FRAME:
967
        case CATCH_FRAME:
968 969
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
970 971 972 973 974 975 976
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default:
977
            barf("Invalid object *c in pop(): %d", get_itbl(se->c)->type);
978 979
            return;
        }
Ben Gamari's avatar
Ben Gamari committed
980
    } while (true);
981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999
}

/* -----------------------------------------------------------------------------
 * 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
1000
#if defined(SECOND_APPROACH)
1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
    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.
 * -------------------------------------------------------------------------- */
1019
static INLINE void
1020 1021 1022
maybeInitRetainerSet( StgClosure *c )
{
    if (!isRetainerSetFieldValid(c)) {
1023
        setRetainerSetToNull(c);
1024 1025 1026 1027
    }
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
1028
 * Returns true if *c is a retainer.
1029 1030 1031 1032 1033 1034
 * In general the retainers are the objects that may be the roots of the
 * collection. Basically this roots represents programmers threads
 * (TSO) with their stack and thunks.
 *
 * In addition we mark all mutable objects as a retainers, the reason for
 * that decision is lost in time.
1035
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
1036
static INLINE bool
1037 1038 1039
isRetainer( StgClosure *c )
{
    switch (get_itbl(c)->type) {
1040 1041 1042 1043
        //
        //  True case
        //
        // TSOs MUST be retainers: they constitute the set of roots.
1044
    case TSO:
1045
    case STACK:
1046

1047
        // mutable objects
1048
    case MUT_PRIM:
1049 1050
    case MVAR_CLEAN:
    case MVAR_DIRTY:
1051
    case TVAR:
1052 1053
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
1054 1055
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
1056 1057 1058
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case BLOCKING_QUEUE:
1059

1060
        // thunks are retainers.
1061 1062 1063 1064 1065 1066 1067
    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:
1068 1069
    case AP:
    case AP_STACK:
1070

1071
        // Static thunks, or CAFS, are obviously retainers.
1072 1073
    case THUNK_STATIC:

1074 1075
        // WEAK objects are roots; there is separate code in which traversing
        // begins from WEAK objects.
1076
    case WEAK:
Ben Gamari's avatar
Ben Gamari committed
1077
        return true;
1078

1079 1080 1081
        //
        // False case
        //
1082

1083
        // constructors
1084
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
1085
    case CONSTR_NOCAF:
1086 1087 1088 1089 1090
    case CONSTR_1_0:
    case CONSTR_0_1:
    case CONSTR_2_0:
    case CONSTR_1_1:
    case CONSTR_0_2:
1091
        // functions
1092 1093 1094 1095 1096 1097
    case FUN:
    case FUN_1_0:
    case FUN_0_1:
    case FUN_2_0:
    case FUN_1_1:
    case FUN_0_2:
1098
        // partial applications
1099
    case PAP:
1100
        // indirection
Ian Lynagh's avatar
Ian Lynagh committed
1101 1102 1103 1104
    // 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:
1105
    case BLACKHOLE:
1106
    case WHITEHOLE:
1107
        // static objects
1108
    case FUN_STATIC:
1109
        // misc
1110
    case PRIM:
1111 1112
    case BCO:
    case ARR_WORDS:
1113
    case COMPACT_NFDATA:
1114
        // STM
1115
    case TREC_CHUNK:
1116
        // immutable arrays
1117 1118 1119 1120
    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
1121
        return false;
1122

1123 1124 1125 1126 1127
        //
        // Error case
        //
        // Stack objects are invalid because they are never treated as
        // legal objects during retainer profiling.
1128 1129
    case UPDATE_FRAME:
    case CATCH_FRAME:
1130 1131
    case CATCH_RETRY_FRAME:
    case CATCH_STM_FRAME: