RetainerProfile.c 72.5 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
Retainer profiling is a profiling technique that gives information why
objects can't be freed and lists the consumers that hold pointers to
41
the heap objects. It does not list all the objects that keep references
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
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Exact algorithm and additional information can be found the historical
document 'docs/storage-mgt/rp.tex'.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RTS implementation may be out of date, but the general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 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
/** Note [Profiling heap traversal visited bit]
 *
 * If the RTS is compiled with profiling enabled StgProfHeader can be used by
 * profiling code to store per-heap object information.
 *
 * When using the generic heap traversal code we use this field to store
 * profiler specific information. However we reserve the LSB of the *entire*
 * 'trav' union (which will overlap with the other fields) for the generic
 * traversal code. We use the bit to decide whether we've already visited this
 * closure in this pass or not. We do this as the heap may contain cyclic
 * references, it being a graph and all, so we would likely just infinite loop
 * if we didn't.
 *
 * We assume that at least the LSB of the largest field in the corresponding
 * union is insignificant. This is true at least for the word aligned pointers
 * which the retainer profiler currently stores there and should be maintained
 * by new users of the 'trav' union.
 *
 * Now the way the traversal works is that the interpretation of the "visited?"
 * bit depends on the value of the global 'flip' variable. We don't want to have
 * to do another pass over the heap just to reset the bit to zero so instead on
 * each traversal (i.e. each run of the profiling code) we invert the value of
 * the global 'flip' variable. We interpret this as resetting all the "visited?"
 * flags on the heap.
 *
 * There is one exception to this rule, namely: static objects. There we do just
 * go over the heap and reset the bit manually. See
 * 'resetStaticObjectForRetainerProfiling'.
112
 */
113
StgWord flip = 0;     // flip bit
114 115
                      // must be 0 if DEBUG_RETAINER is on (for static closures)

116 117
#define setTravDataToZero(c) \
  (c)->header.prof.hp.trav.lsb = flip
118

Ben Gamari's avatar
Ben Gamari committed
119
#if defined(DEBUG_RETAINER)
120
static uint32_t sumOfNewCost;        // sum of the cost of each object, computed
121
                                // when the object is first visited
122
static uint32_t sumOfNewCostExtra;   // for those objects not visited during
123
                                // retainer profiling, e.g., MUT_VAR
124
static uint32_t costArray[N_CLOSURE_TYPES];
125

126
uint32_t sumOfCostLinear;            // sum of the costs of all object, computed
127 128
                                // when linearly traversing the heap after
                                // retainer profiling
129
uint32_t costArrayLinear[N_CLOSURE_TYPES];
130 131 132 133 134 135 136 137 138 139 140 141 142
#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 {
143 144
    // Object with fixed layout. Keeps an information about that
    // element was processed. (stackPos.next.step)
145
    posTypeStep,
146 147
    // Description of the pointers-first heap object. Keeps information
    // about layout. (stackPos.next.ptrs)
148
    posTypePtrs,
149
    // Keeps SRT bitmap (stackPos.next.srt)
150
    posTypeSRT,
151 152 153
    // Keeps a new object that was not inspected yet. Keeps a parent
    // element (stackPos.next.parent)
    posTypeFresh
154 155 156 157 158 159 160 161
} nextPosType;

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

    // layout.payload
    struct {
162 163 164
        // See StgClosureInfo in InfoTables.h
        StgHalfWord pos;
        StgHalfWord ptrs;
165
        StgPtr payload;
166 167 168 169
    } ptrs;

    // SRT
    struct {
170
        StgClosure *srt;
171 172 173
    } srt;
} nextPos;

174 175
// Tagged stack element, that keeps information how to process
// the next element in the traverse stack.
176 177 178 179 180
typedef struct {
    nextPosType type;
    nextPos next;
} stackPos;

181 182 183 184 185 186 187
typedef union {
     /**
      * Most recent retainer for the corresponding closure on the stack.
      */
    retainer c_child_r;
} stackData;

188 189
// Element in the traverse stack, keeps the element, information
// how to continue processing the element, and it's retainer set.
190 191
typedef struct {
    stackPos info;
192 193 194
    StgClosure *c;
    StgClosure *cp; // parent of 'c'
    stackData data;
195 196
} stackElement;

197
typedef struct {
198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
/*
  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.
 */
215 216 217
    bdescr *firstStack;
    bdescr *currentStack;
    stackElement *stackBottom, *stackTop, *stackLimit;
218 219 220 221 222 223 224

/*
  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.
 */
225
    stackElement *currentStackBoundary;
226

227
#if defined(DEBUG_RETAINER)
228 229 230 231 232 233 234 235 236
/*
  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
237
    retainer profiling, maxStackSize is some value no greater
238 239
    than the actual depth of the graph.
 */
240
    int stackSize, maxStackSize;
241
#endif
242 243 244 245 246
} traverseState;

traverseState g_retainerTraverseState;


247
static void retainStack(traverseState *, StgClosure *, stackData, StgPtr, StgPtr);
248
static void retainClosure(traverseState *, StgClosure *, StgClosure *, retainer);
249
static void retainPushClosure(traverseState *, StgClosure *, StgClosure *, stackData);
250 251 252 253 254 255 256
static void retainActualPush(traverseState *, stackElement *);

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

257 258 259 260 261 262 263 264 265

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

/* -----------------------------------------------------------------------------
 * Add a new block group to the stack.
 * Invariants:
 *  currentStack->link == s.
 * -------------------------------------------------------------------------- */
266
static INLINE void
267
newStackBlock( traverseState *ts, bdescr *bd )
268
{
269 270 271 272 273
    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;
274 275 276 277 278 279 280
}

/* -----------------------------------------------------------------------------
 * Return to the previous block group.
 * Invariants:
 *   s->link == currentStack.
 * -------------------------------------------------------------------------- */
281
static INLINE void
282
returnToOldStack( traverseState *ts, bdescr *bd )
283
{
284 285 286 287 288
    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;
289 290 291 292 293 294
}

/* -----------------------------------------------------------------------------
 *  Initializes the traverse stack.
 * -------------------------------------------------------------------------- */
static void
295
initializeTraverseStack( traverseState *ts )
296
{
297 298
    if (ts->firstStack != NULL) {
        freeChain(ts->firstStack);
299 300
    }

301 302 303
    ts->firstStack = allocGroup(BLOCKS_IN_STACK);
    ts->firstStack->link = NULL;
    ts->firstStack->u.back = NULL;
304

305
    newStackBlock(ts, ts->firstStack);
306 307 308 309 310 311 312 313
}

/* -----------------------------------------------------------------------------
 * Frees all the block groups in the traverse stack.
 * Invariants:
 *   firstStack != NULL
 * -------------------------------------------------------------------------- */
static void
314
closeTraverseStack( traverseState *ts )
315
{
316 317
    freeChain(ts->firstStack);
    ts->firstStack = NULL;
318 319 320
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
321
 * Returns true if the whole stack is empty.
322
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
323
static INLINE bool
324
isEmptyRetainerStack( traverseState *ts )
325
{
326
    return (ts->firstStack == ts->currentStack) && ts->stackTop == ts->stackLimit;
327 328
}

sof's avatar
sof committed
329 330 331
/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
332
W_
333
retainerStackBlocks( void )
sof's avatar
sof committed
334 335
{
    bdescr* bd;
336
    W_ res = 0;
337
    traverseState *ts = &g_retainerTraverseState;
sof's avatar
sof committed
338

339
    for (bd = ts->firstStack; bd != NULL; bd = bd->link)
sof's avatar
sof committed
340 341 342 343 344
      res += bd->blocks;

    return res;
}

345
/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
346
 * Returns true if stackTop is at the stack boundary of the current stack,
347 348
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
349
static INLINE bool
350
isOnBoundary( traverseState *ts )
351
{
352
    return ts->stackTop == ts->currentStackBoundary;
353 354 355 356 357 358 359
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
360
static INLINE void
361
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
362 363 364 365 366 367 368 369 370 371
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
372
static INLINE StgClosure *
373 374 375
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
376
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
377
    } else {
378
        return NULL;
379 380 381 382 383 384
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
385
static INLINE void
386
init_srt_fun( stackPos *info, const StgFunInfoTable *infoTable )
387
{
388 389 390
    info->type = posTypeSRT;
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_FUN_SRT(infoTable);
391
    } else {
392
        info->next.srt.srt = NULL;
393
    }
394 395
}

396
static INLINE void
397
init_srt_thunk( stackPos *info, const StgThunkInfoTable *infoTable )
398
{
Ömer Sinan Ağacan's avatar
Ömer Sinan Ağacan committed
399
    info->type = posTypeSRT;
400 401
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_SRT(infoTable);
402
    } else {
403
        info->next.srt.srt = NULL;
404
    }
405 406 407 408 409
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
410
static INLINE StgClosure *
411 412 413
find_srt( stackPos *info )
{
    StgClosure *c;
414
    if (info->type == posTypeSRT) {
415 416 417
        c = info->next.srt.srt;
        info->next.srt.srt = NULL;
        return c;
418 419 420
    }
}

421 422 423 424
/* -----------------------------------------------------------------------------
 * Pushes an element onto traverse stack
 * -------------------------------------------------------------------------- */
static void
425
retainActualPush(traverseState *ts, stackElement *se) {
426
    bdescr *nbd;      // Next Block Descriptor
427
    if (ts->stackTop - 1 < ts->stackBottom) {
428 429 430 431 432
#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.
433
        ts->currentStack->free = (StgPtr)ts->stackTop;
434

435
        if (ts->currentStack->link == NULL) {
436 437
            nbd = allocGroup(BLOCKS_IN_STACK);
            nbd->link = NULL;
438 439
            nbd->u.back = ts->currentStack;
            ts->currentStack->link = nbd;
440
        } else
441
            nbd = ts->currentStack->link;
442

443
        newStackBlock(ts, nbd);
444 445 446
    }

    // adjust stackTop (acutal push)
447
    ts->stackTop--;
448 449 450 451
    // 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.
452
    *ts->stackTop = *se;
453 454

#if defined(DEBUG_RETAINER)
455 456 457 458
    ts->stackSize++;
    if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
    ASSERT(ts->stackSize >= 0);
    debugBelch("stackSize = %d\n", ts->stackSize);
459
#endif
460

461 462 463 464 465 466
}

/* 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.
 *
467
 *  *cp - object's parent
468 469 470 471
 *  *c - closure
 *  c_child_r - closure retainer.
 */
static INLINE void
472
retainPushClosure( traverseState *ts, StgClosure *c, StgClosure *cp, stackData data) {
473 474 475
    stackElement se;

    se.c = c;
476 477
    se.cp = cp;
    se.data = data;
478 479
    se.info.type = posTypeFresh;

480
    retainActualPush(ts, &se);
481 482
};

483 484 485 486
/* -----------------------------------------------------------------------------
 *  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
487
 *  child, *c_child is set to that child and nothing is pushed onto
488 489 490 491 492 493
 *  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.
494
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
495 496 497
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
498
static INLINE void
499
push( traverseState *ts, StgClosure *c, stackData data, StgClosure **first_child )
500 501 502 503
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
504
#if defined(DEBUG_RETAINER)
505
    debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
506 507 508
#endif

    ASSERT(get_itbl(c)->type != TSO);
509
    ASSERT(get_itbl(c)->type != AP_STACK);
510 511 512 513 514 515

    //
    // fill in se
    //

    se.c = c;
516 517
    se.data = data;
    // Note: se.cp ommitted on purpose, only retainPushClosure uses that.
518 519 520

    // fill in se.info
    switch (get_itbl(c)->type) {
521
        // no child, no SRT
522 523 524
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
gcampax's avatar
gcampax committed
525
    case COMPACT_NFDATA:
526 527
        *first_child = NULL;
        return;
528

529
        // one child (fixed), no SRT
530 531
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
532 533
        *first_child = ((StgMutVar *)c)->var;
        return;
534
    case THUNK_SELECTOR:
535 536
        *first_child = ((StgSelector *)c)->selectee;
        return;
537
    case BLACKHOLE:
538 539
        *first_child = ((StgInd *)c)->indirectee;
        return;
540 541
    case CONSTR_1_0:
    case CONSTR_1_1:
542 543
        *first_child = c->payload[0];
        return;
544

545 546 547
        // 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;
548

549 550
        // two children (fixed), no SRT
        // need to push a stackElement, but nothing to store in se.info
551
    case CONSTR_2_0:
552
        *first_child = c->payload[0];         // return the first pointer
553 554
        se.info.type = posTypeStep;
        se.info.next.step = 2;            // 2 = second
555
        break;
556

557 558
        // three children (fixed), no SRT
        // need to push a stackElement
559 560
    case MVAR_CLEAN:
    case MVAR_DIRTY:
561 562 563
        // 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;
564
        se.info.type = posTypeStep;
565 566 567 568
        se.info.next.step = 2;            // 2 = second
        break;

        // three children (fixed), no SRT
569
    case WEAK:
570
        *first_child = ((StgWeak *)c)->key;
571
        se.info.type = posTypeStep;
572 573
        se.info.next.step = 2;
        break;
574

575
        // layout.payload.ptrs, no SRT
576
    case TVAR:
577
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
578
    case CONSTR_NOCAF:
579
    case PRIM:
580
    case MUT_PRIM:
581
    case BCO:
582 583 584 585 586 587 588 589
        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
590 591
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
592 593
    case MUT_ARR_PTRS_FROZEN_CLEAN:
    case MUT_ARR_PTRS_FROZEN_DIRTY:
594 595 596 597 598 599 600 601
        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
602 603
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
604 605
    case SMALL_MUT_ARR_PTRS_FROZEN_CLEAN:
    case SMALL_MUT_ARR_PTRS_FROZEN_DIRTY:
606 607 608 609 610 611
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;
612

613
    // layout.payload.ptrs, SRT
614
    case FUN_STATIC:
615 616
    case FUN:           // *c is a heap object.
    case FUN_2_0:
617 618 619 620 621 622
        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;
623

624 625
    case THUNK:
    case THUNK_2_0:
626 627 628 629 630 631 632 633 634
        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
635 636
    case FUN_1_0:
    case FUN_1_1:
637 638 639 640
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;
641

642 643
    case THUNK_1_0:
    case THUNK_1_1:
644 645 646 647
        *first_child = ((StgThunk *)c)->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        break;
648

649
    case FUN_0_1:      // *c is a heap object.
650
    case FUN_0_2:
651 652
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
653 654 655 656
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;
657 658 659

    // SRT only
    case THUNK_STATIC:
660
        ASSERT(get_itbl(c)->srt != 0);
661 662
    case THUNK_0_1:
    case THUNK_0_2:
663 664
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
665 666 667 668 669
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

670
    case TREC_CHUNK:
671
        *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
672
        se.info.type = posTypeStep;
673 674
        se.info.next.step = 0;  // entry no.
        break;
675

676
        // cannot appear
677
    case PAP:
678 679
    case AP:
    case AP_STACK:
680
    case TSO:
681
    case STACK:
682
    case IND_STATIC:
683
        // stack objects
684 685
    case UPDATE_FRAME:
    case CATCH_FRAME:
686
    case UNDERFLOW_FRAME:
687 688 689 690
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
691
        // invalid objects
692 693 694
    case IND:
    case INVALID_OBJECT:
    default:
695
        barf("Invalid object *c in push(): %d", get_itbl(c)->type);
696
        return;
697 698
    }

699 700 701 702
    // 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);

703
    retainActualPush(ts, &se);
704 705 706 707 708 709 710 711 712 713 714 715
}

/* -----------------------------------------------------------------------------
 *  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
716
 *    INLINE function (for the sake of execution speed).  popOffReal()
717 718 719
 *    is called only within popOff() and nowhere else.
 * -------------------------------------------------------------------------- */
static void
720
popOffReal(traverseState *ts)
721 722 723
{
    bdescr *pbd;    // Previous Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
724
#if defined(DEBUG_RETAINER)
725
    debugBelch("pop() to the previous stack.\n");
726 727
#endif

728 729
    ASSERT(ts->stackTop + 1 == ts->stackLimit);
    ASSERT(ts->stackBottom == (stackElement *)ts->currentStack->start);
730

731
    if (ts->firstStack == ts->currentStack) {
732
        // The stack is completely empty.
733 734
        ts->stackTop++;
        ASSERT(ts->stackTop == ts->stackLimit);
Ben Gamari's avatar
Ben Gamari committed
735
#if defined(DEBUG_RETAINER)
736 737 738 739
        ts->stackSize--;
        if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
        ASSERT(ts->stackSize >= 0);
        debugBelch("stackSize = %d\n", ts->stackSize);
740
#endif
741
        return;
742 743 744 745
    }

    // currentStack->free is updated when the active stack is switched back
    // to the previous stack.
746
    ts->currentStack->free = (StgPtr)ts->stackLimit;
747 748

    // find the previous block descriptor
749
    pbd = ts->currentStack->u.back;
750 751
    ASSERT(pbd != NULL);

752
    returnToOldStack(ts, pbd);
753

Ben Gamari's avatar
Ben Gamari committed
754
#if defined(DEBUG_RETAINER)
755 756 757 758
    ts->stackSize--;
    if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
    ASSERT(ts->stackSize >= 0);
    debugBelch("stackSize = %d\n", ts->stackSize);
759 760 761
#endif
}

762
static INLINE void
763
popOff(traverseState *ts) {
Ben Gamari's avatar
Ben Gamari committed
764
#if defined(DEBUG_RETAINER)
765
    debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
766 767
#endif

768 769
    ASSERT(ts->stackTop != ts->stackLimit);
    ASSERT(!isEmptyRetainerStack(ts));
770 771

    // <= (instead of <) is wrong!
772 773
    if (ts->stackTop + 1 < ts->stackLimit) {
        ts->stackTop++;
Ben Gamari's avatar
Ben Gamari committed
774
#if defined(DEBUG_RETAINER)
775 776 777 778
        ts->stackSize--;
        if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
        ASSERT(ts->stackSize >= 0);
        debugBelch("stackSize = %d\n", ts->stackSize);
779
#endif
780
        return;
781 782
    }

783
    popOffReal(ts);
784 785 786 787 788
}

/* -----------------------------------------------------------------------------
 *  Finds the next object to be considered for retainer profiling and store
 *  its pointer to *c.
789 790 791
 *  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,
792
 *  and if so, retrieve the first object and store its pointer to *c. Also,
793 794 795
 *  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.
796 797
 *  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
798
 *  the current stack chunk becomes empty, indicated by true returned by
799 800 801 802 803
 *  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.
 * -------------------------------------------------------------------------- */
804
static INLINE void
805
pop( traverseState *ts, StgClosure **c, StgClosure **cp, stackData *data )
806 807 808
{
    stackElement *se;

Ben Gamari's avatar
Ben Gamari committed
809
#if defined(DEBUG_RETAINER)
810
    debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
811 812 813
#endif

    do {
814
        if (isOnBoundary(ts)) {     // if the current stack chunk is depleted
815 816 817 818
            *c = NULL;
            return;
        }

819
        se = ts->stackTop;
820

821 822
        // If this is a top-level element, you should pop that out.
        if (se->info.type == posTypeFresh) {
823
            *cp = se->cp;
824
            *c = se->c;
825
            *data = se->data;
826
            popOff(ts);
827 828 829
            return;
        }

830 831 832 833 834 835
        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;
836
            *data = se->data;
837
            popOff(ts);
838 839 840 841
            return;

            // three children (fixed), no SRT
            // need to push a stackElement
842 843
        case MVAR_CLEAN:
        case MVAR_DIRTY:
844 845 846 847 848 849
            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;
850
                popOff(ts);
851 852
            }
            *cp = se->c;
853
            *data = se->data;
854 855 856 857 858 859 860 861 862 863
            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;
864
                popOff(ts);
865 866
            }
            *cp = se->c;
867
            *data = se->data;
868 869 870 871 872 873 874 875 876
            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;
877 878
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
879 880
            if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
                *c = NULL;
881
                popOff(ts);
882
                break;
883 884 885 886 887 888 889 890 891 892
            }
            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;
893
            *data = se->data;
894 895 896
            se->info.next.step++;
            return;
        }
897

898 899
        case TVAR:
        case CONSTR:
900 901 902 903 904 905
        case PRIM:
        case MUT_PRIM:
        case BCO:
            // StgMutArrPtr.ptrs, no SRT
        case MUT_ARR_PTRS_CLEAN:
        case MUT_ARR_PTRS_DIRTY:
906 907
        case MUT_ARR_PTRS_FROZEN_CLEAN:
        case MUT_ARR_PTRS_FROZEN_DIRTY:
908 909 910 911
        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:
912 913
            *c = find_ptrs(&se->info);
            if (*c == NULL) {
914
                popOff(ts);
915 916 917
                break;
            }
            *cp = se->c;
918
            *data = se->data;
919 920 921 922
            return;

            // layout.payload.ptrs, SRT
        case FUN:         // always a heap object
923
        case FUN_STATIC:
924 925 926 927 928
        case FUN_2_0:
            if (se->info.type == posTypePtrs) {
                *c = find_ptrs(&se->info);
                if (*c != NULL) {
                    *cp = se->c;
929
                    *data = se->data;
930 931 932 933 934 935 936 937 938 939 940 941
                    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;
942
                    *data = se->data;
943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962
                    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;
963
                *data = se->data;
964 965
                return;
            }
966
            popOff(ts);
967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982
            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:
983 984
        case STACK:
        case IND_STATIC:
Simon Marlow's avatar
Simon Marlow committed
985
        case CONSTR_NOCAF:
986
            // stack objects
987
        case UPDATE_FRAME:
988
        case CATCH_FRAME:
989 990
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
991 992 993 994 995 996 997
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default:
998
            barf("Invalid object *c in pop(): %d", get_itbl(se->c)->type);
999 1000
            return;
        }
Ben Gamari's avatar
Ben Gamari committed
1001
    } while (true);
1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020
}

/* -----------------------------------------------------------------------------
 * 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
1021
#if defined(SECOND_APPROACH)
1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039
    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.
 * -------------------------------------------------------------------------- */
1040
static INLINE void
1041
maybeInitTravData( StgClosure *c )
1042
{
1043 1044
    if (!isTravDataValid(c)) {
        setTravDataToZero(c);
1045 1046 1047 1048
    }
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
1049
 * Returns true if *c is a retainer.
1050 1051 1052 1053 1054 1055
 * 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.
1056
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
1057
static INLINE bool
1058 1059 1060
isRetainer( StgClosure *c )
{
    switch (get_itbl(c)->type) {
1061 1062 1063 1064
        //
        //  True case
        //
        // TSOs MUST be retainers: they constitute the set of roots.
1065
    case TSO:
1066
    case STACK:
1067

1068
        // mutable objects
1069
    case MUT_PRIM:
1070 1071
    case MVAR_CLEAN:
    case MVAR_DIRTY:
1072
    case TVAR:
1073 1074
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
1075 1076
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
1077 1078 1079
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case BLOCKING_QUEUE:
1080

1081
        // thunks are retainers.
1082 1083 1084 1085 1086 1087 1088
    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:
1089 1090
    case AP:
    case AP_STACK:
1091

1092
        // Static thunks, or CAFS, are obviously retainers.
1093 1094
    case THUNK_STATIC:

1095 1096
        // WEAK objects are roots; there is separate code in which traversing
        // begins from WEAK objects.
1097
    case WEAK:
Ben Gamari's avatar
Ben Gamari committed
1098
        return true;
1099

1100 1101 1102
        //
        // False case
        //
1103

1104
        // constructors
1105
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
1106
    case CONSTR_NOCAF:
1107 1108 1109 1110 1111
    case CONSTR_1_0:
    case CONSTR_0_1:
    case CONSTR_2_0:
    case CONSTR_1_1:
    case CONSTR_0_2:
1112
        // functions
1113 1114 1115 1116 1117 1118
    case FUN:
    case FUN_1_0:
    case FUN_0_1:
    case FUN_2_0:
    case FUN_1_1:
    case FUN_0_2:
1119
        // partial applications
1120
    case PAP:
1121
        // indirection
Ian Lynagh's avatar
Ian Lynagh committed
1122 1123 1124 1125
    // 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:
1126
    case BLACKHOLE:
1127
    case WHITEHOLE:
1128
        // static objects
1129
    case FUN_STATIC:
1130
        // misc
1131
    case PRIM:
1132 1133
    case BCO:
    case ARR_WORDS:
1134
    case COMPACT_NFDATA:
1135
        // STM