RetainerProfile.c 61.3 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Exact algorithm and additional information can be found the historical
document 'docs/storage-mgt/rp.tex'.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RTS implementation may be out of date, but the general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
/** 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
111
 * 'resetStaticObjectForProfiling'.
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
118 119 120 121 122 123 124 125 126 127 128 129

/* -----------------------------------------------------------------------------
 * 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 {
130 131
    // Object with fixed layout. Keeps an information about that
    // element was processed. (stackPos.next.step)
132
    posTypeStep,
133 134
    // Description of the pointers-first heap object. Keeps information
    // about layout. (stackPos.next.ptrs)
135
    posTypePtrs,
136
    // Keeps SRT bitmap (stackPos.next.srt)
137
    posTypeSRT,
138 139 140
    // Keeps a new object that was not inspected yet. Keeps a parent
    // element (stackPos.next.parent)
    posTypeFresh
141 142 143 144 145 146 147 148
} nextPosType;

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

    // layout.payload
    struct {
149 150 151
        // See StgClosureInfo in InfoTables.h
        StgHalfWord pos;
        StgHalfWord ptrs;
152
        StgPtr payload;
153 154 155 156
    } ptrs;

    // SRT
    struct {
157
        StgClosure *srt;
158 159 160
    } srt;
} nextPos;

161 162
// Tagged stack element, that keeps information how to process
// the next element in the traverse stack.
163 164 165 166 167
typedef struct {
    nextPosType type;
    nextPos next;
} stackPos;

168 169 170 171 172 173 174
typedef union {
     /**
      * Most recent retainer for the corresponding closure on the stack.
      */
    retainer c_child_r;
} stackData;

175 176
// Element in the traverse stack, keeps the element, information
// how to continue processing the element, and it's retainer set.
177 178
typedef struct {
    stackPos info;
179 180 181
    StgClosure *c;
    StgClosure *cp; // parent of 'c'
    stackData data;
182 183
} stackElement;

184
typedef struct {
185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
/*
  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.
 */
202 203 204
    bdescr *firstStack;
    bdescr *currentStack;
    stackElement *stackBottom, *stackTop, *stackLimit;
205 206 207 208 209 210 211

/*
  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.
 */
212
    stackElement *currentStackBoundary;
213 214 215 216 217 218 219 220 221 222

/*
  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
223
    retainer profiling, maxStackSize is some value no greater
224 225
    than the actual depth of the graph.
 */
226 227 228
    int stackSize, maxStackSize;
} traverseState;

229 230 231 232 233 234 235 236 237 238 239 240 241 242 243
/* Callback called when heap traversal visits a closure.
 *
 * Before this callback is called the profiling header of the visited closure
 * 'c' is zero'd with 'setTravDataToZero' if this closure hasn't been visited in
 * this run yet. See Note [Profiling heap traversal visited bit].
 *
 * Return 'true' when this is not the first visit to this element. The generic
 * traversal code will then skip traversing the children.
 */
typedef bool (*visitClosure_cb) (
    const StgClosure *c,
    const StgClosure *cp,
    const stackData data,
    stackData *child_data);

244 245 246
traverseState g_retainerTraverseState;


247 248 249
static void traverseStack(traverseState *, StgClosure *, stackData, StgPtr, StgPtr);
static void traverseClosure(traverseState *, StgClosure *, StgClosure *, retainer);
static void traversePushClosure(traverseState *, StgClosure *, StgClosure *, stackData);
250

251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266
#if defined(DEBUG)
unsigned int g_traversalDebugLevel = 0;
static inline void debug(const char *s, ...)
{
    va_list ap;

    if(g_traversalDebugLevel == 0)
        return;

    va_start(ap,s);
    vdebugBelch(s, ap);
    va_end(ap);
}
#else
#define debug(...)
#endif
267 268 269 270 271 272 273 274 275

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

/* -----------------------------------------------------------------------------
 * Add a new block group to the stack.
 * Invariants:
 *  currentStack->link == s.
 * -------------------------------------------------------------------------- */
276
static INLINE void
277
newStackBlock( traverseState *ts, bdescr *bd )
278
{
279 280 281 282 283
    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;
284 285 286 287 288 289 290
}

/* -----------------------------------------------------------------------------
 * Return to the previous block group.
 * Invariants:
 *   s->link == currentStack.
 * -------------------------------------------------------------------------- */
291
static INLINE void
292
returnToOldStack( traverseState *ts, bdescr *bd )
293
{
294 295 296 297 298
    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;
299 300 301 302 303 304
}

/* -----------------------------------------------------------------------------
 *  Initializes the traverse stack.
 * -------------------------------------------------------------------------- */
static void
305
initializeTraverseStack( traverseState *ts )
306
{
307 308
    if (ts->firstStack != NULL) {
        freeChain(ts->firstStack);
309 310
    }

311 312 313
    ts->firstStack = allocGroup(BLOCKS_IN_STACK);
    ts->firstStack->link = NULL;
    ts->firstStack->u.back = NULL;
314

315
    newStackBlock(ts, ts->firstStack);
316 317 318 319 320 321 322 323
}

/* -----------------------------------------------------------------------------
 * Frees all the block groups in the traverse stack.
 * Invariants:
 *   firstStack != NULL
 * -------------------------------------------------------------------------- */
static void
324
closeTraverseStack( traverseState *ts )
325
{
326 327
    freeChain(ts->firstStack);
    ts->firstStack = NULL;
328 329 330
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
331
 * Returns true if the whole stack is empty.
332
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
333
static INLINE bool
334
isEmptyWorkStack( traverseState *ts )
335
{
336
    return (ts->firstStack == ts->currentStack) && ts->stackTop == ts->stackLimit;
337 338
}

sof's avatar
sof committed
339 340 341
/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
342
W_
343
traverseWorkStackBlocks(traverseState *ts)
sof's avatar
sof committed
344 345
{
    bdescr* bd;
346
    W_ res = 0;
sof's avatar
sof committed
347

348
    for (bd = ts->firstStack; bd != NULL; bd = bd->link)
sof's avatar
sof committed
349 350 351 352 353
      res += bd->blocks;

    return res;
}

354 355 356 357 358 359
W_
retainerStackBlocks(void)
{
    return traverseWorkStackBlocks(&g_retainerTraverseState);
}

360
/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
361
 * Returns true if stackTop is at the stack boundary of the current stack,
362 363
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
364
static INLINE bool
365
isOnBoundary( traverseState *ts )
366
{
367
    return ts->stackTop == ts->currentStackBoundary;
368 369 370 371 372 373 374
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
375
static INLINE void
376
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
377 378 379 380 381 382 383 384 385 386
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
387
static INLINE StgClosure *
388 389 390
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
391
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
392
    } else {
393
        return NULL;
394 395 396 397 398 399
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
400
static INLINE void
401
init_srt_fun( stackPos *info, const StgFunInfoTable *infoTable )
402
{
403 404 405
    info->type = posTypeSRT;
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_FUN_SRT(infoTable);
406
    } else {
407
        info->next.srt.srt = NULL;
408
    }
409 410
}

411
static INLINE void
412
init_srt_thunk( stackPos *info, const StgThunkInfoTable *infoTable )
413
{
Ömer Sinan Ağacan's avatar
Ömer Sinan Ağacan committed
414
    info->type = posTypeSRT;
415 416
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_SRT(infoTable);
417
    } else {
418
        info->next.srt.srt = NULL;
419
    }
420 421 422 423 424
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
425
static INLINE StgClosure *
426 427 428
find_srt( stackPos *info )
{
    StgClosure *c;
429
    if (info->type == posTypeSRT) {
430 431 432
        c = info->next.srt.srt;
        info->next.srt.srt = NULL;
        return c;
433 434 435
    }
}

436 437 438 439
/* -----------------------------------------------------------------------------
 * Pushes an element onto traverse stack
 * -------------------------------------------------------------------------- */
static void
440 441
pushStackElement(traverseState *ts, stackElement *se)
{
442
    bdescr *nbd;      // Next Block Descriptor
443
    if (ts->stackTop - 1 < ts->stackBottom) {
444
        debug("pop() to the next stack.\n");
445 446
        // currentStack->free is updated when the active stack is switched
        // to the next stack.
447
        ts->currentStack->free = (StgPtr)ts->stackTop;
448

449
        if (ts->currentStack->link == NULL) {
450 451
            nbd = allocGroup(BLOCKS_IN_STACK);
            nbd->link = NULL;
452 453
            nbd->u.back = ts->currentStack;
            ts->currentStack->link = nbd;
454
        } else
455
            nbd = ts->currentStack->link;
456

457
        newStackBlock(ts, nbd);
458 459 460
    }

    // adjust stackTop (acutal push)
461
    ts->stackTop--;
462 463 464 465
    // 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.
466
    *ts->stackTop = *se;
467

468 469 470
    ts->stackSize++;
    if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
    ASSERT(ts->stackSize >= 0);
471
    debug("stackSize = %d\n", ts->stackSize);
472 473 474 475 476 477
}

/* 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.
 *
478
 *  *cp - object's parent
479 480 481 482
 *  *c - closure
 *  c_child_r - closure retainer.
 */
static INLINE void
483
traversePushClosure(traverseState *ts, StgClosure *c, StgClosure *cp, stackData data) {
484 485 486
    stackElement se;

    se.c = c;
487 488
    se.cp = cp;
    se.data = data;
489 490
    se.info.type = posTypeFresh;

491
    pushStackElement(ts, &se);
492 493
};

494 495 496 497
/* -----------------------------------------------------------------------------
 *  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
498
 *  child, *c_child is set to that child and nothing is pushed onto
499 500 501 502 503 504
 *  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.
505
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
506 507 508
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
509
static INLINE void
510
traversePushChildren(traverseState *ts, StgClosure *c, stackData data, StgClosure **first_child)
511 512 513 514
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

515
    debug("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
516 517

    ASSERT(get_itbl(c)->type != TSO);
518
    ASSERT(get_itbl(c)->type != AP_STACK);
519 520 521 522 523 524

    //
    // fill in se
    //

    se.c = c;
525 526
    se.data = data;
    // Note: se.cp ommitted on purpose, only retainPushClosure uses that.
527 528 529

    // fill in se.info
    switch (get_itbl(c)->type) {
530
        // no child, no SRT
531 532 533
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
gcampax's avatar
gcampax committed
534
    case COMPACT_NFDATA:
535 536
        *first_child = NULL;
        return;
537

538
        // one child (fixed), no SRT
539 540
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
541 542
        *first_child = ((StgMutVar *)c)->var;
        return;
543
    case THUNK_SELECTOR:
544 545
        *first_child = ((StgSelector *)c)->selectee;
        return;
546
    case BLACKHOLE:
547 548
        *first_child = ((StgInd *)c)->indirectee;
        return;
549 550
    case CONSTR_1_0:
    case CONSTR_1_1:
551 552
        *first_child = c->payload[0];
        return;
553

554 555 556
        // 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;
557

558 559
        // two children (fixed), no SRT
        // need to push a stackElement, but nothing to store in se.info
560
    case CONSTR_2_0:
561
        *first_child = c->payload[0];         // return the first pointer
562 563
        se.info.type = posTypeStep;
        se.info.next.step = 2;            // 2 = second
564
        break;
565

566 567
        // three children (fixed), no SRT
        // need to push a stackElement
568 569
    case MVAR_CLEAN:
    case MVAR_DIRTY:
570 571 572
        // 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;
573
        se.info.type = posTypeStep;
574 575 576 577
        se.info.next.step = 2;            // 2 = second
        break;

        // three children (fixed), no SRT
578
    case WEAK:
579
        *first_child = ((StgWeak *)c)->key;
580
        se.info.type = posTypeStep;
581 582
        se.info.next.step = 2;
        break;
583

584
        // layout.payload.ptrs, no SRT
585
    case TVAR:
586
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
587
    case CONSTR_NOCAF:
588
    case PRIM:
589
    case MUT_PRIM:
590
    case BCO:
591 592 593 594 595 596 597 598
        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
599 600
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
601 602
    case MUT_ARR_PTRS_FROZEN_CLEAN:
    case MUT_ARR_PTRS_FROZEN_DIRTY:
603 604 605 606 607 608 609 610
        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
611 612
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
613 614
    case SMALL_MUT_ARR_PTRS_FROZEN_CLEAN:
    case SMALL_MUT_ARR_PTRS_FROZEN_DIRTY:
615 616 617 618 619 620
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;
621

622
    // layout.payload.ptrs, SRT
623
    case FUN_STATIC:
624 625
    case FUN:           // *c is a heap object.
    case FUN_2_0:
626 627 628 629 630 631
        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;
632

633 634
    case THUNK:
    case THUNK_2_0:
635 636 637 638 639 640 641 642 643
        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
644 645
    case FUN_1_0:
    case FUN_1_1:
646 647 648 649
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;
650

651 652
    case THUNK_1_0:
    case THUNK_1_1:
653 654 655 656
        *first_child = ((StgThunk *)c)->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        break;
657

658
    case FUN_0_1:      // *c is a heap object.
659
    case FUN_0_2:
660 661
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
662 663 664 665
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;
666 667 668

    // SRT only
    case THUNK_STATIC:
669
        ASSERT(get_itbl(c)->srt != 0);
670 671
    case THUNK_0_1:
    case THUNK_0_2:
672 673
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
674 675 676 677 678
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

679
    case TREC_CHUNK:
680
        *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
681
        se.info.type = posTypeStep;
682 683
        se.info.next.step = 0;  // entry no.
        break;
684

685
        // cannot appear
686
    case PAP:
687 688
    case AP:
    case AP_STACK:
689
    case TSO:
690
    case STACK:
691
    case IND_STATIC:
692
        // stack objects
693 694
    case UPDATE_FRAME:
    case CATCH_FRAME:
695
    case UNDERFLOW_FRAME:
696 697 698 699
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
700
        // invalid objects
701 702 703
    case IND:
    case INVALID_OBJECT:
    default:
704
        barf("Invalid object *c in push(): %d", get_itbl(c)->type);
705
        return;
706 707
    }

708 709 710 711
    // 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);

712
    pushStackElement(ts, &se);
713 714 715 716 717 718 719 720 721 722 723 724
}

/* -----------------------------------------------------------------------------
 *  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
725
 *    INLINE function (for the sake of execution speed).  popOffReal()
726 727 728
 *    is called only within popOff() and nowhere else.
 * -------------------------------------------------------------------------- */
static void
729
popStackElement(traverseState *ts) {
730
    debug("popStackElement(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
731 732 733 734 735 736 737

    ASSERT(ts->stackTop != ts->stackLimit);
    ASSERT(!isEmptyWorkStack(ts));

    // <= (instead of <) is wrong!
    if (ts->stackTop + 1 < ts->stackLimit) {
        ts->stackTop++;
738

739 740 741
        ts->stackSize--;
        if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
        ASSERT(ts->stackSize >= 0);
742 743
        debug("stackSize = (--) %d\n", ts->stackSize);

744 745 746
        return;
    }

747 748
    bdescr *pbd;    // Previous Block Descriptor

749
    debug("pop() to the previous stack.\n");
750

751 752
    ASSERT(ts->stackTop + 1 == ts->stackLimit);
    ASSERT(ts->stackBottom == (stackElement *)ts->currentStack->start);
753

754
    if (ts->firstStack == ts->currentStack) {
755
        // The stack is completely empty.
756 757
        ts->stackTop++;
        ASSERT(ts->stackTop == ts->stackLimit);
758

759 760 761
        ts->stackSize--;
        if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
        ASSERT(ts->stackSize >= 0);
762 763
        debug("stackSize = %d\n", ts->stackSize);

764
        return;
765 766 767 768
    }

    // currentStack->free is updated when the active stack is switched back
    // to the previous stack.
769
    ts->currentStack->free = (StgPtr)ts->stackLimit;
770 771

    // find the previous block descriptor
772
    pbd = ts->currentStack->u.back;
773 774
    ASSERT(pbd != NULL);

775
    returnToOldStack(ts, pbd);
776

777 778 779
    ts->stackSize--;
    if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
    ASSERT(ts->stackSize >= 0);
780
    debug("stackSize = %d\n", ts->stackSize);
781 782 783 784 785
}

/* -----------------------------------------------------------------------------
 *  Finds the next object to be considered for retainer profiling and store
 *  its pointer to *c.
786 787 788
 *  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,
789
 *  and if so, retrieve the first object and store its pointer to *c. Also,
790 791 792
 *  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.
793 794
 *  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
795
 *  the current stack chunk becomes empty, indicated by true returned by
796 797 798 799 800
 *  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.
 * -------------------------------------------------------------------------- */
801
static INLINE void
802
traversePop(traverseState *ts, StgClosure **c, StgClosure **cp, stackData *data)
803 804 805
{
    stackElement *se;

806
    debug("pop(): stackTop = 0x%x currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
807

808 809 810 811
    // Is this the last internal element? If so instead of modifying the current
    // stackElement in place we actually remove it from the stack.
    bool last = false;

812
    do {
813
        if (isOnBoundary(ts)) {     // if the current stack chunk is depleted
814 815 816 817
            *c = NULL;
            return;
        }

818 819
        // Note: Below every `break`, where the loop condition is true, must be
        // accompanied by a popOff() otherwise this is an infinite loop.
820
        se = ts->stackTop;
821

822 823
        // If this is a top-level element, you should pop that out.
        if (se->info.type == posTypeFresh) {
824
            *cp = se->cp;
825
            *c = se->c;
826
            *data = se->data;
827
            popStackElement(ts);
828 829 830
            return;
        }

831 832
        // Note: The first ptr of all of these was already returned as
        // *fist_child in push(), so we always start with the second field.
833 834 835 836 837
        switch (get_itbl(se->c)->type) {
            // two children (fixed), no SRT
            // nothing in se.info
        case CONSTR_2_0:
            *c = se->c->payload[1];
838 839
            last = true;
            goto out;
840 841 842

            // three children (fixed), no SRT
            // need to push a stackElement
843 844
        case MVAR_CLEAN:
        case MVAR_DIRTY:
845 846 847 848 849 850
            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;
851
                last = true;
852
            }
853
            goto out;
854 855 856 857 858 859 860 861 862

            // 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;
863
                last = true;
864
            }
865
            goto out;
866 867 868 869 870 871 872 873

        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;
874 875
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
876 877
            if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
                *c = NULL;
878
                popStackElement(ts);
879
                break; // this breaks out of the switch not the loop
880 881 882 883 884 885 886 887 888 889
            }
            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;
            }
            se->info.next.step++;
890
            goto out;
891
        }
892

893 894
        case TVAR:
        case CONSTR:
895 896 897 898 899 900
        case PRIM:
        case MUT_PRIM:
        case BCO:
            // StgMutArrPtr.ptrs, no SRT
        case MUT_ARR_PTRS_CLEAN:
        case MUT_ARR_PTRS_DIRTY:
901 902
        case MUT_ARR_PTRS_FROZEN_CLEAN:
        case MUT_ARR_PTRS_FROZEN_DIRTY:
903 904 905 906
        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:
907 908
            *c = find_ptrs(&se->info);
            if (*c == NULL) {
909
                popStackElement(ts);
910
                break; // this breaks out of the switch not the loop
911
            }
912
            goto out;
913 914 915

            // layout.payload.ptrs, SRT
        case FUN:         // always a heap object
916
        case FUN_STATIC:
917 918 919 920
        case FUN_2_0:
            if (se->info.type == posTypePtrs) {
                *c = find_ptrs(&se->info);
                if (*c != NULL) {
921
                    goto out;
922 923 924 925 926 927 928 929 930 931
                }
                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) {
932
                    goto out;
933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949
                }
                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);
950
            if(*c == NULL) {
951
                popStackElement(ts);
952
                break; // this breaks out of the switch not the loop
953
            }
954
            goto out;
955 956 957 958 959 960 961 962 963 964 965 966 967 968 969

            // 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:
970 971
        case STACK:
        case IND_STATIC:
Simon Marlow's avatar
Simon Marlow committed
972
        case CONSTR_NOCAF:
973
            // stack objects
974
        case UPDATE_FRAME:
975
        case CATCH_FRAME:
976 977
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
978 979 980 981 982 983 984
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default:
985
            barf("Invalid object *c in pop(): %d", get_itbl(se->c)->type);
986 987
            return;
        }
988 989 990 991 992 993 994 995 996 997
    } while (*c == NULL);

out:

    ASSERT(*c != NULL);

    *cp = se->c;
    *data = se->data;

    if(last)
998
        popStackElement(ts);
999 1000 1001

    return;

1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037
}

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

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

/* -----------------------------------------------------------------------------
 *  This function must be called before f-closing prof_file.
 * -------------------------------------------------------------------------- */
void
endRetainerProfiling( void )
{
    outputAllRetainerSet(prof_file);
}

/* -----------------------------------------------------------------------------
 *  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.
 * -------------------------------------------------------------------------- */
1038
static INLINE void
1039
traverseMaybeInitClosureData(StgClosure *c)
1040
{
1041 1042
    if (!isTravDataValid(c)) {
        setTravDataToZero(c);
1043 1044 1045 1046
    }
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
1047
 * Returns true if *c is a retainer.
1048 1049 1050 1051 1052 1053
 * 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.
1054
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
1055
static INLINE bool
1056 1057 1058
isRetainer( StgClosure *c )
{
    switch (get_itbl(c)->type) {
1059 1060 1061 1062
        //
        //  True case
        //
        // TSOs MUST be retainers: they constitute the set of roots.
1063
    case TSO:
1064
    case STACK:
1065

1066
        // mutable objects
1067
    case MUT_PRIM:
1068 1069
    case MVAR_CLEAN:
    case MVAR_DIRTY:
1070
    case TVAR:
1071 1072
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
1073 1074
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
1075 1076 1077
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case BLOCKING_QUEUE:
1078

1079
        // thunks are retainers.
1080 1081 1082 1083 1084 1085 1086
    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:
1087 1088
    case AP:
    case AP_STACK:
1089

1090
        // Static thunks, or CAFS, are obviously retainers.
1091 1092
    case THUNK_STATIC:

1093 1094
        // WEAK objects are roots; there is separate code in which traversing
        // begins from WEAK objects.
1095
    case WEAK:
Ben Gamari's avatar
Ben Gamari committed
1096
        return true;
1097

1098 1099 1100
        //
        // False case
        //
1101

1102
        // constructors
1103
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
1104
    case CONSTR_NOCAF:
1105 1106 1107 1108 1109
    case CONSTR_1_0:
    case CONSTR_0_1:
    case CONSTR_2_0:
    case CONSTR_1_1:
    case CONSTR_0_2:
1110
        // functions
1111 1112 1113 1114 1115 1116
    case FUN:
    case FUN_1_0:
    case FUN_0_1:
    case FUN_2_0:
    case FUN_1_1:
    case FUN_0_2:
1117
        // partial applications
1118
    case PAP:
1119
        // indirection
Ian Lynagh's avatar
Ian Lynagh committed
1120 1121 1122 1123
    // 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:
1124
    case BLACKHOLE:
1125
    case WHITEHOLE:
1126
        // static objects
1127
    case FUN_STATIC:
1128
        // misc
1129
    case PRIM:
1130 1131
    case BCO:
    case ARR_WORDS:
1132
    case COMPACT_NFDATA:
1133
        // STM
1134
    case TREC_CHUNK:
1135
        // immutable arrays
1136 1137 1138 1139
    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
1140
        return false;
1141

1142 1143 1144 1145 1146
        //
        // Error case
        //
        // Stack objects are invalid because they are never treated as
        // legal objects during retainer profiling.
1147 1148
    case UPDATE_FRAME:
    case CATCH_FRAME:
1149 1150
    case CATCH_RETRY_FRAME:
    case CATCH_STM_FRAME:
1151
    case UNDERFLOW_FRAME:
1152
    case ATOMICALLY_FRAME:
1153 1154 1155 1156
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
1157
    case RET_FUN:
1158
        // other cases
1159 1160 1161
    case IND:
    case INVALID_OBJECT:
    default:
1162
        barf(