RetainerProfile.c 69.8 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Exact algorithm and additional information can be found the historical
document 'docs/storage-mgt/rp.tex'.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RTS implementation may be out of date, but the general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 72 73 74 75
/*
  Note: what to change in order to plug-in a new retainer profiling scheme?
    (1) type retainer in ../includes/StgRetainerProf.h
    (2) retainer function R(), i.e., getRetainerFrom()
    (3) the two hashing functions, hashKeySingleton() and hashKeyAddElement(),
        in RetainerSet.h, if needed.
    (4) printRetainer() and printRetainerSetShort() in RetainerSet.c.
 */

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

76
static uint32_t retainerGeneration;  // generation
77

78 79 80
static uint32_t numObjectVisited;    // total number of objects visited
static uint32_t timesAnyObjectVisited;  // number of times any objects are
                                        // visited
81 82 83 84 85 86 87 88 89

/*
  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().
 */

90
StgWord flip = 0;     // flip bit
91 92 93 94 95
                      // must be 0 if DEBUG_RETAINER is on (for static closures)

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

96
static void retainStack(StgClosure *, retainer, StgPtr, StgPtr);
97
static void retainClosure(StgClosure *, StgClosure *, retainer);
Ben Gamari's avatar
Ben Gamari committed
98
#if defined(DEBUG_RETAINER)
99
static void belongToHeap(StgPtr p);
100
static uint32_t checkHeapSanityForRetainerProfiling( void );
101
#endif
102
static void retainPushClosure( StgClosure *p, StgClosure *c, retainer c_child_r);
103

Ben Gamari's avatar
Ben Gamari committed
104
#if defined(DEBUG_RETAINER)
105
static uint32_t sumOfNewCost;        // sum of the cost of each object, computed
106
                                // when the object is first visited
107
static uint32_t sumOfNewCostExtra;   // for those objects not visited during
108
                                // retainer profiling, e.g., MUT_VAR
109
static uint32_t costArray[N_CLOSURE_TYPES];
110

111
uint32_t sumOfCostLinear;            // sum of the costs of all object, computed
112 113
                                // when linearly traversing the heap after
                                // retainer profiling
114
uint32_t costArrayLinear[N_CLOSURE_TYPES];
115 116 117 118 119 120 121 122 123 124 125 126 127
#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 {
128 129
    // Object with fixed layout. Keeps an information about that
    // element was processed. (stackPos.next.step)
130
    posTypeStep,
131 132
    // Description of the pointers-first heap object. Keeps information
    // about layout. (stackPos.next.ptrs)
133
    posTypePtrs,
134
    // Keeps SRT bitmap (stackPos.next.srt)
135
    posTypeSRT,
136 137 138
    // Keeps a new object that was not inspected yet. Keeps a parent
    // element (stackPos.next.parent)
    posTypeFresh
139 140 141 142 143 144 145 146
} nextPosType;

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

    // layout.payload
    struct {
147 148 149
        // See StgClosureInfo in InfoTables.h
        StgHalfWord pos;
        StgHalfWord ptrs;
150
        StgPtr payload;
151 152 153 154
    } ptrs;

    // SRT
    struct {
155
        StgClosure *srt;
156
    } srt;
157 158 159 160

    // parent of the current object, used
    // when posTypeFresh is set
    StgClosure *parent;
161 162
} nextPos;

163 164
// Tagged stack element, that keeps information how to process
// the next element in the traverse stack.
165 166 167 168 169
typedef struct {
    nextPosType type;
    nextPos next;
} stackPos;

170 171
// Element in the traverse stack, keeps the element, information
// how to continue processing the element, and it's retainer set.
172 173
typedef struct {
    StgClosure *c;
174
    retainer c_child_r;
175 176 177
    stackPos info;
} stackElement;

178 179
static void retainActualPush( stackElement *se);

180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
/*
  Invariants:
    firstStack points to the first block group.
    currentStack points to the block group currently being used.
    currentStack->free == stackLimit.
    stackTop points to the topmost byte in the stack of currentStack.
    Unless the whole stack is empty, stackTop must point to the topmost
    object (or byte) in the whole stack. Thus, it is only when the whole stack
    is empty that stackTop == stackLimit (not during the execution of push()
    and pop()).
    stackBottom == currentStack->start.
    stackLimit == currentStack->start + BLOCK_SIZE_W * currentStack->blocks.
  Note:
    When a current stack becomes empty, stackTop is set to point to
    the topmost element on the previous block group so as to satisfy
    the invariants described above.
 */
sof's avatar
sof committed
197
static bdescr *firstStack = NULL;
198 199 200 201 202 203 204 205 206 207 208
static bdescr *currentStack;
static stackElement *stackBottom, *stackTop, *stackLimit;

/*
  currentStackBoundary is used to mark the current stack chunk.
  If stackTop == currentStackBoundary, it means that the current stack chunk
  is empty. It is the responsibility of the user to keep currentStackBoundary
  valid all the time if it is to be employed.
 */
static stackElement *currentStackBoundary;

209
#if defined(DEBUG_RETAINER)
210 211 212 213 214 215 216 217 218
/*
  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
219
    retainer profiling, maxStackSize is some value no greater
220 221 222 223 224 225 226 227 228 229 230 231 232
    than the actual depth of the graph.
 */
static int stackSize, maxStackSize;
#endif

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

/* -----------------------------------------------------------------------------
 * Add a new block group to the stack.
 * Invariants:
 *  currentStack->link == s.
 * -------------------------------------------------------------------------- */
233
static INLINE void
234 235 236 237 238 239 240 241 242 243 244 245 246 247
newStackBlock( bdescr *bd )
{
    currentStack = bd;
    stackTop     = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
    stackBottom  = (stackElement *)bd->start;
    stackLimit   = (stackElement *)stackTop;
    bd->free     = (StgPtr)stackLimit;
}

/* -----------------------------------------------------------------------------
 * Return to the previous block group.
 * Invariants:
 *   s->link == currentStack.
 * -------------------------------------------------------------------------- */
248
static INLINE void
249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
returnToOldStack( bdescr *bd )
{
    currentStack = bd;
    stackTop = (stackElement *)bd->free;
    stackBottom = (stackElement *)bd->start;
    stackLimit = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
    bd->free = (StgPtr)stackLimit;
}

/* -----------------------------------------------------------------------------
 *  Initializes the traverse stack.
 * -------------------------------------------------------------------------- */
static void
initializeTraverseStack( void )
{
    if (firstStack != NULL) {
265
        freeChain(firstStack);
266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287
    }

    firstStack = allocGroup(BLOCKS_IN_STACK);
    firstStack->link = NULL;
    firstStack->u.back = NULL;

    newStackBlock(firstStack);
}

/* -----------------------------------------------------------------------------
 * Frees all the block groups in the traverse stack.
 * Invariants:
 *   firstStack != NULL
 * -------------------------------------------------------------------------- */
static void
closeTraverseStack( void )
{
    freeChain(firstStack);
    firstStack = NULL;
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
288
 * Returns true if the whole stack is empty.
289
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
290
static INLINE bool
291 292 293 294 295
isEmptyRetainerStack( void )
{
    return (firstStack == currentStack) && stackTop == stackLimit;
}

sof's avatar
sof committed
296 297 298
/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
299
W_
300
retainerStackBlocks( void )
sof's avatar
sof committed
301 302
{
    bdescr* bd;
303
    W_ res = 0;
sof's avatar
sof committed
304

305
    for (bd = firstStack; bd != NULL; bd = bd->link)
sof's avatar
sof committed
306 307 308 309 310
      res += bd->blocks;

    return res;
}

311
/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
312
 * Returns true if stackTop is at the stack boundary of the current stack,
313 314
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
315
static INLINE bool
316 317 318 319 320 321 322 323 324 325
isOnBoundary( void )
{
    return stackTop == currentStackBoundary;
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
326
static INLINE void
327
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
328 329 330 331 332 333 334 335 336 337
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
338
static INLINE StgClosure *
339 340 341
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
342
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
343
    } else {
344
        return NULL;
345 346 347 348 349 350
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
351
static INLINE void
352
init_srt_fun( stackPos *info, const StgFunInfoTable *infoTable )
353
{
354 355 356
    info->type = posTypeSRT;
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_FUN_SRT(infoTable);
357
    } else {
358
        info->next.srt.srt = NULL;
359
    }
360 361
}

362
static INLINE void
363
init_srt_thunk( stackPos *info, const StgThunkInfoTable *infoTable )
364
{
Ömer Sinan Ağacan's avatar
Ömer Sinan Ağacan committed
365
    info->type = posTypeSRT;
366 367
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_SRT(infoTable);
368
    } else {
369
        info->next.srt.srt = NULL;
370
    }
371 372 373 374 375
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
376
static INLINE StgClosure *
377 378 379
find_srt( stackPos *info )
{
    StgClosure *c;
380
    if (info->type == posTypeSRT) {
381 382 383
        c = info->next.srt.srt;
        info->next.srt.srt = NULL;
        return c;
384 385 386
    }
}

387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422
/* -----------------------------------------------------------------------------
 * Pushes an element onto traverse stack
 * -------------------------------------------------------------------------- */
static void
retainActualPush(stackElement *se) {
    bdescr *nbd;      // Next Block Descriptor
    if (stackTop - 1 < stackBottom) {
#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.
        currentStack->free = (StgPtr)stackTop;

        if (currentStack->link == NULL) {
            nbd = allocGroup(BLOCKS_IN_STACK);
            nbd->link = NULL;
            nbd->u.back = currentStack;
            currentStack->link = nbd;
        } else
            nbd = currentStack->link;

        newStackBlock(nbd);
    }

    // adjust stackTop (acutal push)
    stackTop--;
    // If the size of stackElement was huge, we would better replace the
    // following statement by either a memcpy() call or a switch statement
    // on the type of the element. Currently, the size of stackElement is
    // small enough (5 words) that this direct assignment seems to be enough.
    *stackTop = *se;

#if defined(DEBUG_RETAINER)
    stackSize++;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
423 424
    ASSERT(stackSize >= 0);
    debugBelch("stackSize = %d\n", stackSize);
425
#endif
426

427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448
}

/* 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.
 *
 *  *p - object's parent
 *  *c - closure
 *  c_child_r - closure retainer.
 */
static INLINE void
retainPushClosure( StgClosure *c, StgClosure *p, retainer c_child_r) {
    stackElement se;

    se.c = c;
    se.c_child_r = c_child_r;
    se.info.next.parent = p;
    se.info.type = posTypeFresh;

    retainActualPush(&se);
};

449 450 451 452
/* -----------------------------------------------------------------------------
 *  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
453
 *  child, *c_child is set to that child and nothing is pushed onto
454 455 456 457 458 459
 *  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.
460
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
461 462 463
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
464
static INLINE void
465
push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
466 467 468 469
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
470
#if defined(DEBUG_RETAINER)
471
    debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
472 473 474
#endif

    ASSERT(get_itbl(c)->type != TSO);
475
    ASSERT(get_itbl(c)->type != AP_STACK);
476 477 478 479 480 481 482 483 484 485

    //
    // fill in se
    //

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

    // fill in se.info
    switch (get_itbl(c)->type) {
486
        // no child, no SRT
487 488 489
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
gcampax's avatar
gcampax committed
490
    case COMPACT_NFDATA:
491 492
        *first_child = NULL;
        return;
493

494
        // one child (fixed), no SRT
495 496
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
497 498
        *first_child = ((StgMutVar *)c)->var;
        return;
499
    case THUNK_SELECTOR:
500 501
        *first_child = ((StgSelector *)c)->selectee;
        return;
502
    case BLACKHOLE:
503 504
        *first_child = ((StgInd *)c)->indirectee;
        return;
505 506
    case CONSTR_1_0:
    case CONSTR_1_1:
507 508
        *first_child = c->payload[0];
        return;
509

510 511 512
        // 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;
513

514 515
        // two children (fixed), no SRT
        // need to push a stackElement, but nothing to store in se.info
516
    case CONSTR_2_0:
517
        *first_child = c->payload[0];         // return the first pointer
518 519
        se.info.type = posTypeStep;
        se.info.next.step = 2;            // 2 = second
520
        break;
521

522 523
        // three children (fixed), no SRT
        // need to push a stackElement
524 525
    case MVAR_CLEAN:
    case MVAR_DIRTY:
526 527 528
        // 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;
529
        se.info.type = posTypeStep;
530 531 532 533
        se.info.next.step = 2;            // 2 = second
        break;

        // three children (fixed), no SRT
534
    case WEAK:
535
        *first_child = ((StgWeak *)c)->key;
536
        se.info.type = posTypeStep;
537 538
        se.info.next.step = 2;
        break;
539

540
        // layout.payload.ptrs, no SRT
541
    case TVAR:
542
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
543
    case CONSTR_NOCAF:
544
    case PRIM:
545
    case MUT_PRIM:
546
    case BCO:
547 548 549 550 551 552 553 554
        init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs,
                  (StgPtr)c->payload);
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;   // no child
        break;

        // StgMutArrPtr.ptrs, no SRT
555 556
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
557 558
    case MUT_ARR_PTRS_FROZEN_CLEAN:
    case MUT_ARR_PTRS_FROZEN_DIRTY:
559 560 561 562 563 564 565 566
        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
567 568
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
569 570
    case SMALL_MUT_ARR_PTRS_FROZEN_CLEAN:
    case SMALL_MUT_ARR_PTRS_FROZEN_DIRTY:
571 572 573 574 575 576
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;
577

578
    // layout.payload.ptrs, SRT
579
    case FUN_STATIC:
580 581
    case FUN:           // *c is a heap object.
    case FUN_2_0:
582 583 584 585 586 587
        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;
588

589 590
    case THUNK:
    case THUNK_2_0:
591 592 593 594 595 596 597 598 599
        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
600 601
    case FUN_1_0:
    case FUN_1_1:
602 603 604 605
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;
606

607 608
    case THUNK_1_0:
    case THUNK_1_1:
609 610 611 612
        *first_child = ((StgThunk *)c)->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        break;
613

614
    case FUN_0_1:      // *c is a heap object.
615
    case FUN_0_2:
616 617
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
618 619 620 621
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;
622 623 624

    // SRT only
    case THUNK_STATIC:
625
        ASSERT(get_itbl(c)->srt != 0);
626 627
    case THUNK_0_1:
    case THUNK_0_2:
628 629
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
630 631 632 633 634
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

635
    case TREC_CHUNK:
636
        *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
637
        se.info.type = posTypeStep;
638 639
        se.info.next.step = 0;  // entry no.
        break;
640

641
        // cannot appear
642
    case PAP:
643 644
    case AP:
    case AP_STACK:
645
    case TSO:
646
    case STACK:
647
    case IND_STATIC:
648
        // stack objects
649 650
    case UPDATE_FRAME:
    case CATCH_FRAME:
651
    case UNDERFLOW_FRAME:
652 653 654 655
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
656
        // invalid objects
657 658 659
    case IND:
    case INVALID_OBJECT:
    default:
660
        barf("Invalid object *c in push(): %d", get_itbl(c)->type);
661
        return;
662 663
    }

664
    retainActualPush(&se);
665 666 667 668 669 670 671 672 673 674 675 676
}

/* -----------------------------------------------------------------------------
 *  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
677
 *    INLINE function (for the sake of execution speed).  popOffReal()
678 679 680 681 682 683 684
 *    is called only within popOff() and nowhere else.
 * -------------------------------------------------------------------------- */
static void
popOffReal(void)
{
    bdescr *pbd;    // Previous Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
685
#if defined(DEBUG_RETAINER)
686
    debugBelch("pop() to the previous stack.\n");
687 688 689 690 691 692
#endif

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

    if (firstStack == currentStack) {
693 694 695
        // The stack is completely empty.
        stackTop++;
        ASSERT(stackTop == stackLimit);
Ben Gamari's avatar
Ben Gamari committed
696
#if defined(DEBUG_RETAINER)
697 698
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
699 700
        ASSERT(stackSize >= 0);
        debugBelch("stackSize = %d\n", stackSize);
701
#endif
702
        return;
703 704 705 706 707 708 709 710 711 712 713 714
    }

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

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

    returnToOldStack(pbd);

Ben Gamari's avatar
Ben Gamari committed
715
#if defined(DEBUG_RETAINER)
716 717
    stackSize--;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
718 719
    ASSERT(stackSize >= 0);
    debugBelch("stackSize = %d\n", stackSize);
720 721 722
#endif
}

723
static INLINE void
724
popOff(void) {
Ben Gamari's avatar
Ben Gamari committed
725
#if defined(DEBUG_RETAINER)
726
    debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
727 728 729 730 731 732 733
#endif

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

    // <= (instead of <) is wrong!
    if (stackTop + 1 < stackLimit) {
734
        stackTop++;
Ben Gamari's avatar
Ben Gamari committed
735
#if defined(DEBUG_RETAINER)
736 737
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
738 739
        ASSERT(stackSize >= 0);
        debugBelch("stackSize = %d\n", stackSize);
740
#endif
741
        return;
742 743 744 745 746 747 748 749
    }

    popOffReal();
}

/* -----------------------------------------------------------------------------
 *  Finds the next object to be considered for retainer profiling and store
 *  its pointer to *c.
750 751 752
 *  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,
753 754 755 756 757 758
 *  and if so, retrieve the first object and store its pointer to *c. Also,
 *  set *cp and *r appropriately, both of which are stored in the stack element.
 *  The topmost stack element then is overwritten so as for it to now denote
 *  the next object.
 *  If the topmost stack element indicates no more objects are left, pop
 *  off the stack element until either an object can be retrieved or
Ben Gamari's avatar
Ben Gamari committed
759
 *  the current stack chunk becomes empty, indicated by true returned by
760 761 762 763 764
 *  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.
 * -------------------------------------------------------------------------- */
765
static INLINE void
766
pop( StgClosure **c, StgClosure **cp, retainer *r )
767 768 769
{
    stackElement *se;

Ben Gamari's avatar
Ben Gamari committed
770
#if defined(DEBUG_RETAINER)
771
    debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
772 773 774
#endif

    do {
775 776 777 778 779 780 781
        if (isOnBoundary()) {     // if the current stack chunk is depleted
            *c = NULL;
            return;
        }

        se = stackTop;

782 783 784 785 786 787 788 789 790
        // If this is a top-level element, you should pop that out.
        if (se->info.type == posTypeFresh) {
            *cp = se->info.next.parent;
            *c = se->c;
            *r = se->c_child_r;
            popOff();
            return;
        }

791 792 793 794 795 796 797 798 799 800 801 802
        switch (get_itbl(se->c)->type) {
            // two children (fixed), no SRT
            // nothing in se.info
        case CONSTR_2_0:
            *c = se->c->payload[1];
            *cp = se->c;
            *r = se->c_child_r;
            popOff();
            return;

            // three children (fixed), no SRT
            // need to push a stackElement
803 804
        case MVAR_CLEAN:
        case MVAR_DIRTY:
805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837
            if (se->info.next.step == 2) {
                *c = (StgClosure *)((StgMVar *)se->c)->tail;
                se->info.next.step++;             // move to the next step
                // no popOff
            } else {
                *c = ((StgMVar *)se->c)->value;
                popOff();
            }
            *cp = se->c;
            *r = se->c_child_r;
            return;

            // three children (fixed), no SRT
        case WEAK:
            if (se->info.next.step == 2) {
                *c = ((StgWeak *)se->c)->value;
                se->info.next.step++;
                // no popOff
            } else {
                *c = ((StgWeak *)se->c)->finalizer;
                popOff();
            }
            *cp = se->c;
            *r = se->c_child_r;
            return;

        case TREC_CHUNK: {
            // These are pretty complicated: we have N entries, each
            // of which contains 3 fields that we want to follow.  So
            // we divide the step counter: the 2 low bits indicate
            // which field, and the rest of the bits indicate the
            // entry number (starting from zero).
            TRecEntry *entry;
838 839
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
840 841 842
            if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
                *c = NULL;
                popOff();
843
                break;
844 845 846 847 848 849 850 851 852 853 854 855 856 857
            }
            entry = &((StgTRecChunk *)se->c)->entries[entry_no];
            if (field_no == 0) {
                *c = (StgClosure *)entry->tvar;
            } else if (field_no == 1) {
                *c = entry->expected_value;
            } else {
                *c = entry->new_value;
            }
            *cp = se->c;
            *r = se->c_child_r;
            se->info.next.step++;
            return;
        }
858

859 860
        case TVAR:
        case CONSTR:
861 862 863 864 865 866
        case PRIM:
        case MUT_PRIM:
        case BCO:
            // StgMutArrPtr.ptrs, no SRT
        case MUT_ARR_PTRS_CLEAN:
        case MUT_ARR_PTRS_DIRTY:
867 868
        case MUT_ARR_PTRS_FROZEN_CLEAN:
        case MUT_ARR_PTRS_FROZEN_DIRTY:
869 870 871 872
        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:
873 874 875 876 877 878 879 880 881 882 883
            *c = find_ptrs(&se->info);
            if (*c == NULL) {
                popOff();
                break;
            }
            *cp = se->c;
            *r = se->c_child_r;
            return;

            // layout.payload.ptrs, SRT
        case FUN:         // always a heap object
884
        case FUN_STATIC:
885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943
        case FUN_2_0:
            if (se->info.type == posTypePtrs) {
                *c = find_ptrs(&se->info);
                if (*c != NULL) {
                    *cp = se->c;
                    *r = se->c_child_r;
                    return;
                }
                init_srt_fun(&se->info, get_fun_itbl(se->c));
            }
            goto do_srt;

        case THUNK:
        case THUNK_2_0:
            if (se->info.type == posTypePtrs) {
                *c = find_ptrs(&se->info);
                if (*c != NULL) {
                    *cp = se->c;
                    *r = se->c_child_r;
                    return;
                }
                init_srt_thunk(&se->info, get_thunk_itbl(se->c));
            }
            goto do_srt;

            // SRT
        do_srt:
        case THUNK_STATIC:
        case FUN_0_1:
        case FUN_0_2:
        case THUNK_0_1:
        case THUNK_0_2:
        case FUN_1_0:
        case FUN_1_1:
        case THUNK_1_0:
        case THUNK_1_1:
            *c = find_srt(&se->info);
            if (*c != NULL) {
                *cp = se->c;
                *r = se->c_child_r;
                return;
            }
            popOff();
            break;

            // no child (fixed), no SRT
        case CONSTR_0_1:
        case CONSTR_0_2:
        case ARR_WORDS:
            // one child (fixed), no SRT
        case MUT_VAR_CLEAN:
        case MUT_VAR_DIRTY:
        case THUNK_SELECTOR:
        case CONSTR_1_1:
            // cannot appear
        case PAP:
        case AP:
        case AP_STACK:
        case TSO:
944 945
        case STACK:
        case IND_STATIC:
Simon Marlow's avatar
Simon Marlow committed
946
        case CONSTR_NOCAF:
947
            // stack objects
948
        case UPDATE_FRAME:
949
        case CATCH_FRAME:
950 951
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
952 953 954 955 956 957 958
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default:
959
            barf("Invalid object *c in pop(): %d", get_itbl(se->c)->type);
960 961
            return;
        }
Ben Gamari's avatar
Ben Gamari committed
962
    } while (true);
963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981
}

/* -----------------------------------------------------------------------------
 * 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
982
#if defined(SECOND_APPROACH)
983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000
    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.
 * -------------------------------------------------------------------------- */
1001
static INLINE void
1002 1003 1004
maybeInitRetainerSet( StgClosure *c )
{
    if (!isRetainerSetFieldValid(c)) {
1005
        setRetainerSetToNull(c);
1006 1007 1008 1009
    }
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
1010
 * Returns true if *c is a retainer.
1011 1012 1013 1014 1015 1016
 * 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.
1017
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
1018
static INLINE bool
1019 1020 1021
isRetainer( StgClosure *c )
{
    switch (get_itbl(c)->type) {
1022 1023 1024 1025
        //
        //  True case
        //
        // TSOs MUST be retainers: they constitute the set of roots.
1026
    case TSO:
1027
    case STACK:
1028

1029
        // mutable objects
1030
    case MUT_PRIM:
1031 1032
    case MVAR_CLEAN:
    case MVAR_DIRTY:
1033
    case TVAR:
1034 1035
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
1036 1037
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
1038 1039 1040
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case BLOCKING_QUEUE:
1041

1042
        // thunks are retainers.
1043 1044 1045 1046 1047 1048 1049
    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:
1050 1051
    case AP:
    case AP_STACK:
1052

1053
        // Static thunks, or CAFS, are obviously retainers.
1054 1055
    case THUNK_STATIC:

1056 1057
        // WEAK objects are roots; there is separate code in which traversing
        // begins from WEAK objects.
1058
    case WEAK:
Ben Gamari's avatar
Ben Gamari committed
1059
        return true;
1060

1061 1062 1063
        //
        // False case
        //
1064

1065
        // constructors
1066
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
1067
    case CONSTR_NOCAF:
1068 1069 1070 1071 1072
    case CONSTR_1_0:
    case CONSTR_0_1:
    case CONSTR_2_0:
    case CONSTR_1_1:
    case CONSTR_0_2:
1073
        // functions
1074 1075 1076 1077 1078 1079
    case FUN:
    case FUN_1_0:
    case FUN_0_1:
    case FUN_2_0:
    case FUN_1_1:
    case FUN_0_2:
1080
        // partial applications
1081
    case PAP:
1082
        // indirection
Ian Lynagh's avatar
Ian Lynagh committed
1083 1084 1085 1086
    // 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:
1087
    case BLACKHOLE:
1088
    case WHITEHOLE:
1089
        // static objects
1090
    case FUN_STATIC:
1091
        // misc
1092
    case PRIM:
1093 1094
    case BCO:
    case ARR_WORDS:
1095
    case COMPACT_NFDATA:
1096
        // STM
1097
    case TREC_CHUNK:
1098
        // immutable arrays
1099 1100 1101 1102
    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
1103
        return false;
1104

1105 1106 1107 1108 1109
        //
        // Error case
        //
        // Stack objects are invalid because they are never treated as
        // legal objects during retainer profiling.
1110 1111
    case UPDATE_FRAME:
    case CATCH_FRAME:
1112 1113
    case CATCH_RETRY_FRAME:
    case CATCH_STM_FRAME:
1114
    case UNDERFLOW_FRAME:
1115
    case ATOMICALLY_FRAME:
1116 1117 1118 1119
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
1120
    case RET_FUN:
1121
        // other cases
1122 1123 1124
    case IND:
    case INVALID_OBJECT:
    default:
1125
        barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
Ben Gamari's avatar
Ben Gamari committed
1126
        return false;
1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138
    }
}

/* -----------------------------------------------------------------------------
 *  Returns the retainer function value for the closure *c, i.e., R(*c).
 *  This function does NOT return the retainer(s) of *c.
 *  Invariants:
 *    *c must be a retainer.
 *  Note:
 *    Depending on the definition of this function, the maintenance of retainer
 *    sets can be made easier. If most retainer sets are likely to be created
 *    again across garbage collections, refreshAllRetainerSet() in
1139
 *    RetainerSet.c can simply do nothing.
1140 1141 1142 1143
 *    If this is not the case, we can free all the retainer sets and
 *    re-initialize the hash table.
 *    See refreshAllRetainerSet() in RetainerSet.c.
 * -------------------------------------------------------------------------- */
1144
static INLINE retainer
1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158
getRetainerFrom( StgClosure *c )
{
    ASSERT(isRetainer(c));

    return c->header.prof.ccs;
}

/* -----------------------------------------------------------------------------
 *  Associates the retainer set *s with the closure *c, that is, *s becomes
 *  the retainer set of *c.
 *  Invariants:
 *    c != NULL
 *    s != NULL
 * -------------------------------------------------------------------------- */
1159
static INLINE void
1160
associate( StgClosure *c, RetainerSet *s )
1161 1162 1163 1164 1165 1166
{
    // StgWord has the same size as pointers, so the following type
    // casting is okay.
    RSET(c) = (RetainerSet *)((StgWord)s | flip);
}

1167
/* -----------------------------------------------------------------------------
1168
   Call retainPushClosure for each of the closures covered by a large bitmap.
1169 1170 1171
   -------------------------------------------------------------------------- */

static void
1172
retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, uint32_t size,
1173
                     StgClosure *c, retainer c_child_r)
1174
{
1175
    uint32_t i, b;
1176
    StgWord bitmap;
1177

1178 1179 1180
    b = 0;
    bitmap = large_bitmap->bitmap[b];
    for (i = 0; i < size; ) {
1181
        if ((bitmap & 1) == 0) {
1182
            retainPushClosure((StgClosure *)*p, c, c_child_r);
1183 1184 1185 1186 1187 1188 1189 1190 1191
        }
        i++;
        p++;
        if (i % BITS_IN(W_) == 0) {
            b++;
            bitmap = large_bitmap->bitmap[b];
        } else {
            bitmap = bitmap >> 1;
        }
1192 1193 1194
    }
}

1195
static INLINE StgPtr
1196
retain_small_bitmap (StgPtr p, uint32_t size, StgWord bitmap,
1197
                     StgClosure *c, retainer c_child_r)
1198 1199
{
    while (size > 0) {
1200
        if ((bitmap & 1) == 0) {
1201
            retainPushClosure((StgClosure *)*p, c, c_child_r);
1202 1203 1204 1205
        }
        p++;
        bitmap = bitmap >> 1;
        size--;
1206 1207 1208 1209
    }
    return p;
}

1210 1211 1212 1213 1214 1215
/* -----------------------------------------------------------------------------
 *  Process all the objects in the stack chunk from stackStart to stackEnd
 *  with *c and *c_child_r being their parent and their most recent retainer,
 *  respectively. Treat stackOptionalFun as another child of *c if it is
 *  not NULL.
 *  Invariants:
1216
 *    *c is one of the following: TSO, AP_STACK.
1217 1218 1219 1220 1221
 *    If *c is TSO, c == c_child_r.
 *    stackStart < stackEnd.
 *    RSET(c) and RSET(c_child_r) are valid, i.e., their
 *    interpretation conforms to the current value of flip (even when they
 *    are interpreted to be NULL).
1222
 *    If *c is TSO, its state is not ThreadComplete,or ThreadKilled,
1223
 *    which means that its stack is ready to process.
1224 1225
 *  Note:
 *    This code was almost plagiarzied from GC.c! For each pointer,
1226
 *    retainPushClosure() is invoked instead of evacuate().
1227 1228
 * -------------------------------------------------------------------------- */
static void
1229
retainStack( StgClosure *c, retainer c_child_r,
1230
             StgPtr stackStart, StgPtr stackEnd )
1231 1232
{
    stackElement *oldStackBoundary;
1233
    StgPtr p;
1234
    const StgRetInfoTable *info;
1235
    StgWord bitmap;
1236
    uint32_t size;
1237 1238 1239 1240 1241 1242 1243 1244 1245 1246

    /*
      Each invocation of retainStack() creates a new virtual
      stack. Since all such stacks share a single common stack, we
      record the current currentStackBoundary, which will be restored
      at the exit.
    */
    oldStackBoundary = currentStackBoundary;
    currentStackBoundary = stackTop;

Ben Gamari's avatar
Ben Gamari committed
1247
#if defined(DEBUG_RETAINER)
1248 1249
    debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n",
        oldStackBoundary, currentStackBoundary);
1250 1251
#endif

1252 1253
    ASSERT(get_itbl(c)->type == STACK);

1254 1255
    p = stackStart;
    while (p < stackEnd) {
1256
        info = get_ret_itbl((StgClosure *)p);
1257

1258
        switch(info->i.type) {
1259

1260
        case UPDATE_FRAME:
1261
            retainPushClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
1262 1263
            p += sizeofW(StgUpdateFrame);
            continue;
1264

1265 1266
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277
        case CATCH_FRAME:
        case CATCH_STM_FRAME:
        case CATCH_RETRY_FRAME:
        case ATOMICALLY_FRAME:
        case RET_SMALL:
            bitmap = BITMAP_BITS(info->i.layout.bitmap);
            size   = BITMAP_SIZE(info->i.layout.bitmap);
            p++;
            p = retain_small_bitmap(p, size, bitmap, c, c_child_r);

        follow_srt:
1278
            if (info->i.srt) {
1279
                retainPushClosure(GET_SRT(info), c, c_child_r);
1280
            }
1281 1282 1283 1284 1285 1286
            continue;

        case RET_BCO: {
            StgBCO *bco;

            p++;
1287
            retainPushClosure((StgClosure*)*p, c, c_child_r);
1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304
            bco = (StgBCO *)*p;
            p++;
            size = BCO_BITMAP_SIZE(bco);
            retain_large_bitmap(p, BCO_BITMAP(bco), size, c, c_child_r);
            p += size;
            continue;
        }

            // large bitmap (> 32 entries, or > 64 on a 64-bit machine)
        case RET_BIG:
            size = GET_LARGE_BITMAP(&info->i)->size;
            p++;
            retain_large_bitmap(p, GET_LARGE_BITMAP(&info->i),
                                size, c, c_child_r);</