RetainerProfile.c 70.3 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 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Exact algorithm and additional information can be found the historical
document 'docs/storage-mgt/rp.tex'.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RTS implementation may be out of date, but the general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)
105 106 107 108 109 110 111
/*
  cStackSize records how many times retainStack() has been invoked recursively,
  that is, the number of activation records for retainStack() on the C stack.
  maxCStackSize records its max value.
  Invariants:
    cStackSize <= maxCStackSize
 */
112
static uint32_t cStackSize, maxCStackSize;
113

114
static uint32_t sumOfNewCost;        // sum of the cost of each object, computed
115
                                // when the object is first visited
116
static uint32_t sumOfNewCostExtra;   // for those objects not visited during
117
                                // retainer profiling, e.g., MUT_VAR
118
static uint32_t costArray[N_CLOSURE_TYPES];
119

120
uint32_t sumOfCostLinear;            // sum of the costs of all object, computed
121 122
                                // when linearly traversing the heap after
                                // retainer profiling
123
uint32_t costArrayLinear[N_CLOSURE_TYPES];
124 125 126 127 128 129 130 131 132 133 134 135 136
#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 {
137 138
    // Object with fixed layout. Keeps an information about that
    // element was processed. (stackPos.next.step)
139
    posTypeStep,
140 141
    // Description of the pointers-first heap object. Keeps information
    // about layout. (stackPos.next.ptrs)
142
    posTypePtrs,
143
    // Keeps SRT bitmap (stackPos.next.srt)
144
    posTypeSRT,
145 146 147
    // Keeps a new object that was not inspected yet. Keeps a parent
    // element (stackPos.next.parent)
    posTypeFresh
148 149 150 151 152 153 154 155
} nextPosType;

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

    // layout.payload
    struct {
156 157 158
        // See StgClosureInfo in InfoTables.h
        StgHalfWord pos;
        StgHalfWord ptrs;
159
        StgPtr payload;
160 161 162 163
    } ptrs;

    // SRT
    struct {
164
        StgClosure *srt;
165
    } srt;
166 167 168 169

    // parent of the current object, used
    // when posTypeFresh is set
    StgClosure *parent;
170 171
} nextPos;

172 173
// Tagged stack element, that keeps information how to process
// the next element in the traverse stack.
174 175 176 177 178
typedef struct {
    nextPosType type;
    nextPos next;
} stackPos;

179 180
// Element in the traverse stack, keeps the element, information
// how to continue processing the element, and it's retainer set.
181 182
typedef struct {
    StgClosure *c;
183
    retainer c_child_r;
184 185 186
    stackPos info;
} stackElement;

187 188
static void retainActualPush( stackElement *se);

189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
/*
  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
206
static bdescr *firstStack = NULL;
207 208 209 210 211 212 213 214 215 216 217
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;

218
#if defined(DEBUG_RETAINER)
219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241
/*
  stackSize records the current size of the stack.
  maxStackSize records its high water mark.
  Invariants:
    stackSize <= maxStackSize
  Note:
    stackSize is just an estimate measure of the depth of the graph. The reason
    is that some heap objects have only a single child and may not result
    in a new element being pushed onto the stack. Therefore, at the end of
    retainer profiling, maxStackSize + maxCStackSize is some value no greater
    than the actual depth of the graph.
 */
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.
 * -------------------------------------------------------------------------- */
242
static INLINE void
243 244 245 246 247 248 249 250 251 252 253 254 255 256
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.
 * -------------------------------------------------------------------------- */
257
static INLINE void
258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273
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) {
274
        freeChain(firstStack);
275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296
    }

    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
297
 * Returns true if the whole stack is empty.
298
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
299
static INLINE bool
300 301 302 303 304
isEmptyRetainerStack( void )
{
    return (firstStack == currentStack) && stackTop == stackLimit;
}

sof's avatar
sof committed
305 306 307
/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
308
W_
309
retainerStackBlocks( void )
sof's avatar
sof committed
310 311
{
    bdescr* bd;
312
    W_ res = 0;
sof's avatar
sof committed
313

314
    for (bd = firstStack; bd != NULL; bd = bd->link)
sof's avatar
sof committed
315 316 317 318 319
      res += bd->blocks;

    return res;
}

320
/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
321
 * Returns true if stackTop is at the stack boundary of the current stack,
322 323
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
324
static INLINE bool
325 326 327 328 329 330 331 332 333 334
isOnBoundary( void )
{
    return stackTop == currentStackBoundary;
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
335
static INLINE void
336
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
337 338 339 340 341 342 343 344 345 346
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
347
static INLINE StgClosure *
348 349 350
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
351
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
352
    } else {
353
        return NULL;
354 355 356 357 358 359
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
360
static INLINE void
361
init_srt_fun( stackPos *info, const StgFunInfoTable *infoTable )
362
{
363 364 365
    info->type = posTypeSRT;
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_FUN_SRT(infoTable);
366
    } else {
367
        info->next.srt.srt = NULL;
368
    }
369 370
}

371
static INLINE void
372
init_srt_thunk( stackPos *info, const StgThunkInfoTable *infoTable )
373
{
Ömer Sinan Ağacan's avatar
Ömer Sinan Ağacan committed
374
    info->type = posTypeSRT;
375 376
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_SRT(infoTable);
377
    } else {
378
        info->next.srt.srt = NULL;
379
    }
380 381 382 383 384
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
385
static INLINE StgClosure *
386 387 388
find_srt( stackPos *info )
{
    StgClosure *c;
389
    if (info->type == posTypeSRT) {
390 391 392
        c = info->next.srt.srt;
        info->next.srt.srt = NULL;
        return c;
393 394 395
    }
}

396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431
/* -----------------------------------------------------------------------------
 * 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;
432 433
    ASSERT(stackSize >= 0);
    debugBelch("stackSize = %d\n", stackSize);
434
#endif
435

436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457
}

/* 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);
};

458 459 460 461
/* -----------------------------------------------------------------------------
 *  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
462
 *  child, *c_child is set to that child and nothing is pushed onto
463 464 465 466 467 468
 *  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.
469
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
470 471 472
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
473
static INLINE void
474
push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
475 476 477 478
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
479
#if defined(DEBUG_RETAINER)
480
    debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
481 482 483
#endif

    ASSERT(get_itbl(c)->type != TSO);
484
    ASSERT(get_itbl(c)->type != AP_STACK);
485 486 487 488 489 490 491 492 493 494

    //
    // fill in se
    //

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

    // fill in se.info
    switch (get_itbl(c)->type) {
495
        // no child, no SRT
496 497 498
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
gcampax's avatar
gcampax committed
499
    case COMPACT_NFDATA:
500 501
        *first_child = NULL;
        return;
502

503
        // one child (fixed), no SRT
504 505
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
506 507
        *first_child = ((StgMutVar *)c)->var;
        return;
508
    case THUNK_SELECTOR:
509 510
        *first_child = ((StgSelector *)c)->selectee;
        return;
511
    case BLACKHOLE:
512 513
        *first_child = ((StgInd *)c)->indirectee;
        return;
514 515
    case CONSTR_1_0:
    case CONSTR_1_1:
516 517
        *first_child = c->payload[0];
        return;
518

519 520 521
        // 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;
522

523 524
        // two children (fixed), no SRT
        // need to push a stackElement, but nothing to store in se.info
525
    case CONSTR_2_0:
526
        *first_child = c->payload[0];         // return the first pointer
527 528
        se.info.type = posTypeStep;
        se.info.next.step = 2;            // 2 = second
529
        break;
530

531 532
        // three children (fixed), no SRT
        // need to push a stackElement
533 534
    case MVAR_CLEAN:
    case MVAR_DIRTY:
535 536 537
        // 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;
538
        se.info.type = posTypeStep;
539 540 541 542
        se.info.next.step = 2;            // 2 = second
        break;

        // three children (fixed), no SRT
543
    case WEAK:
544
        *first_child = ((StgWeak *)c)->key;
545
        se.info.type = posTypeStep;
546 547
        se.info.next.step = 2;
        break;
548

549
        // layout.payload.ptrs, no SRT
550
    case TVAR:
551
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
552
    case CONSTR_NOCAF:
553
    case PRIM:
554
    case MUT_PRIM:
555
    case BCO:
556 557 558 559 560 561 562 563
        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
564 565
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
566 567
    case MUT_ARR_PTRS_FROZEN_CLEAN:
    case MUT_ARR_PTRS_FROZEN_DIRTY:
568 569 570 571 572 573 574 575
        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
576 577
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
578 579
    case SMALL_MUT_ARR_PTRS_FROZEN_CLEAN:
    case SMALL_MUT_ARR_PTRS_FROZEN_DIRTY:
580 581 582 583 584 585
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;
586

587
    // layout.payload.ptrs, SRT
588
    case FUN_STATIC:
589 590
    case FUN:           // *c is a heap object.
    case FUN_2_0:
591 592 593 594 595 596
        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;
597

598 599
    case THUNK:
    case THUNK_2_0:
600 601 602 603 604 605 606 607 608
        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
609 610
    case FUN_1_0:
    case FUN_1_1:
611 612 613 614
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;
615

616 617
    case THUNK_1_0:
    case THUNK_1_1:
618 619 620 621
        *first_child = ((StgThunk *)c)->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        break;
622

623
    case FUN_0_1:      // *c is a heap object.
624
    case FUN_0_2:
625 626
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
627 628 629 630
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;
631 632 633

    // SRT only
    case THUNK_STATIC:
634
        ASSERT(get_itbl(c)->srt != 0);
635 636
    case THUNK_0_1:
    case THUNK_0_2:
637 638
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
639 640 641 642 643
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

644
    case TREC_CHUNK:
645
        *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
646
        se.info.type = posTypeStep;
647 648
        se.info.next.step = 0;  // entry no.
        break;
649

650
        // cannot appear
651
    case PAP:
652 653
    case AP:
    case AP_STACK:
654
    case TSO:
655
    case STACK:
656
    case IND_STATIC:
657
        // stack objects
658 659
    case UPDATE_FRAME:
    case CATCH_FRAME:
660
    case UNDERFLOW_FRAME:
661 662 663 664
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
665
        // invalid objects
666 667 668
    case IND:
    case INVALID_OBJECT:
    default:
669
        barf("Invalid object *c in push(): %d", get_itbl(c)->type);
670
        return;
671 672
    }

673
    retainActualPush(&se);
674 675 676 677 678 679 680 681 682 683 684 685
}

/* -----------------------------------------------------------------------------
 *  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
686
 *    INLINE function (for the sake of execution speed).  popOffReal()
687 688 689 690 691 692 693
 *    is called only within popOff() and nowhere else.
 * -------------------------------------------------------------------------- */
static void
popOffReal(void)
{
    bdescr *pbd;    // Previous Block Descriptor

Ben Gamari's avatar
Ben Gamari committed
694
#if defined(DEBUG_RETAINER)
695
    debugBelch("pop() to the previous stack.\n");
696 697 698 699 700 701
#endif

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

    if (firstStack == currentStack) {
702 703 704
        // The stack is completely empty.
        stackTop++;
        ASSERT(stackTop == stackLimit);
Ben Gamari's avatar
Ben Gamari committed
705
#if defined(DEBUG_RETAINER)
706 707
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
708 709
        ASSERT(stackSize >= 0);
        debugBelch("stackSize = %d\n", stackSize);
710
#endif
711
        return;
712 713 714 715 716 717 718 719 720 721 722 723
    }

    // 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
724
#if defined(DEBUG_RETAINER)
725 726
    stackSize--;
    if (stackSize > maxStackSize) maxStackSize = stackSize;
727 728
    ASSERT(stackSize >= 0);
    debugBelch("stackSize = %d\n", stackSize);
729 730 731
#endif
}

732
static INLINE void
733
popOff(void) {
Ben Gamari's avatar
Ben Gamari committed
734
#if defined(DEBUG_RETAINER)
735
    debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
736 737 738 739 740 741 742
#endif

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

    // <= (instead of <) is wrong!
    if (stackTop + 1 < stackLimit) {
743
        stackTop++;
Ben Gamari's avatar
Ben Gamari committed
744
#if defined(DEBUG_RETAINER)
745 746
        stackSize--;
        if (stackSize > maxStackSize) maxStackSize = stackSize;
747 748
        ASSERT(stackSize >= 0);
        debugBelch("stackSize = %d\n", stackSize);
749
#endif
750
        return;
751 752 753 754 755 756 757 758
    }

    popOffReal();
}

/* -----------------------------------------------------------------------------
 *  Finds the next object to be considered for retainer profiling and store
 *  its pointer to *c.
759 760 761
 *  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,
762 763 764 765 766 767
 *  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
768
 *  the current stack chunk becomes empty, indicated by true returned by
769 770 771 772 773
 *  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.
 * -------------------------------------------------------------------------- */
774
static INLINE void
775
pop( StgClosure **c, StgClosure **cp, retainer *r )
776 777 778
{
    stackElement *se;

Ben Gamari's avatar
Ben Gamari committed
779
#if defined(DEBUG_RETAINER)
780
    debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
781 782 783
#endif

    do {
784 785 786 787 788 789 790
        if (isOnBoundary()) {     // if the current stack chunk is depleted
            *c = NULL;
            return;
        }

        se = stackTop;

791 792 793 794 795 796 797 798 799
        // 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;
        }

800 801 802 803 804 805 806 807 808 809 810 811
        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
812 813
        case MVAR_CLEAN:
        case MVAR_DIRTY:
814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846
            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;
847 848
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866
            if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
                *c = NULL;
                popOff();
                return;
            }
            entry = &((StgTRecChunk *)se->c)->entries[entry_no];
            if (field_no == 0) {
                *c = (StgClosure *)entry->tvar;
            } else if (field_no == 1) {
                *c = entry->expected_value;
            } else {
                *c = entry->new_value;
            }
            *cp = se->c;
            *r = se->c_child_r;
            se->info.next.step++;
            return;
        }
867

868 869
        case TVAR:
        case CONSTR:
870 871 872 873 874 875
        case PRIM:
        case MUT_PRIM:
        case BCO:
            // StgMutArrPtr.ptrs, no SRT
        case MUT_ARR_PTRS_CLEAN:
        case MUT_ARR_PTRS_DIRTY:
876 877
        case MUT_ARR_PTRS_FROZEN_CLEAN:
        case MUT_ARR_PTRS_FROZEN_DIRTY:
878 879 880 881
        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:
882 883 884 885 886 887 888 889 890 891 892
            *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
893
        case FUN_STATIC:
894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952
        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:
953 954
        case STACK:
        case IND_STATIC:
Simon Marlow's avatar
Simon Marlow committed
955
        case CONSTR_NOCAF:
956
            // stack objects
957
        case UPDATE_FRAME:
958
        case CATCH_FRAME:
959 960
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
961 962 963 964 965 966 967
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default:
968
            barf("Invalid object *c in pop(): %d", get_itbl(se->c)->type);
969 970
            return;
        }
Ben Gamari's avatar
Ben Gamari committed
971
    } while (true);
972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990
}

/* -----------------------------------------------------------------------------
 * 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
991
#if defined(SECOND_APPROACH)
992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009
    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.
 * -------------------------------------------------------------------------- */
1010
static INLINE void
1011 1012 1013
maybeInitRetainerSet( StgClosure *c )
{
    if (!isRetainerSetFieldValid(c)) {
1014
        setRetainerSetToNull(c);
1015 1016 1017 1018
    }
}

/* -----------------------------------------------------------------------------
Ben Gamari's avatar
Ben Gamari committed
1019
 * Returns true if *c is a retainer.
1020 1021 1022 1023 1024 1025
 * 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.
1026
 * -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
1027
static INLINE bool
1028 1029 1030
isRetainer( StgClosure *c )
{
    switch (get_itbl(c)->type) {
1031 1032 1033 1034
        //
        //  True case
        //
        // TSOs MUST be retainers: they constitute the set of roots.
1035
    case TSO:
1036
    case STACK:
1037

1038
        // mutable objects
1039
    case MUT_PRIM:
1040 1041
    case MVAR_CLEAN:
    case MVAR_DIRTY:
1042
    case TVAR:
1043 1044
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
1045 1046
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
1047 1048 1049
    case SMALL_MUT_ARR_PTRS_CLEAN:
    case SMALL_MUT_ARR_PTRS_DIRTY:
    case BLOCKING_QUEUE:
1050

1051
        // thunks are retainers.
1052 1053 1054 1055 1056 1057 1058
    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:
1059 1060
    case AP:
    case AP_STACK:
1061

1062
        // Static thunks, or CAFS, are obviously retainers.
1063 1064
    case THUNK_STATIC:

1065 1066
        // WEAK objects are roots; there is separate code in which traversing
        // begins from WEAK objects.
1067
    case WEAK:
Ben Gamari's avatar
Ben Gamari committed
1068
        return true;
1069

1070 1071 1072
        //
        // False case
        //
1073

1074
        // constructors
1075
    case CONSTR:
Simon Marlow's avatar
Simon Marlow committed
1076
    case CONSTR_NOCAF:
1077 1078 1079 1080 1081
    case CONSTR_1_0:
    case CONSTR_0_1:
    case CONSTR_2_0:
    case CONSTR_1_1:
    case CONSTR_0_2:
1082
        // functions
1083 1084 1085 1086 1087 1088
    case FUN:
    case FUN_1_0:
    case FUN_0_1:
    case FUN_2_0:
    case FUN_1_1:
    case FUN_0_2:
1089
        // partial applications
1090
    case PAP:
1091
        // indirection
Ian Lynagh's avatar
Ian Lynagh committed
1092 1093 1094 1095
    // 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:
1096
    case BLACKHOLE:
1097
    case WHITEHOLE:
1098
        // static objects
1099
    case FUN_STATIC:
1100
        // misc
1101
    case PRIM:
1102 1103
    case BCO:
    case ARR_WORDS:
1104
    case COMPACT_NFDATA:
1105
        // STM
1106
    case TREC_CHUNK:
1107
        // immutable arrays
1108 1109 1110 1111
    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
1112
        return false;
1113

1114 1115 1116 1117 1118
        //
        // Error case
        //
        // Stack objects are invalid because they are never treated as
        // legal objects during retainer profiling.
1119 1120
    case UPDATE_FRAME:
    case CATCH_FRAME:
1121 1122
    case CATCH_RETRY_FRAME:
    case CATCH_STM_FRAME:
1123
    case UNDERFLOW_FRAME:
1124
    case ATOMICALLY_FRAME:
1125 1126 1127 1128
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
1129
    case RET_FUN:
1130
        // other cases
1131 1132 1133
    case IND:
    case INVALID_OBJECT:
    default:
1134
        barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
Ben Gamari's avatar
Ben Gamari committed
1135
        return false;
1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
    }
}

/* -----------------------------------------------------------------------------
 *  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
1148
 *    RetainerSet.c can simply do nothing.
1149 1150 1151 1152
 *    If this is not the case, we can free all the retainer sets and
 *    re-initialize the hash table.
 *    See refreshAllRetainerSet() in RetainerSet.c.
 * -------------------------------------------------------------------------- */
1153
static INLINE retainer
1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167
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
 * -------------------------------------------------------------------------- */
1168
static INLINE void
1169
associate( StgClosure *c, RetainerSet *s )
1170 1171 1172 1173 1174 1175
{
    // StgWord has the same size as pointers, so the following type
    // casting is okay.
    RSET(c) = (RetainerSet *)((StgWord)s | flip);
}

1176
/* -----------------------------------------------------------------------------
1177
   Call retainPushClosure for each of the closures covered by a large bitmap.
1178 1179 1180
   -------------------------------------------------------------------------- */

static void
1181
retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, uint32_t size,
1182
                     StgClosure *c, retainer c_child_r)
1183
{
1184
    uint32_t i, b;
1185
    StgWord bitmap;
1186

1187 1188 1189
    b = 0;
    bitmap = large_bitmap->bitmap[b];
    for (i = 0; i < size; ) {
1190
        if ((bitmap & 1) == 0) {
1191
            retainPushClosure((StgClosure *)*p, c, c_child_r);
1192 1193 1194 1195 1196 1197 1198 1199 1200
        }
        i++;
        p++;
        if (i % BITS_IN(W_) == 0) {
            b++;
            bitmap = large_bitmap->bitmap[b];
        } else {
            bitmap = bitmap >> 1;
        }
1201 1202 1203
    }
}

1204
static INLINE StgPtr
1205
retain_small_bitmap (StgPtr p, uint32_t size, StgWord bitmap,
1206
                     StgClosure *c, retainer c_child_r)
1207 1208
{
    while (size > 0) {
1209
        if ((bitmap & 1) == 0) {
1210
            retainPushClosure((StgClosure *)*p, c, c_child_r);
1211 1212 1213 1214
        }
        p++;
        bitmap = bitmap >> 1;
        size--;
1215 1216 1217 1218
    }
    return p;
}

1219 1220 1221 1222 1223 1224
/* -----------------------------------------------------------------------------
 *  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:
1225
 *    *c is one of the following: TSO, AP_STACK.
1226 1227 1228 1229 1230
 *    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).
1231
 *    If *c is TSO, its state is not ThreadComplete,or ThreadKilled,
1232
 *    which means that its stack is ready to process.
1233 1234
 *  Note:
 *    This code was almost plagiarzied from GC.c! For each pointer,
1235
 *    retainPushClosure() is invoked instead of evacuate().
1236 1237
 * -------------------------------------------------------------------------- */
static void
1238
retainStack( StgClosure *c, retainer c_child_r,
1239
             StgPtr stackStart, StgPtr stackEnd )
1240 1241
{
    stackElement *oldStackBoundary;
1242
    StgPtr p;
1243
    const StgRetInfoTable *info;
1244
    StgWord bitmap;
1245
    uint32_t size;
1246

Ben Gamari's avatar
Ben Gamari committed
1247
#if defined(DEBUG_RETAINER)
1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260
    cStackSize++;
    if (cStackSize > maxCStackSize) maxCStackSize = cStackSize;
#endif

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

Ben Gamari's avatar
Ben Gamari committed
1261
#if defined(DEBUG_RETAINER)
1262 1263
    debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n",
        oldStackBoundary, currentStackBoundary);
1264 1265
#endif

1266 1267
    ASSERT(get_itbl(c)->type == STACK);

1268 1269
    p = stackStart;
    while (p < stackEnd) {
1270
        info = get_ret_itbl((StgClosure *)p);
1271

1272
        switch(info->i.type) {
1273

1274
        case UPDATE_FRAME:
1275
            retainPushClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
1276 1277
            p += sizeofW(StgUpdateFrame);
            continue;
1278

1279 1280
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291
        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:
1292
            if (info->i.srt) {
1293
                retainPushClosure(GET_SRT(info), c, c_child_r);
1294
            }
1295 1296 1297 1298 1299 1300
            continue;

        case RET_BCO: {
            StgBCO *bco;

            p++;
1301
            retainPushClosure((StgClosure*)*p, c, c_child_r);
1302 1303 1304 <