RetainerProfile.c

/* -----------------------------------------------------------------------------
 *
 * (c) The GHC Team, 2001
 * Author: Sungwoo Park
 *
 * Retainer profiling.
 *
 * ---------------------------------------------------------------------------*/

#if defined(PROFILING)

// Turn off inlining when debugging - it obfuscates things
#if defined(DEBUG)
#define INLINE
#else
#define INLINE inline
#endif

#include "PosixSource.h"
#include "Rts.h"

#include "RtsUtils.h"
#include "RetainerProfile.h"
#include "RetainerSet.h"
#include "Schedule.h"
#include "Printer.h"
#include "Weak.h"
#include "sm/Sanity.h"
#include "Profiling.h"
#include "Stats.h"
#include "ProfHeap.h"
#include "Apply.h"
#include "StablePtr.h" /* markStablePtrTable */
#include "StableName.h" /* rememberOldStableNameAddresses */
#include "sm/Storage.h" // for END_OF_STATIC_LIST

/* Note [What is a retainer?]
   ~~~~~~~~~~~~~~~~~~~~~~~~~~
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 keep 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Exact algorithm and additional information can be found the historical
document 'docs/storage-mgt/rp.tex'.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RTS implementation may be out of date, but the general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.
*/


/*
  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.
 */

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

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

static uint32_t retainerGeneration;  // generation

static uint32_t numObjectVisited;    // total number of objects visited
static uint32_t timesAnyObjectVisited;  // number of times any objects are
                                        // visited

/** Note [Profiling heap traversal visited bit]
 *
 * If the RTS is compiled with profiling enabled StgProfHeader can be used by
 * profiling code to store per-heap object information.
 *
 * When using the generic heap traversal code we use this field to store
 * profiler specific information. However we reserve the LSB of the *entire*
 * 'trav' union (which will overlap with the other fields) for the generic
 * traversal code. We use the bit to decide whether we've already visited this
 * closure in this pass or not. We do this as the heap may contain cyclic
 * references, it being a graph and all, so we would likely just infinite loop
 * if we didn't.
 *
 * We assume that at least the LSB of the largest field in the corresponding
 * union is insignificant. This is true at least for the word aligned pointers
 * which the retainer profiler currently stores there and should be maintained
 * by new users of the 'trav' union.
 *
 * Now the way the traversal works is that the interpretation of the "visited?"
 * bit depends on the value of the global 'flip' variable. We don't want to have
 * to do another pass over the heap just to reset the bit to zero so instead on
 * each traversal (i.e. each run of the profiling code) we invert the value of
 * the global 'flip' variable. We interpret this as resetting all the "visited?"
 * flags on the heap.
 *
 * There is one exception to this rule, namely: static objects. There we do just
 * go over the heap and reset the bit manually. See
 * 'resetStaticObjectForRetainerProfiling'.
 */
StgWord flip = 0;     // flip bit
                      // must be 0 if DEBUG_RETAINER is on (for static closures)

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

#if defined(DEBUG_RETAINER)
static uint32_t sumOfNewCost;        // sum of the cost of each object, computed
                                // when the object is first visited
static uint32_t sumOfNewCostExtra;   // for those objects not visited during
                                // retainer profiling, e.g., MUT_VAR
static uint32_t costArray[N_CLOSURE_TYPES];

uint32_t sumOfCostLinear;            // sum of the costs of all object, computed
                                // when linearly traversing the heap after
                                // retainer profiling
uint32_t costArrayLinear[N_CLOSURE_TYPES];
#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 {
    // Object with fixed layout. Keeps an information about that
    // element was processed. (stackPos.next.step)
    posTypeStep,
    // Description of the pointers-first heap object. Keeps information
    // about layout. (stackPos.next.ptrs)
    posTypePtrs,
    // Keeps SRT bitmap (stackPos.next.srt)
    posTypeSRT,
    // Keeps a new object that was not inspected yet. Keeps a parent
    // element (stackPos.next.parent)
    posTypeFresh
} nextPosType;

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

    // layout.payload
    struct {
        // See StgClosureInfo in InfoTables.h
        StgHalfWord pos;
        StgHalfWord ptrs;
        StgPtr payload;
    } ptrs;

    // SRT
    struct {
        StgClosure *srt;
    } srt;
} nextPos;

// Tagged stack element, that keeps information how to process
// the next element in the traverse stack.
typedef struct {
    nextPosType type;
    nextPos next;
} stackPos;

typedef union {
     /**
      * Most recent retainer for the corresponding closure on the stack.
      */
    retainer c_child_r;
} stackData;

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

typedef struct {
/*
  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.
 */
    bdescr *firstStack;
    bdescr *currentStack;
    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.
 */
    stackElement *currentStackBoundary;

#if defined(DEBUG_RETAINER)
/*
  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 is some value no greater
    than the actual depth of the graph.
 */
    int stackSize, maxStackSize;
#endif
} traverseState;

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

traverseState g_retainerTraverseState;


static void retainStack(traverseState *, StgClosure *, stackData, StgPtr, StgPtr);
static void retainClosure(traverseState *, StgClosure *, StgClosure *, retainer);
static void retainPushClosure(traverseState *, StgClosure *, StgClosure *, stackData);
static void retainActualPush(traverseState *, stackElement *);

#if defined(DEBUG_RETAINER)
static void belongToHeap(StgPtr p);
static uint32_t checkHeapSanityForRetainerProfiling( void );
#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.
 * -------------------------------------------------------------------------- */
static INLINE void
newStackBlock( traverseState *ts, bdescr *bd )
{
    ts->currentStack = bd;
    ts->stackTop     = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
    ts->stackBottom  = (stackElement *)bd->start;
    ts->stackLimit   = (stackElement *)ts->stackTop;
    bd->free     = (StgPtr)ts->stackLimit;
}

/* -----------------------------------------------------------------------------
 * Return to the previous block group.
 * Invariants:
 *   s->link == currentStack.
 * -------------------------------------------------------------------------- */
static INLINE void
returnToOldStack( traverseState *ts, bdescr *bd )
{
    ts->currentStack = bd;
    ts->stackTop = (stackElement *)bd->free;
    ts->stackBottom = (stackElement *)bd->start;
    ts->stackLimit = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
    bd->free = (StgPtr)ts->stackLimit;
}

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

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

    newStackBlock(ts, ts->firstStack);
}

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

/* -----------------------------------------------------------------------------
 * Returns true if the whole stack is empty.
 * -------------------------------------------------------------------------- */
static INLINE bool
isEmptyRetainerStack( traverseState *ts )
{
    return (ts->firstStack == ts->currentStack) && ts->stackTop == ts->stackLimit;
}

/* -----------------------------------------------------------------------------
 * Returns size of stack
 * -------------------------------------------------------------------------- */
W_
retainerStackBlocks( void )
{
    bdescr* bd;
    W_ res = 0;
    traverseState *ts = &g_retainerTraverseState;

    for (bd = ts->firstStack; bd != NULL; bd = bd->link)
      res += bd->blocks;

    return res;
}

/* -----------------------------------------------------------------------------
 * Returns true if stackTop is at the stack boundary of the current stack,
 * i.e., if the current stack chunk is empty.
 * -------------------------------------------------------------------------- */
static INLINE bool
isOnBoundary( traverseState *ts )
{
    return ts->stackTop == ts->currentStackBoundary;
}

/* -----------------------------------------------------------------------------
 * Initializes *info from ptrs and payload.
 * Invariants:
 *   payload[] begins with ptrs pointers followed by non-pointers.
 * -------------------------------------------------------------------------- */
static INLINE void
init_ptrs( stackPos *info, uint32_t ptrs, StgPtr payload )
{
    info->type              = posTypePtrs;
    info->next.ptrs.pos     = 0;
    info->next.ptrs.ptrs    = ptrs;
    info->next.ptrs.payload = payload;
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
static INLINE StgClosure *
find_ptrs( stackPos *info )
{
    if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
        return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
    } else {
        return NULL;
    }
}

/* -----------------------------------------------------------------------------
 *  Initializes *info from SRT information stored in *infoTable.
 * -------------------------------------------------------------------------- */
static INLINE void
init_srt_fun( stackPos *info, const StgFunInfoTable *infoTable )
{
    info->type = posTypeSRT;
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_FUN_SRT(infoTable);
    } else {
        info->next.srt.srt = NULL;
    }
}

static INLINE void
init_srt_thunk( stackPos *info, const StgThunkInfoTable *infoTable )
{
    info->type = posTypeSRT;
    if (infoTable->i.srt) {
        info->next.srt.srt = (StgClosure*)GET_SRT(infoTable);
    } else {
        info->next.srt.srt = NULL;
    }
}

/* -----------------------------------------------------------------------------
 * Find the next object from *info.
 * -------------------------------------------------------------------------- */
static INLINE StgClosure *
find_srt( stackPos *info )
{
    StgClosure *c;
    if (info->type == posTypeSRT) {
        c = info->next.srt.srt;
        info->next.srt.srt = NULL;
        return c;
    }
}

/* -----------------------------------------------------------------------------
 * Pushes an element onto traverse stack
 * -------------------------------------------------------------------------- */
static void
retainActualPush(traverseState *ts, stackElement *se) {
    bdescr *nbd;      // Next Block Descriptor
    if (ts->stackTop - 1 < ts->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.
        ts->currentStack->free = (StgPtr)ts->stackTop;

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

        newStackBlock(ts, nbd);
    }

    // adjust stackTop (acutal push)
    ts->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.
    *ts->stackTop = *se;

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

}

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

    se.c = c;
    se.cp = cp;
    se.data = data;
    se.info.type = posTypeFresh;

    retainActualPush(ts, &se);
};

/* -----------------------------------------------------------------------------
 *  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
 *  child, *c_child is set to that child and nothing is pushed onto
 *  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.
 *     *c is not any of TSO, AP, PAP, AP_STACK, which means that
 *        there cannot be any stack objects.
 *  Note: SRTs are considered to  be children as well.
 * -------------------------------------------------------------------------- */
static INLINE void
push( traverseState *ts, StgClosure *c, stackData data, StgClosure **first_child )
{
    stackElement se;
    bdescr *nbd;      // Next Block Descriptor

#if defined(DEBUG_RETAINER)
    debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
#endif

    ASSERT(get_itbl(c)->type != TSO);
    ASSERT(get_itbl(c)->type != AP_STACK);

    //
    // fill in se
    //

    se.c = c;
    se.data = data;
    // Note: se.cp ommitted on purpose, only retainPushClosure uses that.

    // fill in se.info
    switch (get_itbl(c)->type) {
        // no child, no SRT
    case CONSTR_0_1:
    case CONSTR_0_2:
    case ARR_WORDS:
    case COMPACT_NFDATA:
        *first_child = NULL;
        return;

        // one child (fixed), no SRT
    case MUT_VAR_CLEAN:
    case MUT_VAR_DIRTY:
        *first_child = ((StgMutVar *)c)->var;
        return;
    case THUNK_SELECTOR:
        *first_child = ((StgSelector *)c)->selectee;
        return;
    case BLACKHOLE:
        *first_child = ((StgInd *)c)->indirectee;
        return;
    case CONSTR_1_0:
    case CONSTR_1_1:
        *first_child = c->payload[0];
        return;

        // 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;

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

        // three children (fixed), no SRT
        // need to push a stackElement
    case MVAR_CLEAN:
    case MVAR_DIRTY:
        // 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;
        se.info.type = posTypeStep;
        se.info.next.step = 2;            // 2 = second
        break;

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

        // layout.payload.ptrs, no SRT
    case TVAR:
    case CONSTR:
    case CONSTR_NOCAF:
    case PRIM:
    case MUT_PRIM:
    case BCO:
        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
    case MUT_ARR_PTRS_CLEAN:
    case MUT_ARR_PTRS_DIRTY:
    case MUT_ARR_PTRS_FROZEN_CLEAN:
    case MUT_ARR_PTRS_FROZEN_DIRTY:
        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
    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:
        init_ptrs(&se.info, ((StgSmallMutArrPtrs *)c)->ptrs,
                  (StgPtr)(((StgSmallMutArrPtrs *)c)->payload));
        *first_child = find_ptrs(&se.info);
        if (*first_child == NULL)
            return;
        break;

    // layout.payload.ptrs, SRT
    case FUN_STATIC:
    case FUN:           // *c is a heap object.
    case FUN_2_0:
        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;

    case THUNK:
    case THUNK_2_0:
        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
    case FUN_1_0:
    case FUN_1_1:
        *first_child = c->payload[0];
        ASSERT(*first_child != NULL);
        init_srt_fun(&se.info, get_fun_itbl(c));
        break;

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

    case FUN_0_1:      // *c is a heap object.
    case FUN_0_2:
    fun_srt_only:
        init_srt_fun(&se.info, get_fun_itbl(c));
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

    // SRT only
    case THUNK_STATIC:
        ASSERT(get_itbl(c)->srt != 0);
    case THUNK_0_1:
    case THUNK_0_2:
    thunk_srt_only:
        init_srt_thunk(&se.info, get_thunk_itbl(c));
        *first_child = find_srt(&se.info);
        if (*first_child == NULL)
            return;     // no child
        break;

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

        // cannot appear
    case PAP:
    case AP:
    case AP_STACK:
    case TSO:
    case STACK:
    case IND_STATIC:
        // stack objects
    case UPDATE_FRAME:
    case CATCH_FRAME:
    case UNDERFLOW_FRAME:
    case STOP_FRAME:
    case RET_BCO:
    case RET_SMALL:
    case RET_BIG:
        // invalid objects
    case IND:
    case INVALID_OBJECT:
    default:
        barf("Invalid object *c in push(): %d", get_itbl(c)->type);
        return;
    }

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

    retainActualPush(ts, &se);
}

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

#if defined(DEBUG_RETAINER)
    debugBelch("pop() to the previous stack.\n");
#endif

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

    if (ts->firstStack == ts->currentStack) {
        // The stack is completely empty.
        ts->stackTop++;
        ASSERT(ts->stackTop == ts->stackLimit);
#if defined(DEBUG_RETAINER)
        ts->stackSize--;
        if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
        ASSERT(ts->stackSize >= 0);
        debugBelch("stackSize = %d\n", ts->stackSize);
#endif
        return;
    }

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

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

    returnToOldStack(ts, pbd);

#if defined(DEBUG_RETAINER)
    ts->stackSize--;
    if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
    ASSERT(ts->stackSize >= 0);
    debugBelch("stackSize = %d\n", ts->stackSize);
#endif
}

static INLINE void
popOff(traverseState *ts) {
#if defined(DEBUG_RETAINER)
    debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
#endif

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

    // <= (instead of <) is wrong!
    if (ts->stackTop + 1 < ts->stackLimit) {
        ts->stackTop++;
#if defined(DEBUG_RETAINER)
        ts->stackSize--;
        if (ts->stackSize > ts->maxStackSize) ts->maxStackSize = ts->stackSize;
        ASSERT(ts->stackSize >= 0);
        debugBelch("stackSize = %d\n", ts->stackSize);
#endif
        return;
    }

    popOffReal(ts);
}

/* -----------------------------------------------------------------------------
 *  Finds the next object to be considered for retainer profiling and store
 *  its pointer to *c.
 *  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,
 *  and if so, retrieve the first object and store its pointer to *c. Also,
 *  set *cp and *data appropriately, both of which are stored in the stack
 *  element.  The topmost stack element then is overwritten so as for it to now
 *  denote the next object.
 *  If the topmost stack element indicates no more objects are left, pop
 *  off the stack element until either an object can be retrieved or
 *  the current stack chunk becomes empty, indicated by true returned by
 *  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.
 * -------------------------------------------------------------------------- */
static INLINE void
pop( traverseState *ts, StgClosure **c, StgClosure **cp, stackData *data )
{
    stackElement *se;

#if defined(DEBUG_RETAINER)
    debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", ts->stackTop, ts->currentStackBoundary);
#endif

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

        se = ts->stackTop;

        // If this is a top-level element, you should pop that out.
        if (se->info.type == posTypeFresh) {
            *cp = se->cp;
            *c = se->c;
            *data = se->data;
            popOff(ts);
            return;
        }

        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;
            *data = se->data;
            popOff(ts);
            return;

            // three children (fixed), no SRT
            // need to push a stackElement
        case MVAR_CLEAN:
        case MVAR_DIRTY:
            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(ts);
            }
            *cp = se->c;
            *data = se->data;
            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(ts);
            }
            *cp = se->c;
            *data = se->data;
            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;
            uint32_t entry_no = se->info.next.step >> 2;
            uint32_t field_no = se->info.next.step & 3;
            if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
                *c = NULL;
                popOff(ts);
                break;
            }
            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;
            *data = se->data;
            se->info.next.step++;
            return;
        }

        case TVAR:
        case CONSTR:
        case PRIM:
        case MUT_PRIM:
        case BCO:
            // StgMutArrPtr.ptrs, no SRT
        case MUT_ARR_PTRS_CLEAN:
        case MUT_ARR_PTRS_DIRTY:
        case MUT_ARR_PTRS_FROZEN_CLEAN:
        case MUT_ARR_PTRS_FROZEN_DIRTY:
        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:
            *c = find_ptrs(&se->info);
            if (*c == NULL) {
                popOff(ts);
                break;
            }
            *cp = se->c;
            *data = se->data;
            return;

            // layout.payload.ptrs, SRT
        case FUN:         // always a heap object
        case FUN_STATIC:
        case FUN_2_0:
            if (se->info.type == posTypePtrs) {
                *c = find_ptrs(&se->info);
                if (*c != NULL) {
                    *cp = se->c;
                    *data = se->data;
                    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;
                    *data = se->data;
                    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;
                *data = se->data;
                return;
            }
            popOff(ts);
            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:
        case STACK:
        case IND_STATIC:
        case CONSTR_NOCAF:
            // stack objects
        case UPDATE_FRAME:
        case CATCH_FRAME:
        case UNDERFLOW_FRAME:
        case STOP_FRAME:
        case RET_BCO:
        case RET_SMALL:
        case RET_BIG:
            // invalid objects
        case IND:
        case INVALID_OBJECT:
        default: