1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 2001
8 * ---------------------------------------------------------------------------*/
12 // Turn off inlining when debugging - it obfuscates things
19 #include "PosixSource.h"
23 #include "RetainerProfile.h"
24 #include "RetainerSet.h"
28 #include "sm/Sanity.h"
29 #include "Profiling.h"
33 #include "sm/Storage.h" // for END_OF_STATIC_LIST
36 Note: what to change in order to plug-in a new retainer profiling scheme?
37 (1) type retainer in ../includes/StgRetainerProf.h
38 (2) retainer function R(), i.e., getRetainerFrom()
39 (3) the two hashing functions, hashKeySingleton() and hashKeyAddElement(),
40 in RetainerSet.h, if needed.
41 (4) printRetainer() and printRetainerSetShort() in RetainerSet.c.
44 /* -----------------------------------------------------------------------------
46 * -------------------------------------------------------------------------- */
48 static nat retainerGeneration; // generation
50 static nat numObjectVisited; // total number of objects visited
51 static nat timesAnyObjectVisited; // number of times any objects are visited
54 The rs field in the profile header of any object points to its retainer
55 set in an indirect way: if flip is 0, it points to the retainer set;
56 if flip is 1, it points to the next byte after the retainer set (even
57 for NULL pointers). Therefore, with flip 1, (rs ^ 1) is the actual
58 pointer. See retainerSetOf().
61 StgWord flip = 0; // flip bit
62 // must be 0 if DEBUG_RETAINER is on (for static closures)
64 #define setRetainerSetToNull(c) \
65 (c)->header.prof.hp.rs = (RetainerSet *)((StgWord)NULL | flip)
67 static void retainStack(StgClosure *, retainer, StgPtr, StgPtr);
68 static void retainClosure(StgClosure *, StgClosure *, retainer);
70 static void belongToHeap(StgPtr p);
75 cStackSize records how many times retainStack() has been invoked recursively,
76 that is, the number of activation records for retainStack() on the C stack.
77 maxCStackSize records its max value.
79 cStackSize <= maxCStackSize
81 static nat cStackSize, maxCStackSize;
83 static nat sumOfNewCost; // sum of the cost of each object, computed
84 // when the object is first visited
85 static nat sumOfNewCostExtra; // for those objects not visited during
86 // retainer profiling, e.g., MUT_VAR
87 static nat costArray[N_CLOSURE_TYPES];
89 nat sumOfCostLinear; // sum of the costs of all object, computed
90 // when linearly traversing the heap after
92 nat costArrayLinear[N_CLOSURE_TYPES];
95 /* -----------------------------------------------------------------------------
96 * Retainer stack - header
98 * Although the retainer stack implementation could be separated *
99 * from the retainer profiling engine, there does not seem to be
100 * any advantage in doing that; retainer stack is an integral part
101 * of retainer profiling engine and cannot be use elsewhere at
103 * -------------------------------------------------------------------------- */
113 // fixed layout or layout specified by a field in the closure
118 // See StgClosureInfo in InfoTables.h
119 #if SIZEOF_VOID_P == 8
156 firstStack points to the first block group.
157 currentStack points to the block group currently being used.
158 currentStack->free == stackLimit.
159 stackTop points to the topmost byte in the stack of currentStack.
160 Unless the whole stack is empty, stackTop must point to the topmost
161 object (or byte) in the whole stack. Thus, it is only when the whole stack
162 is empty that stackTop == stackLimit (not during the execution of push()
164 stackBottom == currentStack->start.
165 stackLimit == currentStack->start + BLOCK_SIZE_W * currentStack->blocks.
167 When a current stack becomes empty, stackTop is set to point to
168 the topmost element on the previous block group so as to satisfy
169 the invariants described above.
171 static bdescr *firstStack = NULL;
172 static bdescr *currentStack;
173 static stackElement *stackBottom, *stackTop, *stackLimit;
176 currentStackBoundary is used to mark the current stack chunk.
177 If stackTop == currentStackBoundary, it means that the current stack chunk
178 is empty. It is the responsibility of the user to keep currentStackBoundary
179 valid all the time if it is to be employed.
181 static stackElement *currentStackBoundary;
184 stackSize records the current size of the stack.
185 maxStackSize records its high water mark.
187 stackSize <= maxStackSize
189 stackSize is just an estimate measure of the depth of the graph. The reason
190 is that some heap objects have only a single child and may not result
191 in a new element being pushed onto the stack. Therefore, at the end of
192 retainer profiling, maxStackSize + maxCStackSize is some value no greater
193 than the actual depth of the graph.
195 #ifdef DEBUG_RETAINER
196 static int stackSize, maxStackSize;
199 // number of blocks allocated for one stack
200 #define BLOCKS_IN_STACK 1
202 /* -----------------------------------------------------------------------------
203 * Add a new block group to the stack.
205 * currentStack->link == s.
206 * -------------------------------------------------------------------------- */
208 newStackBlock( bdescr *bd )
211 stackTop = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
212 stackBottom = (stackElement *)bd->start;
213 stackLimit = (stackElement *)stackTop;
214 bd->free = (StgPtr)stackLimit;
217 /* -----------------------------------------------------------------------------
218 * Return to the previous block group.
220 * s->link == currentStack.
221 * -------------------------------------------------------------------------- */
223 returnToOldStack( bdescr *bd )
226 stackTop = (stackElement *)bd->free;
227 stackBottom = (stackElement *)bd->start;
228 stackLimit = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
229 bd->free = (StgPtr)stackLimit;
232 /* -----------------------------------------------------------------------------
233 * Initializes the traverse stack.
234 * -------------------------------------------------------------------------- */
236 initializeTraverseStack( void )
238 if (firstStack != NULL) {
239 freeChain(firstStack);
242 firstStack = allocGroup(BLOCKS_IN_STACK);
243 firstStack->link = NULL;
244 firstStack->u.back = NULL;
246 newStackBlock(firstStack);
249 /* -----------------------------------------------------------------------------
250 * Frees all the block groups in the traverse stack.
253 * -------------------------------------------------------------------------- */
255 closeTraverseStack( void )
257 freeChain(firstStack);
261 /* -----------------------------------------------------------------------------
262 * Returns rtsTrue if the whole stack is empty.
263 * -------------------------------------------------------------------------- */
264 static INLINE rtsBool
265 isEmptyRetainerStack( void )
267 return (firstStack == currentStack) && stackTop == stackLimit;
270 /* -----------------------------------------------------------------------------
271 * Returns size of stack
272 * -------------------------------------------------------------------------- */
275 retainerStackBlocks( void )
280 for (bd = firstStack; bd != NULL; bd = bd->link)
287 /* -----------------------------------------------------------------------------
288 * Returns rtsTrue if stackTop is at the stack boundary of the current stack,
289 * i.e., if the current stack chunk is empty.
290 * -------------------------------------------------------------------------- */
291 static INLINE rtsBool
294 return stackTop == currentStackBoundary;
297 /* -----------------------------------------------------------------------------
298 * Initializes *info from ptrs and payload.
300 * payload[] begins with ptrs pointers followed by non-pointers.
301 * -------------------------------------------------------------------------- */
303 init_ptrs( stackPos *info, nat ptrs, StgPtr payload )
305 info->type = posTypePtrs;
306 info->next.ptrs.pos = 0;
307 info->next.ptrs.ptrs = ptrs;
308 info->next.ptrs.payload = payload;
311 /* -----------------------------------------------------------------------------
312 * Find the next object from *info.
313 * -------------------------------------------------------------------------- */
314 static INLINE StgClosure *
315 find_ptrs( stackPos *info )
317 if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
318 return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
324 /* -----------------------------------------------------------------------------
325 * Initializes *info from SRT information stored in *infoTable.
326 * -------------------------------------------------------------------------- */
328 init_srt_fun( stackPos *info, StgFunInfoTable *infoTable )
330 if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
331 info->type = posTypeLargeSRT;
332 info->next.large_srt.srt = (StgLargeSRT *)GET_FUN_SRT(infoTable);
333 info->next.large_srt.offset = 0;
335 info->type = posTypeSRT;
336 info->next.srt.srt = (StgClosure **)GET_FUN_SRT(infoTable);
337 info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
342 init_srt_thunk( stackPos *info, StgThunkInfoTable *infoTable )
344 if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
345 info->type = posTypeLargeSRT;
346 info->next.large_srt.srt = (StgLargeSRT *)GET_SRT(infoTable);
347 info->next.large_srt.offset = 0;
349 info->type = posTypeSRT;
350 info->next.srt.srt = (StgClosure **)GET_SRT(infoTable);
351 info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
355 /* -----------------------------------------------------------------------------
356 * Find the next object from *info.
357 * -------------------------------------------------------------------------- */
358 static INLINE StgClosure *
359 find_srt( stackPos *info )
364 if (info->type == posTypeSRT) {
366 bitmap = info->next.srt.srt_bitmap;
367 while (bitmap != 0) {
368 if ((bitmap & 1) != 0) {
369 #if defined(__PIC__) && defined(mingw32_TARGET_OS)
370 if ((unsigned long)(*(info->next.srt.srt)) & 0x1)
371 c = (* (StgClosure **)((unsigned long)*(info->next.srt.srt)) & ~0x1);
373 c = *(info->next.srt.srt);
375 c = *(info->next.srt.srt);
377 bitmap = bitmap >> 1;
378 info->next.srt.srt++;
379 info->next.srt.srt_bitmap = bitmap;
382 bitmap = bitmap >> 1;
383 info->next.srt.srt++;
385 // bitmap is now zero...
390 nat i = info->next.large_srt.offset;
393 // Follow the pattern from GC.c:scavenge_large_srt_bitmap().
394 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
395 bitmap = bitmap >> (i % BITS_IN(StgWord));
396 while (i < info->next.large_srt.srt->l.size) {
397 if ((bitmap & 1) != 0) {
398 c = ((StgClosure **)info->next.large_srt.srt->srt)[i];
400 info->next.large_srt.offset = i;
404 if (i % BITS_IN(W_) == 0) {
405 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
407 bitmap = bitmap >> 1;
410 // reached the end of this bitmap.
411 info->next.large_srt.offset = i;
416 /* -----------------------------------------------------------------------------
417 * push() pushes a stackElement representing the next child of *c
418 * onto the traverse stack. If *c has no child, *first_child is set
419 * to NULL and nothing is pushed onto the stack. If *c has only one
420 * child, *c_chlid is set to that child and nothing is pushed onto
421 * the stack. If *c has more than two children, *first_child is set
422 * to the first child and a stackElement representing the second
423 * child is pushed onto the stack.
426 * *c_child_r is the most recent retainer of *c's children.
427 * *c is not any of TSO, AP, PAP, AP_STACK, which means that
428 * there cannot be any stack objects.
429 * Note: SRTs are considered to be children as well.
430 * -------------------------------------------------------------------------- */
432 push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
435 bdescr *nbd; // Next Block Descriptor
437 #ifdef DEBUG_RETAINER
438 // debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
441 ASSERT(get_itbl(c)->type != TSO);
442 ASSERT(get_itbl(c)->type != AP_STACK);
449 se.c_child_r = c_child_r;
452 switch (get_itbl(c)->type) {
462 // one child (fixed), no SRT
465 *first_child = ((StgMutVar *)c)->var;
468 *first_child = ((StgSelector *)c)->selectee;
471 case IND_OLDGEN_PERM:
473 *first_child = ((StgInd *)c)->indirectee;
477 *first_child = c->payload[0];
480 // For CONSTR_2_0 and MVAR, we use se.info.step to record the position
481 // of the next child. We do not write a separate initialization code.
482 // Also we do not have to initialize info.type;
484 // two children (fixed), no SRT
485 // need to push a stackElement, but nothing to store in se.info
487 *first_child = c->payload[0]; // return the first pointer
488 // se.info.type = posTypeStep;
489 // se.info.next.step = 2; // 2 = second
492 // three children (fixed), no SRT
493 // need to push a stackElement
496 // head must be TSO and the head of a linked list of TSOs.
497 // Shoule it be a child? Seems to be yes.
498 *first_child = (StgClosure *)((StgMVar *)c)->head;
499 // se.info.type = posTypeStep;
500 se.info.next.step = 2; // 2 = second
503 // three children (fixed), no SRT
505 *first_child = ((StgWeak *)c)->key;
506 // se.info.type = posTypeStep;
507 se.info.next.step = 2;
510 // layout.payload.ptrs, no SRT
515 init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs,
517 *first_child = find_ptrs(&se.info);
518 if (*first_child == NULL)
522 // StgMutArrPtr.ptrs, no SRT
523 case MUT_ARR_PTRS_CLEAN:
524 case MUT_ARR_PTRS_DIRTY:
525 case MUT_ARR_PTRS_FROZEN:
526 case MUT_ARR_PTRS_FROZEN0:
527 init_ptrs(&se.info, ((StgMutArrPtrs *)c)->ptrs,
528 (StgPtr)(((StgMutArrPtrs *)c)->payload));
529 *first_child = find_ptrs(&se.info);
530 if (*first_child == NULL)
534 // layout.payload.ptrs, SRT
535 case FUN: // *c is a heap object.
537 init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs, (StgPtr)c->payload);
538 *first_child = find_ptrs(&se.info);
539 if (*first_child == NULL)
540 // no child from ptrs, so check SRT
546 init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs,
547 (StgPtr)((StgThunk *)c)->payload);
548 *first_child = find_ptrs(&se.info);
549 if (*first_child == NULL)
550 // no child from ptrs, so check SRT
554 // 1 fixed child, SRT
557 *first_child = c->payload[0];
558 ASSERT(*first_child != NULL);
559 init_srt_fun(&se.info, get_fun_itbl(c));
564 *first_child = ((StgThunk *)c)->payload[0];
565 ASSERT(*first_child != NULL);
566 init_srt_thunk(&se.info, get_thunk_itbl(c));
569 case FUN_STATIC: // *c is a heap object.
570 ASSERT(get_itbl(c)->srt_bitmap != 0);
574 init_srt_fun(&se.info, get_fun_itbl(c));
575 *first_child = find_srt(&se.info);
576 if (*first_child == NULL)
582 ASSERT(get_itbl(c)->srt_bitmap != 0);
586 init_srt_thunk(&se.info, get_thunk_itbl(c));
587 *first_child = find_srt(&se.info);
588 if (*first_child == NULL)
592 case TVAR_WATCH_QUEUE:
593 *first_child = (StgClosure *)((StgTVarWatchQueue *)c)->closure;
594 se.info.next.step = 2; // 2 = second
597 *first_child = (StgClosure *)((StgTVar *)c)->current_value;
600 *first_child = (StgClosure *)((StgTRecHeader *)c)->enclosing_trec;
603 *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
604 se.info.next.step = 0; // entry no.
613 case CONSTR_NOCAF_STATIC:
626 barf("Invalid object *c in push()");
630 if (stackTop - 1 < stackBottom) {
631 #ifdef DEBUG_RETAINER
632 // debugBelch("push() to the next stack.\n");
634 // currentStack->free is updated when the active stack is switched
635 // to the next stack.
636 currentStack->free = (StgPtr)stackTop;
638 if (currentStack->link == NULL) {
639 nbd = allocGroup(BLOCKS_IN_STACK);
641 nbd->u.back = currentStack;
642 currentStack->link = nbd;
644 nbd = currentStack->link;
649 // adjust stackTop (acutal push)
651 // If the size of stackElement was huge, we would better replace the
652 // following statement by either a memcpy() call or a switch statement
653 // on the type of the element. Currently, the size of stackElement is
654 // small enough (5 words) that this direct assignment seems to be enough.
656 // ToDo: The line below leads to the warning:
657 // warning: 'se.info.type' may be used uninitialized in this function
658 // This is caused by the fact that there are execution paths through the
659 // large switch statement above where some cases do not initialize this
660 // field. Is this really harmless? Can we avoid the warning?
663 #ifdef DEBUG_RETAINER
665 if (stackSize > maxStackSize) maxStackSize = stackSize;
666 // ASSERT(stackSize >= 0);
667 // debugBelch("stackSize = %d\n", stackSize);
671 /* -----------------------------------------------------------------------------
672 * popOff() and popOffReal(): Pop a stackElement off the traverse stack.
674 * stackTop cannot be equal to stackLimit unless the whole stack is
675 * empty, in which case popOff() is not allowed.
677 * You can think of popOffReal() as a part of popOff() which is
678 * executed at the end of popOff() in necessary. Since popOff() is
679 * likely to be executed quite often while popOffReal() is not, we
680 * separate popOffReal() from popOff(), which is declared as an
681 * INLINE function (for the sake of execution speed). popOffReal()
682 * is called only within popOff() and nowhere else.
683 * -------------------------------------------------------------------------- */
687 bdescr *pbd; // Previous Block Descriptor
689 #ifdef DEBUG_RETAINER
690 // debugBelch("pop() to the previous stack.\n");
693 ASSERT(stackTop + 1 == stackLimit);
694 ASSERT(stackBottom == (stackElement *)currentStack->start);
696 if (firstStack == currentStack) {
697 // The stack is completely empty.
699 ASSERT(stackTop == stackLimit);
700 #ifdef DEBUG_RETAINER
702 if (stackSize > maxStackSize) maxStackSize = stackSize;
704 ASSERT(stackSize >= 0);
705 debugBelch("stackSize = %d\n", stackSize);
711 // currentStack->free is updated when the active stack is switched back
712 // to the previous stack.
713 currentStack->free = (StgPtr)stackLimit;
715 // find the previous block descriptor
716 pbd = currentStack->u.back;
719 returnToOldStack(pbd);
721 #ifdef DEBUG_RETAINER
723 if (stackSize > maxStackSize) maxStackSize = stackSize;
725 ASSERT(stackSize >= 0);
726 debugBelch("stackSize = %d\n", stackSize);
733 #ifdef DEBUG_RETAINER
734 // debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
737 ASSERT(stackTop != stackLimit);
738 ASSERT(!isEmptyRetainerStack());
740 // <= (instead of <) is wrong!
741 if (stackTop + 1 < stackLimit) {
743 #ifdef DEBUG_RETAINER
745 if (stackSize > maxStackSize) maxStackSize = stackSize;
747 ASSERT(stackSize >= 0);
748 debugBelch("stackSize = %d\n", stackSize);
757 /* -----------------------------------------------------------------------------
758 * Finds the next object to be considered for retainer profiling and store
760 * Test if the topmost stack element indicates that more objects are left,
761 * and if so, retrieve the first object and store its pointer to *c. Also,
762 * set *cp and *r appropriately, both of which are stored in the stack element.
763 * The topmost stack element then is overwritten so as for it to now denote
765 * If the topmost stack element indicates no more objects are left, pop
766 * off the stack element until either an object can be retrieved or
767 * the current stack chunk becomes empty, indicated by rtsTrue returned by
768 * isOnBoundary(), in which case *c is set to NULL.
770 * It is okay to call this function even when the current stack chunk
772 * -------------------------------------------------------------------------- */
774 pop( StgClosure **c, StgClosure **cp, retainer *r )
778 #ifdef DEBUG_RETAINER
779 // debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
783 if (isOnBoundary()) { // if the current stack chunk is depleted
790 switch (get_itbl(se->c)->type) {
791 // two children (fixed), no SRT
792 // nothing in se.info
794 *c = se->c->payload[1];
800 // three children (fixed), no SRT
801 // need to push a stackElement
804 if (se->info.next.step == 2) {
805 *c = (StgClosure *)((StgMVar *)se->c)->tail;
806 se->info.next.step++; // move to the next step
809 *c = ((StgMVar *)se->c)->value;
816 // three children (fixed), no SRT
818 if (se->info.next.step == 2) {
819 *c = ((StgWeak *)se->c)->value;
820 se->info.next.step++;
823 *c = ((StgWeak *)se->c)->finalizer;
830 case TVAR_WATCH_QUEUE:
831 if (se->info.next.step == 2) {
832 *c = (StgClosure *)((StgTVarWatchQueue *)se->c)->next_queue_entry;
833 se->info.next.step++; // move to the next step
836 *c = (StgClosure *)((StgTVarWatchQueue *)se->c)->prev_queue_entry;
844 *c = (StgClosure *)((StgTVar *)se->c)->first_watch_queue_entry;
851 *c = (StgClosure *)((StgTRecHeader *)se->c)->current_chunk;
858 // These are pretty complicated: we have N entries, each
859 // of which contains 3 fields that we want to follow. So
860 // we divide the step counter: the 2 low bits indicate
861 // which field, and the rest of the bits indicate the
862 // entry number (starting from zero).
864 nat entry_no = se->info.next.step >> 2;
865 nat field_no = se->info.next.step & 3;
866 if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
871 entry = &((StgTRecChunk *)se->c)->entries[entry_no];
873 *c = (StgClosure *)entry->tvar;
874 } else if (field_no == 1) {
875 *c = entry->expected_value;
877 *c = entry->new_value;
881 se->info.next.step++;
889 // StgMutArrPtr.ptrs, no SRT
890 case MUT_ARR_PTRS_CLEAN:
891 case MUT_ARR_PTRS_DIRTY:
892 case MUT_ARR_PTRS_FROZEN:
893 case MUT_ARR_PTRS_FROZEN0:
894 *c = find_ptrs(&se->info);
903 // layout.payload.ptrs, SRT
904 case FUN: // always a heap object
906 if (se->info.type == posTypePtrs) {
907 *c = find_ptrs(&se->info);
913 init_srt_fun(&se->info, get_fun_itbl(se->c));
919 if (se->info.type == posTypePtrs) {
920 *c = find_ptrs(&se->info);
926 init_srt_thunk(&se->info, get_thunk_itbl(se->c));
942 *c = find_srt(&se->info);
951 // no child (fixed), no SRT
957 // one child (fixed), no SRT
962 case IND_OLDGEN_PERM:
971 case CONSTR_NOCAF_STATIC:
984 barf("Invalid object *c in pop()");
990 /* -----------------------------------------------------------------------------
991 * RETAINER PROFILING ENGINE
992 * -------------------------------------------------------------------------- */
995 initRetainerProfiling( void )
997 initializeAllRetainerSet();
998 retainerGeneration = 0;
1001 /* -----------------------------------------------------------------------------
1002 * This function must be called before f-closing prof_file.
1003 * -------------------------------------------------------------------------- */
1005 endRetainerProfiling( void )
1007 #ifdef SECOND_APPROACH
1008 outputAllRetainerSet(prof_file);
1012 /* -----------------------------------------------------------------------------
1013 * Returns the actual pointer to the retainer set of the closure *c.
1014 * It may adjust RSET(c) subject to flip.
1016 * RSET(c) is initialized to NULL if its current value does not
1019 * Even though this function has side effects, they CAN be ignored because
1020 * subsequent calls to retainerSetOf() always result in the same return value
1021 * and retainerSetOf() is the only way to retrieve retainerSet of a given
1023 * We have to perform an XOR (^) operation each time a closure is examined.
1024 * The reason is that we do not know when a closure is visited last.
1025 * -------------------------------------------------------------------------- */
1027 maybeInitRetainerSet( StgClosure *c )
1029 if (!isRetainerSetFieldValid(c)) {
1030 setRetainerSetToNull(c);
1034 /* -----------------------------------------------------------------------------
1035 * Returns rtsTrue if *c is a retainer.
1036 * -------------------------------------------------------------------------- */
1037 static INLINE rtsBool
1038 isRetainer( StgClosure *c )
1040 switch (get_itbl(c)->type) {
1044 // TSOs MUST be retainers: they constitute the set of roots.
1052 case MUT_ARR_PTRS_CLEAN:
1053 case MUT_ARR_PTRS_DIRTY:
1054 case MUT_ARR_PTRS_FROZEN:
1055 case MUT_ARR_PTRS_FROZEN0:
1057 // thunks are retainers.
1064 case THUNK_SELECTOR:
1068 // Static thunks, or CAFS, are obviously retainers.
1071 // WEAK objects are roots; there is separate code in which traversing
1072 // begins from WEAK objects.
1075 // Since the other mutvar-type things are retainers, seems
1076 // like the right thing to do:
1098 // partial applications
1105 case IND_OLDGEN_PERM:
1115 case TVAR_WATCH_QUEUE:
1123 // IND_STATIC cannot be *c, *cp, *r in the retainer profiling loop.
1125 // CONSTR_NOCAF_STATIC
1126 // cannot be *c, *cp, *r in the retainer profiling loop.
1127 case CONSTR_NOCAF_STATIC:
1128 // Stack objects are invalid because they are never treated as
1129 // legal objects during retainer profiling.
1139 case INVALID_OBJECT:
1141 barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
1146 /* -----------------------------------------------------------------------------
1147 * Returns the retainer function value for the closure *c, i.e., R(*c).
1148 * This function does NOT return the retainer(s) of *c.
1150 * *c must be a retainer.
1152 * Depending on the definition of this function, the maintenance of retainer
1153 * sets can be made easier. If most retainer sets are likely to be created
1154 * again across garbage collections, refreshAllRetainerSet() in
1155 * RetainerSet.c can simply do nothing.
1156 * If this is not the case, we can free all the retainer sets and
1157 * re-initialize the hash table.
1158 * See refreshAllRetainerSet() in RetainerSet.c.
1159 * -------------------------------------------------------------------------- */
1160 static INLINE retainer
1161 getRetainerFrom( StgClosure *c )
1163 ASSERT(isRetainer(c));
1165 #if defined(RETAINER_SCHEME_INFO)
1166 // Retainer scheme 1: retainer = info table
1168 #elif defined(RETAINER_SCHEME_CCS)
1169 // Retainer scheme 2: retainer = cost centre stack
1170 return c->header.prof.ccs;
1171 #elif defined(RETAINER_SCHEME_CC)
1172 // Retainer scheme 3: retainer = cost centre
1173 return c->header.prof.ccs->cc;
1177 /* -----------------------------------------------------------------------------
1178 * Associates the retainer set *s with the closure *c, that is, *s becomes
1179 * the retainer set of *c.
1183 * -------------------------------------------------------------------------- */
1185 associate( StgClosure *c, RetainerSet *s )
1187 // StgWord has the same size as pointers, so the following type
1189 RSET(c) = (RetainerSet *)((StgWord)s | flip);
1192 /* -----------------------------------------------------------------------------
1193 Call retainClosure for each of the closures covered by a large bitmap.
1194 -------------------------------------------------------------------------- */
1197 retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, nat size,
1198 StgClosure *c, retainer c_child_r)
1204 bitmap = large_bitmap->bitmap[b];
1205 for (i = 0; i < size; ) {
1206 if ((bitmap & 1) == 0) {
1207 retainClosure((StgClosure *)*p, c, c_child_r);
1211 if (i % BITS_IN(W_) == 0) {
1213 bitmap = large_bitmap->bitmap[b];
1215 bitmap = bitmap >> 1;
1220 static INLINE StgPtr
1221 retain_small_bitmap (StgPtr p, nat size, StgWord bitmap,
1222 StgClosure *c, retainer c_child_r)
1225 if ((bitmap & 1) == 0) {
1226 retainClosure((StgClosure *)*p, c, c_child_r);
1229 bitmap = bitmap >> 1;
1235 /* -----------------------------------------------------------------------------
1236 * Call retainClosure for each of the closures in an SRT.
1237 * ------------------------------------------------------------------------- */
1240 retain_large_srt_bitmap (StgLargeSRT *srt, StgClosure *c, retainer c_child_r)
1247 p = (StgClosure **)srt->srt;
1249 bitmap = srt->l.bitmap[b];
1250 for (i = 0; i < size; ) {
1251 if ((bitmap & 1) != 0) {
1252 retainClosure((StgClosure *)*p, c, c_child_r);
1256 if (i % BITS_IN(W_) == 0) {
1258 bitmap = srt->l.bitmap[b];
1260 bitmap = bitmap >> 1;
1266 retainSRT (StgClosure **srt, nat srt_bitmap, StgClosure *c, retainer c_child_r)
1271 bitmap = srt_bitmap;
1274 if (bitmap == (StgHalfWord)(-1)) {
1275 retain_large_srt_bitmap( (StgLargeSRT *)srt, c, c_child_r );
1279 while (bitmap != 0) {
1280 if ((bitmap & 1) != 0) {
1281 #if defined(__PIC__) && defined(mingw32_TARGET_OS)
1282 if ( (unsigned long)(*srt) & 0x1 ) {
1283 retainClosure(* (StgClosure**) ((unsigned long) (*srt) & ~0x1),
1286 retainClosure(*srt,c,c_child_r);
1289 retainClosure(*srt,c,c_child_r);
1293 bitmap = bitmap >> 1;
1297 /* -----------------------------------------------------------------------------
1298 * Process all the objects in the stack chunk from stackStart to stackEnd
1299 * with *c and *c_child_r being their parent and their most recent retainer,
1300 * respectively. Treat stackOptionalFun as another child of *c if it is
1303 * *c is one of the following: TSO, AP_STACK.
1304 * If *c is TSO, c == c_child_r.
1305 * stackStart < stackEnd.
1306 * RSET(c) and RSET(c_child_r) are valid, i.e., their
1307 * interpretation conforms to the current value of flip (even when they
1308 * are interpreted to be NULL).
1309 * If *c is TSO, its state is not any of ThreadRelocated, ThreadComplete,
1310 * or ThreadKilled, which means that its stack is ready to process.
1312 * This code was almost plagiarzied from GC.c! For each pointer,
1313 * retainClosure() is invoked instead of evacuate().
1314 * -------------------------------------------------------------------------- */
1316 retainStack( StgClosure *c, retainer c_child_r,
1317 StgPtr stackStart, StgPtr stackEnd )
1319 stackElement *oldStackBoundary;
1321 StgRetInfoTable *info;
1325 #ifdef DEBUG_RETAINER
1327 if (cStackSize > maxCStackSize) maxCStackSize = cStackSize;
1331 Each invocation of retainStack() creates a new virtual
1332 stack. Since all such stacks share a single common stack, we
1333 record the current currentStackBoundary, which will be restored
1336 oldStackBoundary = currentStackBoundary;
1337 currentStackBoundary = stackTop;
1339 #ifdef DEBUG_RETAINER
1340 // debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n", oldStackBoundary, currentStackBoundary);
1343 ASSERT(get_itbl(c)->type != TSO ||
1344 (((StgTSO *)c)->what_next != ThreadRelocated &&
1345 ((StgTSO *)c)->what_next != ThreadComplete &&
1346 ((StgTSO *)c)->what_next != ThreadKilled));
1349 while (p < stackEnd) {
1350 info = get_ret_itbl((StgClosure *)p);
1352 switch(info->i.type) {
1355 retainClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
1356 p += sizeofW(StgUpdateFrame);
1361 case CATCH_STM_FRAME:
1362 case CATCH_RETRY_FRAME:
1363 case ATOMICALLY_FRAME:
1365 bitmap = BITMAP_BITS(info->i.layout.bitmap);
1366 size = BITMAP_SIZE(info->i.layout.bitmap);
1368 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1371 retainSRT((StgClosure **)GET_SRT(info), info->i.srt_bitmap, c, c_child_r);
1378 retainClosure((StgClosure *)*p, c, c_child_r);
1381 size = BCO_BITMAP_SIZE(bco);
1382 retain_large_bitmap(p, BCO_BITMAP(bco), size, c, c_child_r);
1387 // large bitmap (> 32 entries, or > 64 on a 64-bit machine)
1389 size = GET_LARGE_BITMAP(&info->i)->size;
1391 retain_large_bitmap(p, GET_LARGE_BITMAP(&info->i),
1392 size, c, c_child_r);
1394 // and don't forget to follow the SRT
1397 // Dynamic bitmap: the mask is stored on the stack
1400 dyn = ((StgRetDyn *)p)->liveness;
1402 // traverse the bitmap first
1403 bitmap = RET_DYN_LIVENESS(dyn);
1404 p = (P_)&((StgRetDyn *)p)->payload[0];
1405 size = RET_DYN_BITMAP_SIZE;
1406 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1408 // skip over the non-ptr words
1409 p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
1411 // follow the ptr words
1412 for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
1413 retainClosure((StgClosure *)*p, c, c_child_r);
1420 StgRetFun *ret_fun = (StgRetFun *)p;
1421 StgFunInfoTable *fun_info;
1423 retainClosure(ret_fun->fun, c, c_child_r);
1424 fun_info = get_fun_itbl(UNTAG_CLOSURE(ret_fun->fun));
1426 p = (P_)&ret_fun->payload;
1427 switch (fun_info->f.fun_type) {
1429 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1430 size = BITMAP_SIZE(fun_info->f.b.bitmap);
1431 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1434 size = GET_FUN_LARGE_BITMAP(fun_info)->size;
1435 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1436 size, c, c_child_r);
1440 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1441 size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
1442 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1449 barf("Invalid object found in retainStack(): %d",
1450 (int)(info->i.type));
1454 // restore currentStackBoundary
1455 currentStackBoundary = oldStackBoundary;
1456 #ifdef DEBUG_RETAINER
1457 // debugBelch("retainStack() finished: currentStackBoundary = 0x%x\n", currentStackBoundary);
1460 #ifdef DEBUG_RETAINER
1465 /* ----------------------------------------------------------------------------
1466 * Call retainClosure for each of the children of a PAP/AP
1467 * ------------------------------------------------------------------------- */
1469 static INLINE StgPtr
1470 retain_PAP_payload (StgClosure *pap, /* NOT tagged */
1471 retainer c_child_r, /* NOT tagged */
1472 StgClosure *fun, /* tagged */
1473 StgClosure** payload, StgWord n_args)
1477 StgFunInfoTable *fun_info;
1479 retainClosure(fun, pap, c_child_r);
1480 fun = UNTAG_CLOSURE(fun);
1481 fun_info = get_fun_itbl(fun);
1482 ASSERT(fun_info->i.type != PAP);
1484 p = (StgPtr)payload;
1486 switch (fun_info->f.fun_type) {
1488 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1489 p = retain_small_bitmap(p, n_args, bitmap,
1493 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1494 n_args, pap, c_child_r);
1498 retain_large_bitmap((StgPtr)payload, BCO_BITMAP(fun),
1499 n_args, pap, c_child_r);
1503 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1504 p = retain_small_bitmap(p, n_args, bitmap, pap, c_child_r);
1510 /* -----------------------------------------------------------------------------
1511 * Compute the retainer set of *c0 and all its desecents by traversing.
1512 * *cp0 is the parent of *c0, and *r0 is the most recent retainer of *c0.
1514 * c0 = cp0 = r0 holds only for root objects.
1515 * RSET(cp0) and RSET(r0) are valid, i.e., their
1516 * interpretation conforms to the current value of flip (even when they
1517 * are interpreted to be NULL).
1518 * However, RSET(c0) may be corrupt, i.e., it may not conform to
1519 * the current value of flip. If it does not, during the execution
1520 * of this function, RSET(c0) must be initialized as well as all
1523 * stackTop must be the same at the beginning and the exit of this function.
1524 * *c0 can be TSO (as well as AP_STACK).
1525 * -------------------------------------------------------------------------- */
1527 retainClosure( StgClosure *c0, StgClosure *cp0, retainer r0 )
1529 // c = Current closure (possibly tagged)
1530 // cp = Current closure's Parent (NOT tagged)
1531 // r = current closures' most recent Retainer (NOT tagged)
1532 // c_child_r = current closure's children's most recent retainer
1533 // first_child = first child of c
1534 StgClosure *c, *cp, *first_child;
1535 RetainerSet *s, *retainerSetOfc;
1536 retainer r, c_child_r;
1539 #ifdef DEBUG_RETAINER
1540 // StgPtr oldStackTop;
1543 #ifdef DEBUG_RETAINER
1544 // oldStackTop = stackTop;
1545 // debugBelch("retainClosure() called: c0 = 0x%x, cp0 = 0x%x, r0 = 0x%x\n", c0, cp0, r0);
1548 // (c, cp, r) = (c0, cp0, r0)
1555 //debugBelch("loop");
1556 // pop to (c, cp, r);
1560 #ifdef DEBUG_RETAINER
1561 // debugBelch("retainClosure() ends: oldStackTop = 0x%x, stackTop = 0x%x\n", oldStackTop, stackTop);
1566 //debugBelch("inner_loop");
1569 c = UNTAG_CLOSURE(c);
1571 // c = current closure under consideration,
1572 // cp = current closure's parent,
1573 // r = current closure's most recent retainer
1575 // Loop invariants (on the meaning of c, cp, r, and their retainer sets):
1576 // RSET(cp) and RSET(r) are valid.
1577 // RSET(c) is valid only if c has been visited before.
1579 // Loop invariants (on the relation between c, cp, and r)
1580 // if cp is not a retainer, r belongs to RSET(cp).
1581 // if cp is a retainer, r == cp.
1583 typeOfc = get_itbl(c)->type;
1585 #ifdef DEBUG_RETAINER
1588 case CONSTR_NOCAF_STATIC:
1594 if (retainerSetOf(c) == NULL) { // first visit?
1595 costArray[typeOfc] += cost(c);
1596 sumOfNewCost += cost(c);
1605 if (((StgTSO *)c)->what_next == ThreadComplete ||
1606 ((StgTSO *)c)->what_next == ThreadKilled) {
1607 #ifdef DEBUG_RETAINER
1608 debugBelch("ThreadComplete or ThreadKilled encountered in retainClosure()\n");
1612 if (((StgTSO *)c)->what_next == ThreadRelocated) {
1613 #ifdef DEBUG_RETAINER
1614 debugBelch("ThreadRelocated encountered in retainClosure()\n");
1616 c = (StgClosure *)((StgTSO *)c)->_link;
1622 // We just skip IND_STATIC, so its retainer set is never computed.
1623 c = ((StgIndStatic *)c)->indirectee;
1625 // static objects with no pointers out, so goto loop.
1626 case CONSTR_NOCAF_STATIC:
1627 // It is not just enough not to compute the retainer set for *c; it is
1628 // mandatory because CONSTR_NOCAF_STATIC are not reachable from
1629 // scavenged_static_objects, the list from which is assumed to traverse
1630 // all static objects after major garbage collections.
1634 if (get_itbl(c)->srt_bitmap == 0) {
1635 // No need to compute the retainer set; no dynamic objects
1636 // are reachable from *c.
1638 // Static objects: if we traverse all the live closures,
1639 // including static closures, during each heap census then
1640 // we will observe that some static closures appear and
1641 // disappear. eg. a closure may contain a pointer to a
1642 // static function 'f' which is not otherwise reachable
1643 // (it doesn't indirectly point to any CAFs, so it doesn't
1644 // appear in any SRTs), so we would find 'f' during
1645 // traversal. However on the next sweep there may be no
1646 // closures pointing to 'f'.
1648 // We must therefore ignore static closures whose SRT is
1649 // empty, because these are exactly the closures that may
1650 // "appear". A closure with a non-empty SRT, and which is
1651 // still required, will always be reachable.
1653 // But what about CONSTR_STATIC? Surely these may be able
1654 // to appear, and they don't have SRTs, so we can't
1655 // check. So for now, we're calling
1656 // resetStaticObjectForRetainerProfiling() from the
1657 // garbage collector to reset the retainer sets in all the
1658 // reachable static objects.
1665 // The above objects are ignored in computing the average number of times
1666 // an object is visited.
1667 timesAnyObjectVisited++;
1669 // If this is the first visit to c, initialize its retainer set.
1670 maybeInitRetainerSet(c);
1671 retainerSetOfc = retainerSetOf(c);
1674 // isRetainer(cp) == rtsTrue => s == NULL
1675 // isRetainer(cp) == rtsFalse => s == cp.retainer
1679 s = retainerSetOf(cp);
1681 // (c, cp, r, s) is available.
1683 // (c, cp, r, s, R_r) is available, so compute the retainer set for *c.
1684 if (retainerSetOfc == NULL) {
1685 // This is the first visit to *c.
1689 associate(c, singleton(r));
1691 // s is actually the retainer set of *c!
1694 // compute c_child_r
1695 c_child_r = isRetainer(c) ? getRetainerFrom(c) : r;
1697 // This is not the first visit to *c.
1698 if (isMember(r, retainerSetOfc))
1699 goto loop; // no need to process child
1702 associate(c, addElement(r, retainerSetOfc));
1704 // s is not NULL and cp is not a retainer. This means that
1705 // each time *cp is visited, so is *c. Thus, if s has
1706 // exactly one more element in its retainer set than c, s
1707 // is also the new retainer set for *c.
1708 if (s->num == retainerSetOfc->num + 1) {
1711 // Otherwise, just add R_r to the current retainer set of *c.
1713 associate(c, addElement(r, retainerSetOfc));
1718 goto loop; // no need to process child
1720 // compute c_child_r
1724 // now, RSET() of all of *c, *cp, and *r is valid.
1725 // (c, c_child_r) are available.
1729 // Special case closures: we process these all in one go rather
1730 // than attempting to save the current position, because doing so
1734 retainStack(c, c_child_r,
1736 ((StgTSO *)c)->stack + ((StgTSO *)c)->stack_size);
1741 StgPAP *pap = (StgPAP *)c;
1742 retain_PAP_payload(c, c_child_r, pap->fun, pap->payload, pap->n_args);
1748 StgAP *ap = (StgAP *)c;
1749 retain_PAP_payload(c, c_child_r, ap->fun, ap->payload, ap->n_args);
1754 retainClosure(((StgAP_STACK *)c)->fun, c, c_child_r);
1755 retainStack(c, c_child_r,
1756 (StgPtr)((StgAP_STACK *)c)->payload,
1757 (StgPtr)((StgAP_STACK *)c)->payload +
1758 ((StgAP_STACK *)c)->size);
1762 push(c, c_child_r, &first_child);
1764 // If first_child is null, c has no child.
1765 // If first_child is not null, the top stack element points to the next
1766 // object. push() may or may not push a stackElement on the stack.
1767 if (first_child == NULL)
1770 // (c, cp, r) = (first_child, c, c_child_r)
1777 /* -----------------------------------------------------------------------------
1778 * Compute the retainer set for every object reachable from *tl.
1779 * -------------------------------------------------------------------------- */
1781 retainRoot(void *user STG_UNUSED, StgClosure **tl)
1785 // We no longer assume that only TSOs and WEAKs are roots; any closure can
1788 ASSERT(isEmptyRetainerStack());
1789 currentStackBoundary = stackTop;
1791 c = UNTAG_CLOSURE(*tl);
1792 if (c != &stg_END_TSO_QUEUE_closure && isRetainer(c)) {
1793 retainClosure(c, c, getRetainerFrom(c));
1795 retainClosure(c, c, CCS_SYSTEM);
1798 // NOT TRUE: ASSERT(isMember(getRetainerFrom(*tl), retainerSetOf(*tl)));
1799 // *tl might be a TSO which is ThreadComplete, in which
1800 // case we ignore it for the purposes of retainer profiling.
1803 /* -----------------------------------------------------------------------------
1804 * Compute the retainer set for each of the objects in the heap.
1805 * -------------------------------------------------------------------------- */
1807 computeRetainerSet( void )
1814 #ifdef DEBUG_RETAINER
1815 RetainerSet tmpRetainerSet;
1818 markCapabilities(retainRoot, NULL); // for scheduler roots
1820 // This function is called after a major GC, when key, value, and finalizer
1821 // all are guaranteed to be valid, or reachable.
1823 // The following code assumes that WEAK objects are considered to be roots
1824 // for retainer profilng.
1825 for (weak = weak_ptr_list; weak != NULL; weak = weak->link)
1826 // retainRoot((StgClosure *)weak);
1827 retainRoot(NULL, (StgClosure **)&weak);
1829 // Consider roots from the stable ptr table.
1830 markStablePtrTable(retainRoot, NULL);
1832 // The following code resets the rs field of each unvisited mutable
1833 // object (computing sumOfNewCostExtra and updating costArray[] when
1834 // debugging retainer profiler).
1835 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1836 // NOT TRUE: even G0 has a block on its mutable list
1837 // ASSERT(g != 0 || (generations[g].mut_list == NULL));
1839 // Traversing through mut_list is necessary
1840 // because we can find MUT_VAR objects which have not been
1841 // visited during retainer profiling.
1842 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
1843 for (ml = bd->start; ml < bd->free; ml++) {
1845 maybeInitRetainerSet((StgClosure *)*ml);
1846 rtl = retainerSetOf((StgClosure *)*ml);
1848 #ifdef DEBUG_RETAINER
1850 // first visit to *ml
1851 // This is a violation of the interface rule!
1852 RSET(ml) = (RetainerSet *)((StgWord)(&tmpRetainerSet) | flip);
1854 switch (get_itbl((StgClosure *)ml)->type) {
1858 case CONSTR_NOCAF_STATIC:
1862 barf("Invalid object in computeRetainerSet(): %d", get_itbl((StgClosure*)ml)->type);
1866 costArray[get_itbl((StgClosure *)ml)->type] += cost((StgClosure *)ml);
1867 sumOfNewCostExtra += cost((StgClosure *)ml);
1877 /* -----------------------------------------------------------------------------
1878 * Traverse all static objects for which we compute retainer sets,
1879 * and reset their rs fields to NULL, which is accomplished by
1880 * invoking maybeInitRetainerSet(). This function must be called
1881 * before zeroing all objects reachable from scavenged_static_objects
1882 * in the case of major gabage collections. See GarbageCollect() in
1885 * The mut_once_list of the oldest generation must also be traversed?
1886 * Why? Because if the evacuation of an object pointed to by a static
1887 * indirection object fails, it is put back to the mut_once_list of
1888 * the oldest generation.
1889 * However, this is not necessary because any static indirection objects
1890 * are just traversed through to reach dynamic objects. In other words,
1891 * they are not taken into consideration in computing retainer sets.
1892 * -------------------------------------------------------------------------- */
1894 resetStaticObjectForRetainerProfiling( StgClosure *static_objects )
1896 #ifdef DEBUG_RETAINER
1901 #ifdef DEBUG_RETAINER
1905 while (p != END_OF_STATIC_LIST) {
1906 #ifdef DEBUG_RETAINER
1909 switch (get_itbl(p)->type) {
1911 // Since we do not compute the retainer set of any
1912 // IND_STATIC object, we don't have to reset its retainer
1914 p = (StgClosure*)*IND_STATIC_LINK(p);
1917 maybeInitRetainerSet(p);
1918 p = (StgClosure*)*THUNK_STATIC_LINK(p);
1921 maybeInitRetainerSet(p);
1922 p = (StgClosure*)*FUN_STATIC_LINK(p);
1925 maybeInitRetainerSet(p);
1926 p = (StgClosure*)*STATIC_LINK(get_itbl(p), p);
1929 barf("resetStaticObjectForRetainerProfiling: %p (%s)",
1930 p, get_itbl(p)->type);
1934 #ifdef DEBUG_RETAINER
1935 // debugBelch("count in scavenged_static_objects = %d\n", count);
1939 /* -----------------------------------------------------------------------------
1940 * Perform retainer profiling.
1941 * N is the oldest generation being profilied, where the generations are
1942 * numbered starting at 0.
1945 * This function should be called only immediately after major garbage
1947 * ------------------------------------------------------------------------- */
1949 retainerProfile(void)
1951 #ifdef DEBUG_RETAINER
1953 nat totalHeapSize; // total raw heap size (computed by linear scanning)
1956 #ifdef DEBUG_RETAINER
1957 debugBelch(" < retainerProfile() invoked : %d>\n", retainerGeneration);
1962 // We haven't flipped the bit yet.
1963 #ifdef DEBUG_RETAINER
1964 debugBelch("Before traversing:\n");
1965 sumOfCostLinear = 0;
1966 for (i = 0;i < N_CLOSURE_TYPES; i++)
1967 costArrayLinear[i] = 0;
1968 totalHeapSize = checkHeapSanityForRetainerProfiling();
1970 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
1972 debugBelch("costArrayLinear[] = ");
1973 for (i = 0;i < N_CLOSURE_TYPES; i++)
1974 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
1978 ASSERT(sumOfCostLinear == totalHeapSize);
1981 #define pcostArrayLinear(index) \
1982 if (costArrayLinear[index] > 0) \
1983 debugBelch("costArrayLinear[" #index "] = %u\n", costArrayLinear[index])
1984 pcostArrayLinear(THUNK_STATIC);
1985 pcostArrayLinear(FUN_STATIC);
1986 pcostArrayLinear(CONSTR_STATIC);
1987 pcostArrayLinear(CONSTR_NOCAF_STATIC);
1991 // Now we flips flip.
1994 #ifdef DEBUG_RETAINER
2000 numObjectVisited = 0;
2001 timesAnyObjectVisited = 0;
2003 #ifdef DEBUG_RETAINER
2004 debugBelch("During traversing:\n");
2006 sumOfNewCostExtra = 0;
2007 for (i = 0;i < N_CLOSURE_TYPES; i++)
2012 We initialize the traverse stack each time the retainer profiling is
2013 performed (because the traverse stack size varies on each retainer profiling
2014 and this operation is not costly anyhow). However, we just refresh the
2017 initializeTraverseStack();
2018 #ifdef DEBUG_RETAINER
2019 initializeAllRetainerSet();
2021 refreshAllRetainerSet();
2023 computeRetainerSet();
2025 #ifdef DEBUG_RETAINER
2026 debugBelch("After traversing:\n");
2027 sumOfCostLinear = 0;
2028 for (i = 0;i < N_CLOSURE_TYPES; i++)
2029 costArrayLinear[i] = 0;
2030 totalHeapSize = checkHeapSanityForRetainerProfiling();
2032 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
2033 ASSERT(sumOfCostLinear == totalHeapSize);
2035 // now, compare the two results
2038 costArray[] must be exactly the same as costArrayLinear[].
2040 1) Dead weak pointers, whose type is CONSTR. These objects are not
2041 reachable from any roots.
2043 debugBelch("Comparison:\n");
2044 debugBelch("\tcostArrayLinear[] (must be empty) = ");
2045 for (i = 0;i < N_CLOSURE_TYPES; i++)
2046 if (costArray[i] != costArrayLinear[i])
2047 // nothing should be printed except MUT_VAR after major GCs
2048 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
2051 debugBelch("\tsumOfNewCost = %u\n", sumOfNewCost);
2052 debugBelch("\tsumOfNewCostExtra = %u\n", sumOfNewCostExtra);
2053 debugBelch("\tcostArray[] (must be empty) = ");
2054 for (i = 0;i < N_CLOSURE_TYPES; i++)
2055 if (costArray[i] != costArrayLinear[i])
2056 // nothing should be printed except MUT_VAR after major GCs
2057 debugBelch("[%u:%u] ", i, costArray[i]);
2060 // only for major garbage collection
2061 ASSERT(sumOfNewCost + sumOfNewCostExtra == sumOfCostLinear);
2065 closeTraverseStack();
2066 #ifdef DEBUG_RETAINER
2067 closeAllRetainerSet();
2069 // Note that there is no post-processing for the retainer sets.
2071 retainerGeneration++;
2074 retainerGeneration - 1, // retainerGeneration has just been incremented!
2075 #ifdef DEBUG_RETAINER
2076 maxCStackSize, maxStackSize,
2078 (double)timesAnyObjectVisited / numObjectVisited);
2081 /* -----------------------------------------------------------------------------
2083 * -------------------------------------------------------------------------- */
2085 #ifdef DEBUG_RETAINER
2087 #define LOOKS_LIKE_PTR(r) ((LOOKS_LIKE_STATIC_CLOSURE(r) || \
2088 ((HEAP_ALLOCED(r) && ((Bdescr((P_)r)->flags & BF_FREE) == 0)))) && \
2089 ((StgWord)(*(StgPtr)r)!=0xaaaaaaaa))
2092 sanityCheckHeapClosure( StgClosure *c )
2096 ASSERT(LOOKS_LIKE_GHC_INFO(c->header.info));
2097 ASSERT(!closure_STATIC(c));
2098 ASSERT(LOOKS_LIKE_PTR(c));
2100 if ((((StgWord)RSET(c) & 1) ^ flip) != 0) {
2101 if (get_itbl(c)->type == CONSTR &&
2102 !strcmp(GET_PROF_TYPE(get_itbl(c)), "DEAD_WEAK") &&
2103 !strcmp(GET_PROF_DESC(get_itbl(c)), "DEAD_WEAK")) {
2104 debugBelch("\tUnvisited dead weak pointer object found: c = %p\n", c);
2105 costArray[get_itbl(c)->type] += cost(c);
2106 sumOfNewCost += cost(c);
2109 "Unvisited object: flip = %d, c = %p(%d, %s, %s), rs = %p\n",
2110 flip, c, get_itbl(c)->type,
2111 get_itbl(c)->prof.closure_type, GET_PROF_DESC(get_itbl(c)),
2114 // debugBelch("sanityCheckHeapClosure) S: flip = %d, c = %p(%d), rs = %p\n", flip, c, get_itbl(c)->type, RSET(c));
2117 return closure_sizeW(c);
2121 heapCheck( bdescr *bd )
2124 static nat costSum, size;
2127 while (bd != NULL) {
2129 while (p < bd->free) {
2130 size = sanityCheckHeapClosure((StgClosure *)p);
2131 sumOfCostLinear += size;
2132 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2134 // no need for slop check; I think slops are not used currently.
2136 ASSERT(p == bd->free);
2137 costSum += bd->free - bd->start;
2145 smallObjectPoolCheck(void)
2149 static nat costSum, size;
2159 while (p < alloc_Hp) {
2160 size = sanityCheckHeapClosure((StgClosure *)p);
2161 sumOfCostLinear += size;
2162 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2165 ASSERT(p == alloc_Hp);
2166 costSum += alloc_Hp - bd->start;
2169 while (bd != NULL) {
2171 while (p < bd->free) {
2172 size = sanityCheckHeapClosure((StgClosure *)p);
2173 sumOfCostLinear += size;
2174 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2177 ASSERT(p == bd->free);
2178 costSum += bd->free - bd->start;
2186 chainCheck(bdescr *bd)
2191 while (bd != NULL) {
2192 // bd->free - bd->start is not an accurate measurement of the
2193 // object size. Actually it is always zero, so we compute its
2195 size = sanityCheckHeapClosure((StgClosure *)bd->start);
2196 sumOfCostLinear += size;
2197 costArrayLinear[get_itbl((StgClosure *)bd->start)->type] += size;
2206 checkHeapSanityForRetainerProfiling( void )
2211 debugBelch("START: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2212 if (RtsFlags.GcFlags.generations == 1) {
2213 costSum += heapCheck(g0s0->to_blocks);
2214 debugBelch("heapCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2215 costSum += chainCheck(g0s0->large_objects);
2216 debugBelch("chainCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2218 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
2219 for (s = 0; s < generations[g].n_steps; s++) {
2221 After all live objects have been scavenged, the garbage
2222 collector may create some objects in
2223 scheduleFinalizers(). These objects are created throught
2224 allocate(), so the small object pool or the large object
2225 pool of the g0s0 may not be empty.
2227 if (g == 0 && s == 0) {
2228 costSum += smallObjectPoolCheck();
2229 debugBelch("smallObjectPoolCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2230 costSum += chainCheck(generations[g].steps[s].large_objects);
2231 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2233 costSum += heapCheck(generations[g].steps[s].blocks);
2234 debugBelch("heapCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2235 costSum += chainCheck(generations[g].steps[s].large_objects);
2236 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2245 findPointer(StgPtr p)
2251 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2252 for (s = 0; s < generations[g].n_steps; s++) {
2253 // if (g == 0 && s == 0) continue;
2254 bd = generations[g].steps[s].blocks;
2255 for (; bd; bd = bd->link) {
2256 for (q = bd->start; q < bd->free; q++) {
2257 if (*q == (StgWord)p) {
2259 while (!LOOKS_LIKE_GHC_INFO(*r)) r--;
2260 debugBelch("Found in gen[%d], step[%d]: q = %p, r = %p\n", g, s, q, r);
2265 bd = generations[g].steps[s].large_objects;
2266 for (; bd; bd = bd->link) {
2267 e = bd->start + cost((StgClosure *)bd->start);
2268 for (q = bd->start; q < e; q++) {
2269 if (*q == (StgWord)p) {
2271 while (*r == 0 || !LOOKS_LIKE_GHC_INFO(*r)) r--;
2272 debugBelch("Found in gen[%d], large_objects: %p\n", g, r);
2282 belongToHeap(StgPtr p)
2287 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2288 for (s = 0; s < generations[g].n_steps; s++) {
2289 // if (g == 0 && s == 0) continue;
2290 bd = generations[g].steps[s].blocks;
2291 for (; bd; bd = bd->link) {
2292 if (bd->start <= p && p < bd->free) {
2293 debugBelch("Belongs to gen[%d], step[%d]", g, s);
2297 bd = generations[g].steps[s].large_objects;
2298 for (; bd; bd = bd->link) {
2299 if (bd->start <= p && p < bd->start + getHeapClosureSize((StgClosure *)bd->start)) {
2300 debugBelch("Found in gen[%d], large_objects: %p\n", g, bd->start);
2307 #endif /* DEBUG_RETAINER */
2309 #endif /* PROFILING */