1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 2001
8 * ---------------------------------------------------------------------------*/
12 // Turn off inlining when debugging - it obfuscates things
21 #include "RetainerProfile.h"
22 #include "RetainerSet.h"
28 #include "Profiling.h"
34 Note: what to change in order to plug-in a new retainer profiling scheme?
35 (1) type retainer in ../includes/StgRetainerProf.h
36 (2) retainer function R(), i.e., getRetainerFrom()
37 (3) the two hashing functions, hashKeySingleton() and hashKeyAddElement(),
38 in RetainerSet.h, if needed.
39 (4) printRetainer() and printRetainerSetShort() in RetainerSet.c.
42 /* -----------------------------------------------------------------------------
44 * -------------------------------------------------------------------------- */
46 static nat retainerGeneration; // generation
48 static nat numObjectVisited; // total number of objects visited
49 static nat timesAnyObjectVisited; // number of times any objects are visited
52 The rs field in the profile header of any object points to its retainer
53 set in an indirect way: if flip is 0, it points to the retainer set;
54 if flip is 1, it points to the next byte after the retainer set (even
55 for NULL pointers). Therefore, with flip 1, (rs ^ 1) is the actual
56 pointer. See retainerSetOf().
59 StgWord flip = 0; // flip bit
60 // must be 0 if DEBUG_RETAINER is on (for static closures)
62 #define setRetainerSetToNull(c) \
63 (c)->header.prof.hp.rs = (RetainerSet *)((StgWord)NULL | flip)
65 static void retainStack(StgClosure *, retainer, StgPtr, StgPtr);
66 static void retainClosure(StgClosure *, StgClosure *, retainer);
68 static void belongToHeap(StgPtr p);
73 cStackSize records how many times retainStack() has been invoked recursively,
74 that is, the number of activation records for retainStack() on the C stack.
75 maxCStackSize records its max value.
77 cStackSize <= maxCStackSize
79 static nat cStackSize, maxCStackSize;
81 static nat sumOfNewCost; // sum of the cost of each object, computed
82 // when the object is first visited
83 static nat sumOfNewCostExtra; // for those objects not visited during
84 // retainer profiling, e.g., MUT_VAR
85 static nat costArray[N_CLOSURE_TYPES];
87 nat sumOfCostLinear; // sum of the costs of all object, computed
88 // when linearly traversing the heap after
90 nat costArrayLinear[N_CLOSURE_TYPES];
93 /* -----------------------------------------------------------------------------
94 * Retainer stack - header
96 * Although the retainer stack implementation could be separated *
97 * from the retainer profiling engine, there does not seem to be
98 * any advantage in doing that; retainer stack is an integral part
99 * of retainer profiling engine and cannot be use elsewhere at
101 * -------------------------------------------------------------------------- */
111 // fixed layout or layout specified by a field in the closure
116 // See StgClosureInfo in InfoTables.h
117 #if SIZEOF_VOID_P == 8
154 firstStack points to the first block group.
155 currentStack points to the block group currently being used.
156 currentStack->free == stackLimit.
157 stackTop points to the topmost byte in the stack of currentStack.
158 Unless the whole stack is empty, stackTop must point to the topmost
159 object (or byte) in the whole stack. Thus, it is only when the whole stack
160 is empty that stackTop == stackLimit (not during the execution of push()
162 stackBottom == currentStack->start.
163 stackLimit == currentStack->start + BLOCK_SIZE_W * currentStack->blocks.
165 When a current stack becomes empty, stackTop is set to point to
166 the topmost element on the previous block group so as to satisfy
167 the invariants described above.
169 static bdescr *firstStack = NULL;
170 static bdescr *currentStack;
171 static stackElement *stackBottom, *stackTop, *stackLimit;
174 currentStackBoundary is used to mark the current stack chunk.
175 If stackTop == currentStackBoundary, it means that the current stack chunk
176 is empty. It is the responsibility of the user to keep currentStackBoundary
177 valid all the time if it is to be employed.
179 static stackElement *currentStackBoundary;
182 stackSize records the current size of the stack.
183 maxStackSize records its high water mark.
185 stackSize <= maxStackSize
187 stackSize is just an estimate measure of the depth of the graph. The reason
188 is that some heap objects have only a single child and may not result
189 in a new element being pushed onto the stack. Therefore, at the end of
190 retainer profiling, maxStackSize + maxCStackSize is some value no greater
191 than the actual depth of the graph.
193 #ifdef DEBUG_RETAINER
194 static int stackSize, maxStackSize;
197 // number of blocks allocated for one stack
198 #define BLOCKS_IN_STACK 1
200 /* -----------------------------------------------------------------------------
201 * Add a new block group to the stack.
203 * currentStack->link == s.
204 * -------------------------------------------------------------------------- */
206 newStackBlock( bdescr *bd )
209 stackTop = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
210 stackBottom = (stackElement *)bd->start;
211 stackLimit = (stackElement *)stackTop;
212 bd->free = (StgPtr)stackLimit;
215 /* -----------------------------------------------------------------------------
216 * Return to the previous block group.
218 * s->link == currentStack.
219 * -------------------------------------------------------------------------- */
221 returnToOldStack( bdescr *bd )
224 stackTop = (stackElement *)bd->free;
225 stackBottom = (stackElement *)bd->start;
226 stackLimit = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
227 bd->free = (StgPtr)stackLimit;
230 /* -----------------------------------------------------------------------------
231 * Initializes the traverse stack.
232 * -------------------------------------------------------------------------- */
234 initializeTraverseStack( void )
236 if (firstStack != NULL) {
237 freeChain(firstStack);
240 firstStack = allocGroup(BLOCKS_IN_STACK);
241 firstStack->link = NULL;
242 firstStack->u.back = NULL;
244 newStackBlock(firstStack);
247 /* -----------------------------------------------------------------------------
248 * Frees all the block groups in the traverse stack.
251 * -------------------------------------------------------------------------- */
253 closeTraverseStack( void )
255 freeChain(firstStack);
259 /* -----------------------------------------------------------------------------
260 * Returns rtsTrue if the whole stack is empty.
261 * -------------------------------------------------------------------------- */
262 static INLINE rtsBool
263 isEmptyRetainerStack( void )
265 return (firstStack == currentStack) && stackTop == stackLimit;
268 /* -----------------------------------------------------------------------------
269 * Returns size of stack
270 * -------------------------------------------------------------------------- */
273 retainerStackBlocks( void )
278 for (bd = firstStack; bd != NULL; bd = bd->link)
285 /* -----------------------------------------------------------------------------
286 * Returns rtsTrue if stackTop is at the stack boundary of the current stack,
287 * i.e., if the current stack chunk is empty.
288 * -------------------------------------------------------------------------- */
289 static INLINE rtsBool
292 return stackTop == currentStackBoundary;
295 /* -----------------------------------------------------------------------------
296 * Initializes *info from ptrs and payload.
298 * payload[] begins with ptrs pointers followed by non-pointers.
299 * -------------------------------------------------------------------------- */
301 init_ptrs( stackPos *info, nat ptrs, StgPtr payload )
303 info->type = posTypePtrs;
304 info->next.ptrs.pos = 0;
305 info->next.ptrs.ptrs = ptrs;
306 info->next.ptrs.payload = payload;
309 /* -----------------------------------------------------------------------------
310 * Find the next object from *info.
311 * -------------------------------------------------------------------------- */
312 static INLINE StgClosure *
313 find_ptrs( stackPos *info )
315 if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
316 return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
322 /* -----------------------------------------------------------------------------
323 * Initializes *info from SRT information stored in *infoTable.
324 * -------------------------------------------------------------------------- */
326 init_srt_fun( stackPos *info, StgFunInfoTable *infoTable )
328 if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
329 info->type = posTypeLargeSRT;
330 info->next.large_srt.srt = (StgLargeSRT *)GET_FUN_SRT(infoTable);
331 info->next.large_srt.offset = 0;
333 info->type = posTypeSRT;
334 info->next.srt.srt = (StgClosure **)GET_FUN_SRT(infoTable);
335 info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
340 init_srt_thunk( stackPos *info, StgThunkInfoTable *infoTable )
342 if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
343 info->type = posTypeLargeSRT;
344 info->next.large_srt.srt = (StgLargeSRT *)GET_SRT(infoTable);
345 info->next.large_srt.offset = 0;
347 info->type = posTypeSRT;
348 info->next.srt.srt = (StgClosure **)GET_SRT(infoTable);
349 info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
353 /* -----------------------------------------------------------------------------
354 * Find the next object from *info.
355 * -------------------------------------------------------------------------- */
356 static INLINE StgClosure *
357 find_srt( stackPos *info )
362 if (info->type == posTypeSRT) {
364 bitmap = info->next.srt.srt_bitmap;
365 while (bitmap != 0) {
366 if ((bitmap & 1) != 0) {
367 #ifdef ENABLE_WIN32_DLL_SUPPORT
369 if ((unsigned long)(*(info->next.srt.srt)) & 0x1)
370 c = (* (StgClosure **)((unsigned long)*(info->next.srt.srt)) & ~0x1);
372 c = *(info->next.srt.srt);
374 c = *(info->next.srt.srt);
376 bitmap = bitmap >> 1;
377 info->next.srt.srt++;
378 info->next.srt.srt_bitmap = bitmap;
381 bitmap = bitmap >> 1;
382 info->next.srt.srt++;
384 // bitmap is now zero...
389 nat i = info->next.large_srt.offset;
392 // Follow the pattern from GC.c:scavenge_large_srt_bitmap().
393 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
394 bitmap = bitmap >> (i % BITS_IN(StgWord));
395 while (i < info->next.large_srt.srt->l.size) {
396 if ((bitmap & 1) != 0) {
397 c = ((StgClosure **)info->next.large_srt.srt->srt)[i];
399 info->next.large_srt.offset = i;
403 if (i % BITS_IN(W_) == 0) {
404 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
406 bitmap = bitmap >> 1;
409 // reached the end of this bitmap.
410 info->next.large_srt.offset = i;
415 /* -----------------------------------------------------------------------------
416 * push() pushes a stackElement representing the next child of *c
417 * onto the traverse stack. If *c has no child, *first_child is set
418 * to NULL and nothing is pushed onto the stack. If *c has only one
419 * child, *c_chlid is set to that child and nothing is pushed onto
420 * the stack. If *c has more than two children, *first_child is set
421 * to the first child and a stackElement representing the second
422 * child is pushed onto the stack.
425 * *c_child_r is the most recent retainer of *c's children.
426 * *c is not any of TSO, AP, PAP, AP_STACK, which means that
427 * there cannot be any stack objects.
428 * Note: SRTs are considered to be children as well.
429 * -------------------------------------------------------------------------- */
431 push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
434 bdescr *nbd; // Next Block Descriptor
436 #ifdef DEBUG_RETAINER
437 // debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
440 ASSERT(get_itbl(c)->type != TSO);
441 ASSERT(get_itbl(c)->type != AP_STACK);
448 se.c_child_r = c_child_r;
451 switch (get_itbl(c)->type) {
458 case SE_CAF_BLACKHOLE:
463 // one child (fixed), no SRT
466 *first_child = ((StgMutVar *)c)->var;
469 *first_child = ((StgSelector *)c)->selectee;
472 case IND_OLDGEN_PERM:
474 *first_child = ((StgInd *)c)->indirectee;
478 *first_child = c->payload[0];
481 // For CONSTR_2_0 and MVAR, we use se.info.step to record the position
482 // of the next child. We do not write a separate initialization code.
483 // Also we do not have to initialize info.type;
485 // two children (fixed), no SRT
486 // need to push a stackElement, but nothing to store in se.info
488 *first_child = c->payload[0]; // return the first pointer
489 // se.info.type = posTypeStep;
490 // se.info.next.step = 2; // 2 = second
493 // three children (fixed), no SRT
494 // 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:
632 barf("Invalid object *c in push()");
636 if (stackTop - 1 < stackBottom) {
637 #ifdef DEBUG_RETAINER
638 // debugBelch("push() to the next stack.\n");
640 // currentStack->free is updated when the active stack is switched
641 // to the next stack.
642 currentStack->free = (StgPtr)stackTop;
644 if (currentStack->link == NULL) {
645 nbd = allocGroup(BLOCKS_IN_STACK);
647 nbd->u.back = currentStack;
648 currentStack->link = nbd;
650 nbd = currentStack->link;
655 // adjust stackTop (acutal push)
657 // If the size of stackElement was huge, we would better replace the
658 // following statement by either a memcpy() call or a switch statement
659 // on the type of the element. Currently, the size of stackElement is
660 // small enough (5 words) that this direct assignment seems to be enough.
662 // ToDo: The line below leads to the warning:
663 // warning: 'se.info.type' may be used uninitialized in this function
664 // This is caused by the fact that there are execution paths through the
665 // large switch statement above where some cases do not initialize this
666 // field. Is this really harmless? Can we avoid the warning?
669 #ifdef DEBUG_RETAINER
671 if (stackSize > maxStackSize) maxStackSize = stackSize;
672 // ASSERT(stackSize >= 0);
673 // debugBelch("stackSize = %d\n", stackSize);
677 /* -----------------------------------------------------------------------------
678 * popOff() and popOffReal(): Pop a stackElement off the traverse stack.
680 * stackTop cannot be equal to stackLimit unless the whole stack is
681 * empty, in which case popOff() is not allowed.
683 * You can think of popOffReal() as a part of popOff() which is
684 * executed at the end of popOff() in necessary. Since popOff() is
685 * likely to be executed quite often while popOffReal() is not, we
686 * separate popOffReal() from popOff(), which is declared as an
687 * INLINE function (for the sake of execution speed). popOffReal()
688 * is called only within popOff() and nowhere else.
689 * -------------------------------------------------------------------------- */
693 bdescr *pbd; // Previous Block Descriptor
695 #ifdef DEBUG_RETAINER
696 // debugBelch("pop() to the previous stack.\n");
699 ASSERT(stackTop + 1 == stackLimit);
700 ASSERT(stackBottom == (stackElement *)currentStack->start);
702 if (firstStack == currentStack) {
703 // The stack is completely empty.
705 ASSERT(stackTop == stackLimit);
706 #ifdef DEBUG_RETAINER
708 if (stackSize > maxStackSize) maxStackSize = stackSize;
710 ASSERT(stackSize >= 0);
711 debugBelch("stackSize = %d\n", stackSize);
717 // currentStack->free is updated when the active stack is switched back
718 // to the previous stack.
719 currentStack->free = (StgPtr)stackLimit;
721 // find the previous block descriptor
722 pbd = currentStack->u.back;
725 returnToOldStack(pbd);
727 #ifdef DEBUG_RETAINER
729 if (stackSize > maxStackSize) maxStackSize = stackSize;
731 ASSERT(stackSize >= 0);
732 debugBelch("stackSize = %d\n", stackSize);
739 #ifdef DEBUG_RETAINER
740 // debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
743 ASSERT(stackTop != stackLimit);
744 ASSERT(!isEmptyRetainerStack());
746 // <= (instead of <) is wrong!
747 if (stackTop + 1 < stackLimit) {
749 #ifdef DEBUG_RETAINER
751 if (stackSize > maxStackSize) maxStackSize = stackSize;
753 ASSERT(stackSize >= 0);
754 debugBelch("stackSize = %d\n", stackSize);
763 /* -----------------------------------------------------------------------------
764 * Finds the next object to be considered for retainer profiling and store
766 * Test if the topmost stack element indicates that more objects are left,
767 * and if so, retrieve the first object and store its pointer to *c. Also,
768 * set *cp and *r appropriately, both of which are stored in the stack element.
769 * The topmost stack element then is overwritten so as for it to now denote
771 * If the topmost stack element indicates no more objects are left, pop
772 * off the stack element until either an object can be retrieved or
773 * the current stack chunk becomes empty, indicated by rtsTrue returned by
774 * isOnBoundary(), in which case *c is set to NULL.
776 * It is okay to call this function even when the current stack chunk
778 * -------------------------------------------------------------------------- */
780 pop( StgClosure **c, StgClosure **cp, retainer *r )
784 #ifdef DEBUG_RETAINER
785 // debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
789 if (isOnBoundary()) { // if the current stack chunk is depleted
796 switch (get_itbl(se->c)->type) {
797 // two children (fixed), no SRT
798 // nothing in se.info
800 *c = se->c->payload[1];
806 // three children (fixed), no SRT
807 // need to push a stackElement
809 if (se->info.next.step == 2) {
810 *c = (StgClosure *)((StgMVar *)se->c)->tail;
811 se->info.next.step++; // move to the next step
814 *c = ((StgMVar *)se->c)->value;
821 // three children (fixed), no SRT
823 if (se->info.next.step == 2) {
824 *c = ((StgWeak *)se->c)->value;
825 se->info.next.step++;
828 *c = ((StgWeak *)se->c)->finalizer;
835 case TVAR_WATCH_QUEUE:
836 if (se->info.next.step == 2) {
837 *c = (StgClosure *)((StgTVarWatchQueue *)se->c)->next_queue_entry;
838 se->info.next.step++; // move to the next step
841 *c = (StgClosure *)((StgTVarWatchQueue *)se->c)->prev_queue_entry;
849 *c = (StgClosure *)((StgTVar *)se->c)->first_watch_queue_entry;
856 *c = (StgClosure *)((StgTRecHeader *)se->c)->current_chunk;
863 // These are pretty complicated: we have N entries, each
864 // of which contains 3 fields that we want to follow. So
865 // we divide the step counter: the 2 low bits indicate
866 // which field, and the rest of the bits indicate the
867 // entry number (starting from zero).
868 nat entry_no = se->info.next.step >> 2;
869 nat field_no = se->info.next.step & 3;
870 if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
875 TRecEntry *entry = &((StgTRecChunk *)se->c)->entries[entry_no];
877 *c = (StgClosure *)entry->tvar;
878 } else if (field_no == 1) {
879 *c = entry->expected_value;
881 *c = entry->new_value;
885 se->info.next.step++;
893 // StgMutArrPtr.ptrs, no SRT
894 case MUT_ARR_PTRS_CLEAN:
895 case MUT_ARR_PTRS_DIRTY:
896 case MUT_ARR_PTRS_FROZEN:
897 case MUT_ARR_PTRS_FROZEN0:
898 *c = find_ptrs(&se->info);
907 // layout.payload.ptrs, SRT
908 case FUN: // always a heap object
910 if (se->info.type == posTypePtrs) {
911 *c = find_ptrs(&se->info);
917 init_srt_fun(&se->info, get_fun_itbl(se->c));
923 if (se->info.type == posTypePtrs) {
924 *c = find_ptrs(&se->info);
930 init_srt_thunk(&se->info, get_thunk_itbl(se->c));
946 *c = find_srt(&se->info);
955 // no child (fixed), no SRT
961 case SE_CAF_BLACKHOLE:
963 // one child (fixed), no SRT
968 case IND_OLDGEN_PERM:
977 case CONSTR_NOCAF_STATIC:
996 barf("Invalid object *c in pop()");
1002 /* -----------------------------------------------------------------------------
1003 * RETAINER PROFILING ENGINE
1004 * -------------------------------------------------------------------------- */
1007 initRetainerProfiling( void )
1009 initializeAllRetainerSet();
1010 retainerGeneration = 0;
1013 /* -----------------------------------------------------------------------------
1014 * This function must be called before f-closing prof_file.
1015 * -------------------------------------------------------------------------- */
1017 endRetainerProfiling( void )
1019 #ifdef SECOND_APPROACH
1020 outputAllRetainerSet(prof_file);
1024 /* -----------------------------------------------------------------------------
1025 * Returns the actual pointer to the retainer set of the closure *c.
1026 * It may adjust RSET(c) subject to flip.
1028 * RSET(c) is initialized to NULL if its current value does not
1031 * Even though this function has side effects, they CAN be ignored because
1032 * subsequent calls to retainerSetOf() always result in the same return value
1033 * and retainerSetOf() is the only way to retrieve retainerSet of a given
1035 * We have to perform an XOR (^) operation each time a closure is examined.
1036 * The reason is that we do not know when a closure is visited last.
1037 * -------------------------------------------------------------------------- */
1039 maybeInitRetainerSet( StgClosure *c )
1041 if (!isRetainerSetFieldValid(c)) {
1042 setRetainerSetToNull(c);
1046 /* -----------------------------------------------------------------------------
1047 * Returns rtsTrue if *c is a retainer.
1048 * -------------------------------------------------------------------------- */
1049 static INLINE rtsBool
1050 isRetainer( StgClosure *c )
1052 switch (get_itbl(c)->type) {
1056 // TSOs MUST be retainers: they constitute the set of roots.
1063 case MUT_ARR_PTRS_CLEAN:
1064 case MUT_ARR_PTRS_DIRTY:
1065 case MUT_ARR_PTRS_FROZEN:
1066 case MUT_ARR_PTRS_FROZEN0:
1068 // thunks are retainers.
1075 case THUNK_SELECTOR:
1079 // Static thunks, or CAFS, are obviously retainers.
1082 // WEAK objects are roots; there is separate code in which traversing
1083 // begins from WEAK objects.
1086 // Since the other mutvar-type things are retainers, seems
1087 // like the right thing to do:
1109 // partial applications
1115 case SE_CAF_BLACKHOLE:
1118 case IND_OLDGEN_PERM:
1128 case TVAR_WATCH_QUEUE:
1136 // IND_STATIC cannot be *c, *cp, *r in the retainer profiling loop.
1138 // CONSTR_NOCAF_STATIC
1139 // cannot be *c, *cp, *r in the retainer profiling loop.
1140 case CONSTR_NOCAF_STATIC:
1141 // Stack objects are invalid because they are never treated as
1142 // legal objects during retainer profiling.
1158 case INVALID_OBJECT:
1160 barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
1165 /* -----------------------------------------------------------------------------
1166 * Returns the retainer function value for the closure *c, i.e., R(*c).
1167 * This function does NOT return the retainer(s) of *c.
1169 * *c must be a retainer.
1171 * Depending on the definition of this function, the maintenance of retainer
1172 * sets can be made easier. If most retainer sets are likely to be created
1173 * again across garbage collections, refreshAllRetainerSet() in
1174 * RetainerSet.c can simply do nothing.
1175 * If this is not the case, we can free all the retainer sets and
1176 * re-initialize the hash table.
1177 * See refreshAllRetainerSet() in RetainerSet.c.
1178 * -------------------------------------------------------------------------- */
1179 static INLINE retainer
1180 getRetainerFrom( StgClosure *c )
1182 ASSERT(isRetainer(c));
1184 #if defined(RETAINER_SCHEME_INFO)
1185 // Retainer scheme 1: retainer = info table
1187 #elif defined(RETAINER_SCHEME_CCS)
1188 // Retainer scheme 2: retainer = cost centre stack
1189 return c->header.prof.ccs;
1190 #elif defined(RETAINER_SCHEME_CC)
1191 // Retainer scheme 3: retainer = cost centre
1192 return c->header.prof.ccs->cc;
1196 /* -----------------------------------------------------------------------------
1197 * Associates the retainer set *s with the closure *c, that is, *s becomes
1198 * the retainer set of *c.
1202 * -------------------------------------------------------------------------- */
1204 associate( StgClosure *c, RetainerSet *s )
1206 // StgWord has the same size as pointers, so the following type
1208 RSET(c) = (RetainerSet *)((StgWord)s | flip);
1211 /* -----------------------------------------------------------------------------
1212 Call retainClosure for each of the closures covered by a large bitmap.
1213 -------------------------------------------------------------------------- */
1216 retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, nat size,
1217 StgClosure *c, retainer c_child_r)
1223 bitmap = large_bitmap->bitmap[b];
1224 for (i = 0; i < size; ) {
1225 if ((bitmap & 1) == 0) {
1226 retainClosure((StgClosure *)*p, c, c_child_r);
1230 if (i % BITS_IN(W_) == 0) {
1232 bitmap = large_bitmap->bitmap[b];
1234 bitmap = bitmap >> 1;
1239 static INLINE StgPtr
1240 retain_small_bitmap (StgPtr p, nat size, StgWord bitmap,
1241 StgClosure *c, retainer c_child_r)
1244 if ((bitmap & 1) == 0) {
1245 retainClosure((StgClosure *)*p, c, c_child_r);
1248 bitmap = bitmap >> 1;
1254 /* -----------------------------------------------------------------------------
1255 * Call retainClosure for each of the closures in an SRT.
1256 * ------------------------------------------------------------------------- */
1259 retain_large_srt_bitmap (StgLargeSRT *srt, StgClosure *c, retainer c_child_r)
1266 p = (StgClosure **)srt->srt;
1268 bitmap = srt->l.bitmap[b];
1269 for (i = 0; i < size; ) {
1270 if ((bitmap & 1) != 0) {
1271 retainClosure((StgClosure *)*p, c, c_child_r);
1275 if (i % BITS_IN(W_) == 0) {
1277 bitmap = srt->l.bitmap[b];
1279 bitmap = bitmap >> 1;
1285 retainSRT (StgClosure **srt, nat srt_bitmap, StgClosure *c, retainer c_child_r)
1290 bitmap = srt_bitmap;
1293 if (bitmap == (StgHalfWord)(-1)) {
1294 retain_large_srt_bitmap( (StgLargeSRT *)srt, c, c_child_r );
1298 while (bitmap != 0) {
1299 if ((bitmap & 1) != 0) {
1300 #ifdef ENABLE_WIN32_DLL_SUPPORT
1301 if ( (unsigned long)(*srt) & 0x1 ) {
1302 retainClosure(*stgCast(StgClosure**,(stgCast(unsigned long, *srt) & ~0x1)),
1305 retainClosure(*srt,c,c_child_r);
1308 retainClosure(*srt,c,c_child_r);
1312 bitmap = bitmap >> 1;
1316 /* -----------------------------------------------------------------------------
1317 * Process all the objects in the stack chunk from stackStart to stackEnd
1318 * with *c and *c_child_r being their parent and their most recent retainer,
1319 * respectively. Treat stackOptionalFun as another child of *c if it is
1322 * *c is one of the following: TSO, AP_STACK.
1323 * If *c is TSO, c == c_child_r.
1324 * stackStart < stackEnd.
1325 * RSET(c) and RSET(c_child_r) are valid, i.e., their
1326 * interpretation conforms to the current value of flip (even when they
1327 * are interpreted to be NULL).
1328 * If *c is TSO, its state is not any of ThreadRelocated, ThreadComplete,
1329 * or ThreadKilled, which means that its stack is ready to process.
1331 * This code was almost plagiarzied from GC.c! For each pointer,
1332 * retainClosure() is invoked instead of evacuate().
1333 * -------------------------------------------------------------------------- */
1335 retainStack( StgClosure *c, retainer c_child_r,
1336 StgPtr stackStart, StgPtr stackEnd )
1338 stackElement *oldStackBoundary;
1340 StgRetInfoTable *info;
1344 #ifdef DEBUG_RETAINER
1346 if (cStackSize > maxCStackSize) maxCStackSize = cStackSize;
1350 Each invocation of retainStack() creates a new virtual
1351 stack. Since all such stacks share a single common stack, we
1352 record the current currentStackBoundary, which will be restored
1355 oldStackBoundary = currentStackBoundary;
1356 currentStackBoundary = stackTop;
1358 #ifdef DEBUG_RETAINER
1359 // debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n", oldStackBoundary, currentStackBoundary);
1362 ASSERT(get_itbl(c)->type != TSO ||
1363 (((StgTSO *)c)->what_next != ThreadRelocated &&
1364 ((StgTSO *)c)->what_next != ThreadComplete &&
1365 ((StgTSO *)c)->what_next != ThreadKilled));
1368 while (p < stackEnd) {
1369 info = get_ret_itbl((StgClosure *)p);
1371 switch(info->i.type) {
1374 retainClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
1375 p += sizeofW(StgUpdateFrame);
1380 case CATCH_STM_FRAME:
1381 case CATCH_RETRY_FRAME:
1382 case ATOMICALLY_FRAME:
1384 bitmap = BITMAP_BITS(info->i.layout.bitmap);
1385 size = BITMAP_SIZE(info->i.layout.bitmap);
1387 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1390 retainSRT((StgClosure **)GET_SRT(info), info->i.srt_bitmap, c, c_child_r);
1397 retainClosure((StgClosure *)*p, c, c_child_r);
1400 size = BCO_BITMAP_SIZE(bco);
1401 retain_large_bitmap(p, BCO_BITMAP(bco), size, c, c_child_r);
1406 // large bitmap (> 32 entries, or > 64 on a 64-bit machine)
1408 size = GET_LARGE_BITMAP(&info->i)->size;
1410 retain_large_bitmap(p, GET_LARGE_BITMAP(&info->i),
1411 size, c, c_child_r);
1413 // and don't forget to follow the SRT
1416 // Dynamic bitmap: the mask is stored on the stack
1419 dyn = ((StgRetDyn *)p)->liveness;
1421 // traverse the bitmap first
1422 bitmap = RET_DYN_LIVENESS(dyn);
1423 p = (P_)&((StgRetDyn *)p)->payload[0];
1424 size = RET_DYN_BITMAP_SIZE;
1425 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1427 // skip over the non-ptr words
1428 p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
1430 // follow the ptr words
1431 for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
1432 retainClosure((StgClosure *)*p, c, c_child_r);
1439 StgRetFun *ret_fun = (StgRetFun *)p;
1440 StgFunInfoTable *fun_info;
1442 retainClosure(ret_fun->fun, c, c_child_r);
1443 fun_info = get_fun_itbl(ret_fun->fun);
1445 p = (P_)&ret_fun->payload;
1446 switch (fun_info->f.fun_type) {
1448 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1449 size = BITMAP_SIZE(fun_info->f.b.bitmap);
1450 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1453 size = GET_FUN_LARGE_BITMAP(fun_info)->size;
1454 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1455 size, c, c_child_r);
1459 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1460 size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
1461 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1468 barf("Invalid object found in retainStack(): %d",
1469 (int)(info->i.type));
1473 // restore currentStackBoundary
1474 currentStackBoundary = oldStackBoundary;
1475 #ifdef DEBUG_RETAINER
1476 // debugBelch("retainStack() finished: currentStackBoundary = 0x%x\n", currentStackBoundary);
1479 #ifdef DEBUG_RETAINER
1484 /* ----------------------------------------------------------------------------
1485 * Call retainClosure for each of the children of a PAP/AP
1486 * ------------------------------------------------------------------------- */
1488 static INLINE StgPtr
1489 retain_PAP_payload (StgClosure *pap, retainer c_child_r, StgClosure *fun,
1490 StgClosure** payload, StgWord n_args)
1494 StgFunInfoTable *fun_info;
1496 retainClosure(fun, pap, c_child_r);
1497 fun_info = get_fun_itbl(fun);
1498 ASSERT(fun_info->i.type != PAP);
1500 p = (StgPtr)payload;
1502 switch (fun_info->f.fun_type) {
1504 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1505 p = retain_small_bitmap(p, n_args, bitmap,
1509 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1510 n_args, pap, c_child_r);
1514 retain_large_bitmap((StgPtr)payload, BCO_BITMAP(fun),
1515 n_args, pap, c_child_r);
1519 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1520 p = retain_small_bitmap(p, n_args, bitmap, pap, c_child_r);
1526 /* -----------------------------------------------------------------------------
1527 * Compute the retainer set of *c0 and all its desecents by traversing.
1528 * *cp0 is the parent of *c0, and *r0 is the most recent retainer of *c0.
1530 * c0 = cp0 = r0 holds only for root objects.
1531 * RSET(cp0) and RSET(r0) are valid, i.e., their
1532 * interpretation conforms to the current value of flip (even when they
1533 * are interpreted to be NULL).
1534 * However, RSET(c0) may be corrupt, i.e., it may not conform to
1535 * the current value of flip. If it does not, during the execution
1536 * of this function, RSET(c0) must be initialized as well as all
1539 * stackTop must be the same at the beginning and the exit of this function.
1540 * *c0 can be TSO (as well as AP_STACK).
1541 * -------------------------------------------------------------------------- */
1543 retainClosure( StgClosure *c0, StgClosure *cp0, retainer r0 )
1545 // c = Current closure
1546 // cp = Current closure's Parent
1547 // r = current closures' most recent Retainer
1548 // c_child_r = current closure's children's most recent retainer
1549 // first_child = first child of c
1550 StgClosure *c, *cp, *first_child;
1551 RetainerSet *s, *retainerSetOfc;
1552 retainer r, c_child_r;
1555 #ifdef DEBUG_RETAINER
1556 // StgPtr oldStackTop;
1559 #ifdef DEBUG_RETAINER
1560 // oldStackTop = stackTop;
1561 // debugBelch("retainClosure() called: c0 = 0x%x, cp0 = 0x%x, r0 = 0x%x\n", c0, cp0, r0);
1564 // (c, cp, r) = (c0, cp0, r0)
1571 //debugBelch("loop");
1572 // pop to (c, cp, r);
1576 #ifdef DEBUG_RETAINER
1577 // debugBelch("retainClosure() ends: oldStackTop = 0x%x, stackTop = 0x%x\n", oldStackTop, stackTop);
1582 //debugBelch("inner_loop");
1585 // c = current closure under consideration,
1586 // cp = current closure's parent,
1587 // r = current closure's most recent retainer
1589 // Loop invariants (on the meaning of c, cp, r, and their retainer sets):
1590 // RSET(cp) and RSET(r) are valid.
1591 // RSET(c) is valid only if c has been visited before.
1593 // Loop invariants (on the relation between c, cp, and r)
1594 // if cp is not a retainer, r belongs to RSET(cp).
1595 // if cp is a retainer, r == cp.
1597 typeOfc = get_itbl(c)->type;
1599 #ifdef DEBUG_RETAINER
1602 case CONSTR_NOCAF_STATIC:
1608 if (retainerSetOf(c) == NULL) { // first visit?
1609 costArray[typeOfc] += cost(c);
1610 sumOfNewCost += cost(c);
1619 if (((StgTSO *)c)->what_next == ThreadComplete ||
1620 ((StgTSO *)c)->what_next == ThreadKilled) {
1621 #ifdef DEBUG_RETAINER
1622 debugBelch("ThreadComplete or ThreadKilled encountered in retainClosure()\n");
1626 if (((StgTSO *)c)->what_next == ThreadRelocated) {
1627 #ifdef DEBUG_RETAINER
1628 debugBelch("ThreadRelocated encountered in retainClosure()\n");
1630 c = (StgClosure *)((StgTSO *)c)->link;
1636 // We just skip IND_STATIC, so its retainer set is never computed.
1637 c = ((StgIndStatic *)c)->indirectee;
1639 // static objects with no pointers out, so goto loop.
1640 case CONSTR_NOCAF_STATIC:
1641 // It is not just enough not to compute the retainer set for *c; it is
1642 // mandatory because CONSTR_NOCAF_STATIC are not reachable from
1643 // scavenged_static_objects, the list from which is assumed to traverse
1644 // all static objects after major garbage collections.
1648 if (get_itbl(c)->srt_bitmap == 0) {
1649 // No need to compute the retainer set; no dynamic objects
1650 // are reachable from *c.
1652 // Static objects: if we traverse all the live closures,
1653 // including static closures, during each heap census then
1654 // we will observe that some static closures appear and
1655 // disappear. eg. a closure may contain a pointer to a
1656 // static function 'f' which is not otherwise reachable
1657 // (it doesn't indirectly point to any CAFs, so it doesn't
1658 // appear in any SRTs), so we would find 'f' during
1659 // traversal. However on the next sweep there may be no
1660 // closures pointing to 'f'.
1662 // We must therefore ignore static closures whose SRT is
1663 // empty, because these are exactly the closures that may
1664 // "appear". A closure with a non-empty SRT, and which is
1665 // still required, will always be reachable.
1667 // But what about CONSTR_STATIC? Surely these may be able
1668 // to appear, and they don't have SRTs, so we can't
1669 // check. So for now, we're calling
1670 // resetStaticObjectForRetainerProfiling() from the
1671 // garbage collector to reset the retainer sets in all the
1672 // reachable static objects.
1679 // The above objects are ignored in computing the average number of times
1680 // an object is visited.
1681 timesAnyObjectVisited++;
1683 // If this is the first visit to c, initialize its retainer set.
1684 maybeInitRetainerSet(c);
1685 retainerSetOfc = retainerSetOf(c);
1688 // isRetainer(cp) == rtsTrue => s == NULL
1689 // isRetainer(cp) == rtsFalse => s == cp.retainer
1693 s = retainerSetOf(cp);
1695 // (c, cp, r, s) is available.
1697 // (c, cp, r, s, R_r) is available, so compute the retainer set for *c.
1698 if (retainerSetOfc == NULL) {
1699 // This is the first visit to *c.
1703 associate(c, singleton(r));
1705 // s is actually the retainer set of *c!
1708 // compute c_child_r
1709 c_child_r = isRetainer(c) ? getRetainerFrom(c) : r;
1711 // This is not the first visit to *c.
1712 if (isMember(r, retainerSetOfc))
1713 goto loop; // no need to process child
1716 associate(c, addElement(r, retainerSetOfc));
1718 // s is not NULL and cp is not a retainer. This means that
1719 // each time *cp is visited, so is *c. Thus, if s has
1720 // exactly one more element in its retainer set than c, s
1721 // is also the new retainer set for *c.
1722 if (s->num == retainerSetOfc->num + 1) {
1725 // Otherwise, just add R_r to the current retainer set of *c.
1727 associate(c, addElement(r, retainerSetOfc));
1732 goto loop; // no need to process child
1734 // compute c_child_r
1738 // now, RSET() of all of *c, *cp, and *r is valid.
1739 // (c, c_child_r) are available.
1743 // Special case closures: we process these all in one go rather
1744 // than attempting to save the current position, because doing so
1748 retainStack(c, c_child_r,
1750 ((StgTSO *)c)->stack + ((StgTSO *)c)->stack_size);
1755 StgPAP *pap = (StgPAP *)c;
1756 retain_PAP_payload(c, c_child_r, pap->fun, pap->payload, pap->n_args);
1762 StgAP *ap = (StgAP *)c;
1763 retain_PAP_payload(c, c_child_r, ap->fun, ap->payload, ap->n_args);
1768 retainClosure(((StgAP_STACK *)c)->fun, c, c_child_r);
1769 retainStack(c, c_child_r,
1770 (StgPtr)((StgAP_STACK *)c)->payload,
1771 (StgPtr)((StgAP_STACK *)c)->payload +
1772 ((StgAP_STACK *)c)->size);
1776 push(c, c_child_r, &first_child);
1778 // If first_child is null, c has no child.
1779 // If first_child is not null, the top stack element points to the next
1780 // object. push() may or may not push a stackElement on the stack.
1781 if (first_child == NULL)
1784 // (c, cp, r) = (first_child, c, c_child_r)
1791 /* -----------------------------------------------------------------------------
1792 * Compute the retainer set for every object reachable from *tl.
1793 * -------------------------------------------------------------------------- */
1795 retainRoot( StgClosure **tl )
1797 // We no longer assume that only TSOs and WEAKs are roots; any closure can
1800 ASSERT(isEmptyRetainerStack());
1801 currentStackBoundary = stackTop;
1803 if (*tl != &stg_END_TSO_QUEUE_closure && isRetainer(*tl)) {
1804 retainClosure(*tl, *tl, getRetainerFrom(*tl));
1806 retainClosure(*tl, *tl, CCS_SYSTEM);
1809 // NOT TRUE: ASSERT(isMember(getRetainerFrom(*tl), retainerSetOf(*tl)));
1810 // *tl might be a TSO which is ThreadComplete, in which
1811 // case we ignore it for the purposes of retainer profiling.
1814 /* -----------------------------------------------------------------------------
1815 * Compute the retainer set for each of the objects in the heap.
1816 * -------------------------------------------------------------------------- */
1818 computeRetainerSet( void )
1825 #ifdef DEBUG_RETAINER
1826 RetainerSet tmpRetainerSet;
1829 GetRoots(retainRoot); // for scheduler roots
1831 // This function is called after a major GC, when key, value, and finalizer
1832 // all are guaranteed to be valid, or reachable.
1834 // The following code assumes that WEAK objects are considered to be roots
1835 // for retainer profilng.
1836 for (weak = weak_ptr_list; weak != NULL; weak = weak->link)
1837 // retainRoot((StgClosure *)weak);
1838 retainRoot((StgClosure **)&weak);
1840 // Consider roots from the stable ptr table.
1841 markStablePtrTable(retainRoot);
1843 // The following code resets the rs field of each unvisited mutable
1844 // object (computing sumOfNewCostExtra and updating costArray[] when
1845 // debugging retainer profiler).
1846 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1847 // NOT TRUE: even G0 has a block on its mutable list
1848 // ASSERT(g != 0 || (generations[g].mut_list == NULL));
1850 // Traversing through mut_list is necessary
1851 // because we can find MUT_VAR objects which have not been
1852 // visited during retainer profiling.
1853 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
1854 for (ml = bd->start; ml < bd->free; ml++) {
1856 maybeInitRetainerSet((StgClosure *)*ml);
1857 rtl = retainerSetOf((StgClosure *)*ml);
1859 #ifdef DEBUG_RETAINER
1861 // first visit to *ml
1862 // This is a violation of the interface rule!
1863 RSET(ml) = (RetainerSet *)((StgWord)(&tmpRetainerSet) | flip);
1865 switch (get_itbl((StgClosure *)ml)->type) {
1869 case CONSTR_NOCAF_STATIC:
1873 barf("Invalid object in computeRetainerSet(): %d", get_itbl((StgClosure*)ml)->type);
1877 costArray[get_itbl((StgClosure *)ml)->type] += cost((StgClosure *)ml);
1878 sumOfNewCostExtra += cost((StgClosure *)ml);
1888 /* -----------------------------------------------------------------------------
1889 * Traverse all static objects for which we compute retainer sets,
1890 * and reset their rs fields to NULL, which is accomplished by
1891 * invoking maybeInitRetainerSet(). This function must be called
1892 * before zeroing all objects reachable from scavenged_static_objects
1893 * in the case of major gabage collections. See GarbageCollect() in
1896 * The mut_once_list of the oldest generation must also be traversed?
1897 * Why? Because if the evacuation of an object pointed to by a static
1898 * indirection object fails, it is put back to the mut_once_list of
1899 * the oldest generation.
1900 * However, this is not necessary because any static indirection objects
1901 * are just traversed through to reach dynamic objects. In other words,
1902 * they are not taken into consideration in computing retainer sets.
1903 * -------------------------------------------------------------------------- */
1905 resetStaticObjectForRetainerProfiling( void )
1907 #ifdef DEBUG_RETAINER
1912 #ifdef DEBUG_RETAINER
1915 p = scavenged_static_objects;
1916 while (p != END_OF_STATIC_LIST) {
1917 #ifdef DEBUG_RETAINER
1920 switch (get_itbl(p)->type) {
1922 // Since we do not compute the retainer set of any
1923 // IND_STATIC object, we don't have to reset its retainer
1925 p = (StgClosure*)*IND_STATIC_LINK(p);
1928 maybeInitRetainerSet(p);
1929 p = (StgClosure*)*THUNK_STATIC_LINK(p);
1932 maybeInitRetainerSet(p);
1933 p = (StgClosure*)*FUN_STATIC_LINK(p);
1936 maybeInitRetainerSet(p);
1937 p = (StgClosure*)*STATIC_LINK(get_itbl(p), p);
1940 barf("resetStaticObjectForRetainerProfiling: %p (%s)",
1941 p, get_itbl(p)->type);
1945 #ifdef DEBUG_RETAINER
1946 // debugBelch("count in scavenged_static_objects = %d\n", count);
1950 /* -----------------------------------------------------------------------------
1951 * Perform retainer profiling.
1952 * N is the oldest generation being profilied, where the generations are
1953 * numbered starting at 0.
1956 * This function should be called only immediately after major garbage
1958 * ------------------------------------------------------------------------- */
1960 retainerProfile(void)
1962 #ifdef DEBUG_RETAINER
1964 nat totalHeapSize; // total raw heap size (computed by linear scanning)
1967 #ifdef DEBUG_RETAINER
1968 debugBelch(" < retainerProfile() invoked : %d>\n", retainerGeneration);
1973 // We haven't flipped the bit yet.
1974 #ifdef DEBUG_RETAINER
1975 debugBelch("Before traversing:\n");
1976 sumOfCostLinear = 0;
1977 for (i = 0;i < N_CLOSURE_TYPES; i++)
1978 costArrayLinear[i] = 0;
1979 totalHeapSize = checkHeapSanityForRetainerProfiling();
1981 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
1983 debugBelch("costArrayLinear[] = ");
1984 for (i = 0;i < N_CLOSURE_TYPES; i++)
1985 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
1989 ASSERT(sumOfCostLinear == totalHeapSize);
1992 #define pcostArrayLinear(index) \
1993 if (costArrayLinear[index] > 0) \
1994 debugBelch("costArrayLinear[" #index "] = %u\n", costArrayLinear[index])
1995 pcostArrayLinear(THUNK_STATIC);
1996 pcostArrayLinear(FUN_STATIC);
1997 pcostArrayLinear(CONSTR_STATIC);
1998 pcostArrayLinear(CONSTR_NOCAF_STATIC);
2002 // Now we flips flip.
2005 #ifdef DEBUG_RETAINER
2011 numObjectVisited = 0;
2012 timesAnyObjectVisited = 0;
2014 #ifdef DEBUG_RETAINER
2015 debugBelch("During traversing:\n");
2017 sumOfNewCostExtra = 0;
2018 for (i = 0;i < N_CLOSURE_TYPES; i++)
2023 We initialize the traverse stack each time the retainer profiling is
2024 performed (because the traverse stack size varies on each retainer profiling
2025 and this operation is not costly anyhow). However, we just refresh the
2028 initializeTraverseStack();
2029 #ifdef DEBUG_RETAINER
2030 initializeAllRetainerSet();
2032 refreshAllRetainerSet();
2034 computeRetainerSet();
2036 #ifdef DEBUG_RETAINER
2037 debugBelch("After traversing:\n");
2038 sumOfCostLinear = 0;
2039 for (i = 0;i < N_CLOSURE_TYPES; i++)
2040 costArrayLinear[i] = 0;
2041 totalHeapSize = checkHeapSanityForRetainerProfiling();
2043 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
2044 ASSERT(sumOfCostLinear == totalHeapSize);
2046 // now, compare the two results
2049 costArray[] must be exactly the same as costArrayLinear[].
2051 1) Dead weak pointers, whose type is CONSTR. These objects are not
2052 reachable from any roots.
2054 debugBelch("Comparison:\n");
2055 debugBelch("\tcostArrayLinear[] (must be empty) = ");
2056 for (i = 0;i < N_CLOSURE_TYPES; i++)
2057 if (costArray[i] != costArrayLinear[i])
2058 // nothing should be printed except MUT_VAR after major GCs
2059 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
2062 debugBelch("\tsumOfNewCost = %u\n", sumOfNewCost);
2063 debugBelch("\tsumOfNewCostExtra = %u\n", sumOfNewCostExtra);
2064 debugBelch("\tcostArray[] (must be empty) = ");
2065 for (i = 0;i < N_CLOSURE_TYPES; i++)
2066 if (costArray[i] != costArrayLinear[i])
2067 // nothing should be printed except MUT_VAR after major GCs
2068 debugBelch("[%u:%u] ", i, costArray[i]);
2071 // only for major garbage collection
2072 ASSERT(sumOfNewCost + sumOfNewCostExtra == sumOfCostLinear);
2076 closeTraverseStack();
2077 #ifdef DEBUG_RETAINER
2078 closeAllRetainerSet();
2080 // Note that there is no post-processing for the retainer sets.
2082 retainerGeneration++;
2085 retainerGeneration - 1, // retainerGeneration has just been incremented!
2086 #ifdef DEBUG_RETAINER
2087 maxCStackSize, maxStackSize,
2089 (double)timesAnyObjectVisited / numObjectVisited);
2092 /* -----------------------------------------------------------------------------
2094 * -------------------------------------------------------------------------- */
2096 #ifdef DEBUG_RETAINER
2098 #define LOOKS_LIKE_PTR(r) ((LOOKS_LIKE_STATIC_CLOSURE(r) || \
2099 ((HEAP_ALLOCED(r) && ((Bdescr((P_)r)->flags & BF_FREE) == 0)))) && \
2100 ((StgWord)(*(StgPtr)r)!=0xaaaaaaaa))
2103 sanityCheckHeapClosure( StgClosure *c )
2107 ASSERT(LOOKS_LIKE_GHC_INFO(c->header.info));
2108 ASSERT(!closure_STATIC(c));
2109 ASSERT(LOOKS_LIKE_PTR(c));
2111 if ((((StgWord)RSET(c) & 1) ^ flip) != 0) {
2112 if (get_itbl(c)->type == CONSTR &&
2113 !strcmp(get_itbl(c)->prof.closure_type, "DEAD_WEAK") &&
2114 !strcmp(get_itbl(c)->prof.closure_desc, "DEAD_WEAK")) {
2115 debugBelch("\tUnvisited dead weak pointer object found: c = %p\n", c);
2116 costArray[get_itbl(c)->type] += cost(c);
2117 sumOfNewCost += cost(c);
2120 "Unvisited object: flip = %d, c = %p(%d, %s, %s), rs = %p\n",
2121 flip, c, get_itbl(c)->type,
2122 get_itbl(c)->prof.closure_type, get_itbl(c)->prof.closure_desc,
2125 // debugBelch("sanityCheckHeapClosure) S: flip = %d, c = %p(%d), rs = %p\n", flip, c, get_itbl(c)->type, RSET(c));
2128 return closure_sizeW(c);
2132 heapCheck( bdescr *bd )
2135 static nat costSum, size;
2138 while (bd != NULL) {
2140 while (p < bd->free) {
2141 size = sanityCheckHeapClosure((StgClosure *)p);
2142 sumOfCostLinear += size;
2143 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2145 // no need for slop check; I think slops are not used currently.
2147 ASSERT(p == bd->free);
2148 costSum += bd->free - bd->start;
2156 smallObjectPoolCheck(void)
2160 static nat costSum, size;
2162 bd = small_alloc_list;
2170 while (p < alloc_Hp) {
2171 size = sanityCheckHeapClosure((StgClosure *)p);
2172 sumOfCostLinear += size;
2173 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2176 ASSERT(p == alloc_Hp);
2177 costSum += alloc_Hp - bd->start;
2180 while (bd != NULL) {
2182 while (p < bd->free) {
2183 size = sanityCheckHeapClosure((StgClosure *)p);
2184 sumOfCostLinear += size;
2185 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2188 ASSERT(p == bd->free);
2189 costSum += bd->free - bd->start;
2197 chainCheck(bdescr *bd)
2202 while (bd != NULL) {
2203 // bd->free - bd->start is not an accurate measurement of the
2204 // object size. Actually it is always zero, so we compute its
2206 size = sanityCheckHeapClosure((StgClosure *)bd->start);
2207 sumOfCostLinear += size;
2208 costArrayLinear[get_itbl((StgClosure *)bd->start)->type] += size;
2217 checkHeapSanityForRetainerProfiling( void )
2222 debugBelch("START: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2223 if (RtsFlags.GcFlags.generations == 1) {
2224 costSum += heapCheck(g0s0->to_blocks);
2225 debugBelch("heapCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2226 costSum += chainCheck(g0s0->large_objects);
2227 debugBelch("chainCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2229 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
2230 for (s = 0; s < generations[g].n_steps; s++) {
2232 After all live objects have been scavenged, the garbage
2233 collector may create some objects in
2234 scheduleFinalizers(). These objects are created throught
2235 allocate(), so the small object pool or the large object
2236 pool of the g0s0 may not be empty.
2238 if (g == 0 && s == 0) {
2239 costSum += smallObjectPoolCheck();
2240 debugBelch("smallObjectPoolCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2241 costSum += chainCheck(generations[g].steps[s].large_objects);
2242 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2244 costSum += heapCheck(generations[g].steps[s].blocks);
2245 debugBelch("heapCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2246 costSum += chainCheck(generations[g].steps[s].large_objects);
2247 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2256 findPointer(StgPtr p)
2262 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2263 for (s = 0; s < generations[g].n_steps; s++) {
2264 // if (g == 0 && s == 0) continue;
2265 bd = generations[g].steps[s].blocks;
2266 for (; bd; bd = bd->link) {
2267 for (q = bd->start; q < bd->free; q++) {
2268 if (*q == (StgWord)p) {
2270 while (!LOOKS_LIKE_GHC_INFO(*r)) r--;
2271 debugBelch("Found in gen[%d], step[%d]: q = %p, r = %p\n", g, s, q, r);
2276 bd = generations[g].steps[s].large_objects;
2277 for (; bd; bd = bd->link) {
2278 e = bd->start + cost((StgClosure *)bd->start);
2279 for (q = bd->start; q < e; q++) {
2280 if (*q == (StgWord)p) {
2282 while (*r == 0 || !LOOKS_LIKE_GHC_INFO(*r)) r--;
2283 debugBelch("Found in gen[%d], large_objects: %p\n", g, r);
2293 belongToHeap(StgPtr p)
2298 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2299 for (s = 0; s < generations[g].n_steps; s++) {
2300 // if (g == 0 && s == 0) continue;
2301 bd = generations[g].steps[s].blocks;
2302 for (; bd; bd = bd->link) {
2303 if (bd->start <= p && p < bd->free) {
2304 debugBelch("Belongs to gen[%d], step[%d]", g, s);
2308 bd = generations[g].steps[s].large_objects;
2309 for (; bd; bd = bd->link) {
2310 if (bd->start <= p && p < bd->start + getHeapClosureSize((StgClosure *)bd->start)) {
2311 debugBelch("Found in gen[%d], large_objects: %p\n", g, bd->start);
2318 #endif /* DEBUG_RETAINER */
2320 #endif /* PROFILING */