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 #if defined(__PIC__) && defined(mingw32_TARGET_OS)
368 if ((unsigned long)(*(info->next.srt.srt)) & 0x1)
369 c = (* (StgClosure **)((unsigned long)*(info->next.srt.srt)) & ~0x1);
371 c = *(info->next.srt.srt);
373 c = *(info->next.srt.srt);
375 bitmap = bitmap >> 1;
376 info->next.srt.srt++;
377 info->next.srt.srt_bitmap = bitmap;
380 bitmap = bitmap >> 1;
381 info->next.srt.srt++;
383 // bitmap is now zero...
388 nat i = info->next.large_srt.offset;
391 // Follow the pattern from GC.c:scavenge_large_srt_bitmap().
392 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
393 bitmap = bitmap >> (i % BITS_IN(StgWord));
394 while (i < info->next.large_srt.srt->l.size) {
395 if ((bitmap & 1) != 0) {
396 c = ((StgClosure **)info->next.large_srt.srt->srt)[i];
398 info->next.large_srt.offset = i;
402 if (i % BITS_IN(W_) == 0) {
403 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
405 bitmap = bitmap >> 1;
408 // reached the end of this bitmap.
409 info->next.large_srt.offset = i;
414 /* -----------------------------------------------------------------------------
415 * push() pushes a stackElement representing the next child of *c
416 * onto the traverse stack. If *c has no child, *first_child is set
417 * to NULL and nothing is pushed onto the stack. If *c has only one
418 * child, *c_chlid is set to that child and nothing is pushed onto
419 * the stack. If *c has more than two children, *first_child is set
420 * to the first child and a stackElement representing the second
421 * child is pushed onto the stack.
424 * *c_child_r is the most recent retainer of *c's children.
425 * *c is not any of TSO, AP, PAP, AP_STACK, which means that
426 * there cannot be any stack objects.
427 * Note: SRTs are considered to be children as well.
428 * -------------------------------------------------------------------------- */
430 push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
433 bdescr *nbd; // Next Block Descriptor
435 #ifdef DEBUG_RETAINER
436 // debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
439 ASSERT(get_itbl(c)->type != TSO);
440 ASSERT(get_itbl(c)->type != AP_STACK);
447 se.c_child_r = c_child_r;
450 switch (get_itbl(c)->type) {
457 case SE_CAF_BLACKHOLE:
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:
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
810 if (se->info.next.step == 2) {
811 *c = (StgClosure *)((StgMVar *)se->c)->tail;
812 se->info.next.step++; // move to the next step
815 *c = ((StgMVar *)se->c)->value;
822 // three children (fixed), no SRT
824 if (se->info.next.step == 2) {
825 *c = ((StgWeak *)se->c)->value;
826 se->info.next.step++;
829 *c = ((StgWeak *)se->c)->finalizer;
836 case TVAR_WATCH_QUEUE:
837 if (se->info.next.step == 2) {
838 *c = (StgClosure *)((StgTVarWatchQueue *)se->c)->next_queue_entry;
839 se->info.next.step++; // move to the next step
842 *c = (StgClosure *)((StgTVarWatchQueue *)se->c)->prev_queue_entry;
850 *c = (StgClosure *)((StgTVar *)se->c)->first_watch_queue_entry;
857 *c = (StgClosure *)((StgTRecHeader *)se->c)->current_chunk;
864 // These are pretty complicated: we have N entries, each
865 // of which contains 3 fields that we want to follow. So
866 // we divide the step counter: the 2 low bits indicate
867 // which field, and the rest of the bits indicate the
868 // entry number (starting from zero).
870 nat entry_no = se->info.next.step >> 2;
871 nat field_no = se->info.next.step & 3;
872 if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
877 entry = &((StgTRecChunk *)se->c)->entries[entry_no];
879 *c = (StgClosure *)entry->tvar;
880 } else if (field_no == 1) {
881 *c = entry->expected_value;
883 *c = entry->new_value;
887 se->info.next.step++;
895 // StgMutArrPtr.ptrs, no SRT
896 case MUT_ARR_PTRS_CLEAN:
897 case MUT_ARR_PTRS_DIRTY:
898 case MUT_ARR_PTRS_FROZEN:
899 case MUT_ARR_PTRS_FROZEN0:
900 *c = find_ptrs(&se->info);
909 // layout.payload.ptrs, SRT
910 case FUN: // always a heap object
912 if (se->info.type == posTypePtrs) {
913 *c = find_ptrs(&se->info);
919 init_srt_fun(&se->info, get_fun_itbl(se->c));
925 if (se->info.type == posTypePtrs) {
926 *c = find_ptrs(&se->info);
932 init_srt_thunk(&se->info, get_thunk_itbl(se->c));
948 *c = find_srt(&se->info);
957 // no child (fixed), no SRT
963 case SE_CAF_BLACKHOLE:
965 // one child (fixed), no SRT
970 case IND_OLDGEN_PERM:
979 case CONSTR_NOCAF_STATIC:
998 barf("Invalid object *c in pop()");
1004 /* -----------------------------------------------------------------------------
1005 * RETAINER PROFILING ENGINE
1006 * -------------------------------------------------------------------------- */
1009 initRetainerProfiling( void )
1011 initializeAllRetainerSet();
1012 retainerGeneration = 0;
1015 /* -----------------------------------------------------------------------------
1016 * This function must be called before f-closing prof_file.
1017 * -------------------------------------------------------------------------- */
1019 endRetainerProfiling( void )
1021 #ifdef SECOND_APPROACH
1022 outputAllRetainerSet(prof_file);
1026 /* -----------------------------------------------------------------------------
1027 * Returns the actual pointer to the retainer set of the closure *c.
1028 * It may adjust RSET(c) subject to flip.
1030 * RSET(c) is initialized to NULL if its current value does not
1033 * Even though this function has side effects, they CAN be ignored because
1034 * subsequent calls to retainerSetOf() always result in the same return value
1035 * and retainerSetOf() is the only way to retrieve retainerSet of a given
1037 * We have to perform an XOR (^) operation each time a closure is examined.
1038 * The reason is that we do not know when a closure is visited last.
1039 * -------------------------------------------------------------------------- */
1041 maybeInitRetainerSet( StgClosure *c )
1043 if (!isRetainerSetFieldValid(c)) {
1044 setRetainerSetToNull(c);
1048 /* -----------------------------------------------------------------------------
1049 * Returns rtsTrue if *c is a retainer.
1050 * -------------------------------------------------------------------------- */
1051 static INLINE rtsBool
1052 isRetainer( StgClosure *c )
1054 switch (get_itbl(c)->type) {
1058 // TSOs MUST be retainers: they constitute the set of roots.
1066 case MUT_ARR_PTRS_CLEAN:
1067 case MUT_ARR_PTRS_DIRTY:
1068 case MUT_ARR_PTRS_FROZEN:
1069 case MUT_ARR_PTRS_FROZEN0:
1071 // thunks are retainers.
1078 case THUNK_SELECTOR:
1082 // Static thunks, or CAFS, are obviously retainers.
1085 // WEAK objects are roots; there is separate code in which traversing
1086 // begins from WEAK objects.
1089 // Since the other mutvar-type things are retainers, seems
1090 // like the right thing to do:
1112 // partial applications
1118 case SE_CAF_BLACKHOLE:
1121 case IND_OLDGEN_PERM:
1131 case TVAR_WATCH_QUEUE:
1139 // IND_STATIC cannot be *c, *cp, *r in the retainer profiling loop.
1141 // CONSTR_NOCAF_STATIC
1142 // cannot be *c, *cp, *r in the retainer profiling loop.
1143 case CONSTR_NOCAF_STATIC:
1144 // Stack objects are invalid because they are never treated as
1145 // legal objects during retainer profiling.
1161 case INVALID_OBJECT:
1163 barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
1168 /* -----------------------------------------------------------------------------
1169 * Returns the retainer function value for the closure *c, i.e., R(*c).
1170 * This function does NOT return the retainer(s) of *c.
1172 * *c must be a retainer.
1174 * Depending on the definition of this function, the maintenance of retainer
1175 * sets can be made easier. If most retainer sets are likely to be created
1176 * again across garbage collections, refreshAllRetainerSet() in
1177 * RetainerSet.c can simply do nothing.
1178 * If this is not the case, we can free all the retainer sets and
1179 * re-initialize the hash table.
1180 * See refreshAllRetainerSet() in RetainerSet.c.
1181 * -------------------------------------------------------------------------- */
1182 static INLINE retainer
1183 getRetainerFrom( StgClosure *c )
1185 ASSERT(isRetainer(c));
1187 #if defined(RETAINER_SCHEME_INFO)
1188 // Retainer scheme 1: retainer = info table
1190 #elif defined(RETAINER_SCHEME_CCS)
1191 // Retainer scheme 2: retainer = cost centre stack
1192 return c->header.prof.ccs;
1193 #elif defined(RETAINER_SCHEME_CC)
1194 // Retainer scheme 3: retainer = cost centre
1195 return c->header.prof.ccs->cc;
1199 /* -----------------------------------------------------------------------------
1200 * Associates the retainer set *s with the closure *c, that is, *s becomes
1201 * the retainer set of *c.
1205 * -------------------------------------------------------------------------- */
1207 associate( StgClosure *c, RetainerSet *s )
1209 // StgWord has the same size as pointers, so the following type
1211 RSET(c) = (RetainerSet *)((StgWord)s | flip);
1214 /* -----------------------------------------------------------------------------
1215 Call retainClosure for each of the closures covered by a large bitmap.
1216 -------------------------------------------------------------------------- */
1219 retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, nat size,
1220 StgClosure *c, retainer c_child_r)
1226 bitmap = large_bitmap->bitmap[b];
1227 for (i = 0; i < size; ) {
1228 if ((bitmap & 1) == 0) {
1229 retainClosure((StgClosure *)*p, c, c_child_r);
1233 if (i % BITS_IN(W_) == 0) {
1235 bitmap = large_bitmap->bitmap[b];
1237 bitmap = bitmap >> 1;
1242 static INLINE StgPtr
1243 retain_small_bitmap (StgPtr p, nat size, StgWord bitmap,
1244 StgClosure *c, retainer c_child_r)
1247 if ((bitmap & 1) == 0) {
1248 retainClosure((StgClosure *)*p, c, c_child_r);
1251 bitmap = bitmap >> 1;
1257 /* -----------------------------------------------------------------------------
1258 * Call retainClosure for each of the closures in an SRT.
1259 * ------------------------------------------------------------------------- */
1262 retain_large_srt_bitmap (StgLargeSRT *srt, StgClosure *c, retainer c_child_r)
1269 p = (StgClosure **)srt->srt;
1271 bitmap = srt->l.bitmap[b];
1272 for (i = 0; i < size; ) {
1273 if ((bitmap & 1) != 0) {
1274 retainClosure((StgClosure *)*p, c, c_child_r);
1278 if (i % BITS_IN(W_) == 0) {
1280 bitmap = srt->l.bitmap[b];
1282 bitmap = bitmap >> 1;
1288 retainSRT (StgClosure **srt, nat srt_bitmap, StgClosure *c, retainer c_child_r)
1293 bitmap = srt_bitmap;
1296 if (bitmap == (StgHalfWord)(-1)) {
1297 retain_large_srt_bitmap( (StgLargeSRT *)srt, c, c_child_r );
1301 while (bitmap != 0) {
1302 if ((bitmap & 1) != 0) {
1303 #ifdef ENABLE_WIN32_DLL_SUPPORT
1304 if ( (unsigned long)(*srt) & 0x1 ) {
1305 retainClosure(*stgCast(StgClosure**,(stgCast(unsigned long, *srt) & ~0x1)),
1308 retainClosure(*srt,c,c_child_r);
1311 retainClosure(*srt,c,c_child_r);
1315 bitmap = bitmap >> 1;
1319 /* -----------------------------------------------------------------------------
1320 * Process all the objects in the stack chunk from stackStart to stackEnd
1321 * with *c and *c_child_r being their parent and their most recent retainer,
1322 * respectively. Treat stackOptionalFun as another child of *c if it is
1325 * *c is one of the following: TSO, AP_STACK.
1326 * If *c is TSO, c == c_child_r.
1327 * stackStart < stackEnd.
1328 * RSET(c) and RSET(c_child_r) are valid, i.e., their
1329 * interpretation conforms to the current value of flip (even when they
1330 * are interpreted to be NULL).
1331 * If *c is TSO, its state is not any of ThreadRelocated, ThreadComplete,
1332 * or ThreadKilled, which means that its stack is ready to process.
1334 * This code was almost plagiarzied from GC.c! For each pointer,
1335 * retainClosure() is invoked instead of evacuate().
1336 * -------------------------------------------------------------------------- */
1338 retainStack( StgClosure *c, retainer c_child_r,
1339 StgPtr stackStart, StgPtr stackEnd )
1341 stackElement *oldStackBoundary;
1343 StgRetInfoTable *info;
1347 #ifdef DEBUG_RETAINER
1349 if (cStackSize > maxCStackSize) maxCStackSize = cStackSize;
1353 Each invocation of retainStack() creates a new virtual
1354 stack. Since all such stacks share a single common stack, we
1355 record the current currentStackBoundary, which will be restored
1358 oldStackBoundary = currentStackBoundary;
1359 currentStackBoundary = stackTop;
1361 #ifdef DEBUG_RETAINER
1362 // debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n", oldStackBoundary, currentStackBoundary);
1365 ASSERT(get_itbl(c)->type != TSO ||
1366 (((StgTSO *)c)->what_next != ThreadRelocated &&
1367 ((StgTSO *)c)->what_next != ThreadComplete &&
1368 ((StgTSO *)c)->what_next != ThreadKilled));
1371 while (p < stackEnd) {
1372 info = get_ret_itbl((StgClosure *)p);
1374 switch(info->i.type) {
1377 retainClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
1378 p += sizeofW(StgUpdateFrame);
1383 case CATCH_STM_FRAME:
1384 case CATCH_RETRY_FRAME:
1385 case ATOMICALLY_FRAME:
1387 bitmap = BITMAP_BITS(info->i.layout.bitmap);
1388 size = BITMAP_SIZE(info->i.layout.bitmap);
1390 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1393 retainSRT((StgClosure **)GET_SRT(info), info->i.srt_bitmap, c, c_child_r);
1400 retainClosure((StgClosure *)*p, c, c_child_r);
1403 size = BCO_BITMAP_SIZE(bco);
1404 retain_large_bitmap(p, BCO_BITMAP(bco), size, c, c_child_r);
1409 // large bitmap (> 32 entries, or > 64 on a 64-bit machine)
1411 size = GET_LARGE_BITMAP(&info->i)->size;
1413 retain_large_bitmap(p, GET_LARGE_BITMAP(&info->i),
1414 size, c, c_child_r);
1416 // and don't forget to follow the SRT
1419 // Dynamic bitmap: the mask is stored on the stack
1422 dyn = ((StgRetDyn *)p)->liveness;
1424 // traverse the bitmap first
1425 bitmap = RET_DYN_LIVENESS(dyn);
1426 p = (P_)&((StgRetDyn *)p)->payload[0];
1427 size = RET_DYN_BITMAP_SIZE;
1428 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1430 // skip over the non-ptr words
1431 p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
1433 // follow the ptr words
1434 for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
1435 retainClosure((StgClosure *)*p, c, c_child_r);
1442 StgRetFun *ret_fun = (StgRetFun *)p;
1443 StgFunInfoTable *fun_info;
1445 retainClosure(ret_fun->fun, c, c_child_r);
1446 fun_info = get_fun_itbl(UNTAG_CLOSURE(ret_fun->fun));
1448 p = (P_)&ret_fun->payload;
1449 switch (fun_info->f.fun_type) {
1451 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1452 size = BITMAP_SIZE(fun_info->f.b.bitmap);
1453 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1456 size = GET_FUN_LARGE_BITMAP(fun_info)->size;
1457 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1458 size, c, c_child_r);
1462 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1463 size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
1464 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1471 barf("Invalid object found in retainStack(): %d",
1472 (int)(info->i.type));
1476 // restore currentStackBoundary
1477 currentStackBoundary = oldStackBoundary;
1478 #ifdef DEBUG_RETAINER
1479 // debugBelch("retainStack() finished: currentStackBoundary = 0x%x\n", currentStackBoundary);
1482 #ifdef DEBUG_RETAINER
1487 /* ----------------------------------------------------------------------------
1488 * Call retainClosure for each of the children of a PAP/AP
1489 * ------------------------------------------------------------------------- */
1491 static INLINE StgPtr
1492 retain_PAP_payload (StgClosure *pap, /* NOT tagged */
1493 retainer c_child_r, /* NOT tagged */
1494 StgClosure *fun, /* tagged */
1495 StgClosure** payload, StgWord n_args)
1499 StgFunInfoTable *fun_info;
1501 retainClosure(fun, pap, c_child_r);
1502 fun = UNTAG_CLOSURE(fun);
1503 fun_info = get_fun_itbl(fun);
1504 ASSERT(fun_info->i.type != PAP);
1506 p = (StgPtr)payload;
1508 switch (fun_info->f.fun_type) {
1510 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1511 p = retain_small_bitmap(p, n_args, bitmap,
1515 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1516 n_args, pap, c_child_r);
1520 retain_large_bitmap((StgPtr)payload, BCO_BITMAP(fun),
1521 n_args, pap, c_child_r);
1525 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1526 p = retain_small_bitmap(p, n_args, bitmap, pap, c_child_r);
1532 /* -----------------------------------------------------------------------------
1533 * Compute the retainer set of *c0 and all its desecents by traversing.
1534 * *cp0 is the parent of *c0, and *r0 is the most recent retainer of *c0.
1536 * c0 = cp0 = r0 holds only for root objects.
1537 * RSET(cp0) and RSET(r0) are valid, i.e., their
1538 * interpretation conforms to the current value of flip (even when they
1539 * are interpreted to be NULL).
1540 * However, RSET(c0) may be corrupt, i.e., it may not conform to
1541 * the current value of flip. If it does not, during the execution
1542 * of this function, RSET(c0) must be initialized as well as all
1545 * stackTop must be the same at the beginning and the exit of this function.
1546 * *c0 can be TSO (as well as AP_STACK).
1547 * -------------------------------------------------------------------------- */
1549 retainClosure( StgClosure *c0, StgClosure *cp0, retainer r0 )
1551 // c = Current closure (possibly tagged)
1552 // cp = Current closure's Parent (NOT tagged)
1553 // r = current closures' most recent Retainer (NOT tagged)
1554 // c_child_r = current closure's children's most recent retainer
1555 // first_child = first child of c
1556 StgClosure *c, *cp, *first_child;
1557 RetainerSet *s, *retainerSetOfc;
1558 retainer r, c_child_r;
1561 #ifdef DEBUG_RETAINER
1562 // StgPtr oldStackTop;
1565 #ifdef DEBUG_RETAINER
1566 // oldStackTop = stackTop;
1567 // debugBelch("retainClosure() called: c0 = 0x%x, cp0 = 0x%x, r0 = 0x%x\n", c0, cp0, r0);
1570 // (c, cp, r) = (c0, cp0, r0)
1577 //debugBelch("loop");
1578 // pop to (c, cp, r);
1582 #ifdef DEBUG_RETAINER
1583 // debugBelch("retainClosure() ends: oldStackTop = 0x%x, stackTop = 0x%x\n", oldStackTop, stackTop);
1588 //debugBelch("inner_loop");
1591 c = UNTAG_CLOSURE(c);
1593 // c = current closure under consideration,
1594 // cp = current closure's parent,
1595 // r = current closure's most recent retainer
1597 // Loop invariants (on the meaning of c, cp, r, and their retainer sets):
1598 // RSET(cp) and RSET(r) are valid.
1599 // RSET(c) is valid only if c has been visited before.
1601 // Loop invariants (on the relation between c, cp, and r)
1602 // if cp is not a retainer, r belongs to RSET(cp).
1603 // if cp is a retainer, r == cp.
1605 typeOfc = get_itbl(c)->type;
1607 #ifdef DEBUG_RETAINER
1610 case CONSTR_NOCAF_STATIC:
1616 if (retainerSetOf(c) == NULL) { // first visit?
1617 costArray[typeOfc] += cost(c);
1618 sumOfNewCost += cost(c);
1627 if (((StgTSO *)c)->what_next == ThreadComplete ||
1628 ((StgTSO *)c)->what_next == ThreadKilled) {
1629 #ifdef DEBUG_RETAINER
1630 debugBelch("ThreadComplete or ThreadKilled encountered in retainClosure()\n");
1634 if (((StgTSO *)c)->what_next == ThreadRelocated) {
1635 #ifdef DEBUG_RETAINER
1636 debugBelch("ThreadRelocated encountered in retainClosure()\n");
1638 c = (StgClosure *)((StgTSO *)c)->link;
1644 // We just skip IND_STATIC, so its retainer set is never computed.
1645 c = ((StgIndStatic *)c)->indirectee;
1647 // static objects with no pointers out, so goto loop.
1648 case CONSTR_NOCAF_STATIC:
1649 // It is not just enough not to compute the retainer set for *c; it is
1650 // mandatory because CONSTR_NOCAF_STATIC are not reachable from
1651 // scavenged_static_objects, the list from which is assumed to traverse
1652 // all static objects after major garbage collections.
1656 if (get_itbl(c)->srt_bitmap == 0) {
1657 // No need to compute the retainer set; no dynamic objects
1658 // are reachable from *c.
1660 // Static objects: if we traverse all the live closures,
1661 // including static closures, during each heap census then
1662 // we will observe that some static closures appear and
1663 // disappear. eg. a closure may contain a pointer to a
1664 // static function 'f' which is not otherwise reachable
1665 // (it doesn't indirectly point to any CAFs, so it doesn't
1666 // appear in any SRTs), so we would find 'f' during
1667 // traversal. However on the next sweep there may be no
1668 // closures pointing to 'f'.
1670 // We must therefore ignore static closures whose SRT is
1671 // empty, because these are exactly the closures that may
1672 // "appear". A closure with a non-empty SRT, and which is
1673 // still required, will always be reachable.
1675 // But what about CONSTR_STATIC? Surely these may be able
1676 // to appear, and they don't have SRTs, so we can't
1677 // check. So for now, we're calling
1678 // resetStaticObjectForRetainerProfiling() from the
1679 // garbage collector to reset the retainer sets in all the
1680 // reachable static objects.
1687 // The above objects are ignored in computing the average number of times
1688 // an object is visited.
1689 timesAnyObjectVisited++;
1691 // If this is the first visit to c, initialize its retainer set.
1692 maybeInitRetainerSet(c);
1693 retainerSetOfc = retainerSetOf(c);
1696 // isRetainer(cp) == rtsTrue => s == NULL
1697 // isRetainer(cp) == rtsFalse => s == cp.retainer
1701 s = retainerSetOf(cp);
1703 // (c, cp, r, s) is available.
1705 // (c, cp, r, s, R_r) is available, so compute the retainer set for *c.
1706 if (retainerSetOfc == NULL) {
1707 // This is the first visit to *c.
1711 associate(c, singleton(r));
1713 // s is actually the retainer set of *c!
1716 // compute c_child_r
1717 c_child_r = isRetainer(c) ? getRetainerFrom(c) : r;
1719 // This is not the first visit to *c.
1720 if (isMember(r, retainerSetOfc))
1721 goto loop; // no need to process child
1724 associate(c, addElement(r, retainerSetOfc));
1726 // s is not NULL and cp is not a retainer. This means that
1727 // each time *cp is visited, so is *c. Thus, if s has
1728 // exactly one more element in its retainer set than c, s
1729 // is also the new retainer set for *c.
1730 if (s->num == retainerSetOfc->num + 1) {
1733 // Otherwise, just add R_r to the current retainer set of *c.
1735 associate(c, addElement(r, retainerSetOfc));
1740 goto loop; // no need to process child
1742 // compute c_child_r
1746 // now, RSET() of all of *c, *cp, and *r is valid.
1747 // (c, c_child_r) are available.
1751 // Special case closures: we process these all in one go rather
1752 // than attempting to save the current position, because doing so
1756 retainStack(c, c_child_r,
1758 ((StgTSO *)c)->stack + ((StgTSO *)c)->stack_size);
1763 StgPAP *pap = (StgPAP *)c;
1764 retain_PAP_payload(c, c_child_r, pap->fun, pap->payload, pap->n_args);
1770 StgAP *ap = (StgAP *)c;
1771 retain_PAP_payload(c, c_child_r, ap->fun, ap->payload, ap->n_args);
1776 retainClosure(((StgAP_STACK *)c)->fun, c, c_child_r);
1777 retainStack(c, c_child_r,
1778 (StgPtr)((StgAP_STACK *)c)->payload,
1779 (StgPtr)((StgAP_STACK *)c)->payload +
1780 ((StgAP_STACK *)c)->size);
1784 push(c, c_child_r, &first_child);
1786 // If first_child is null, c has no child.
1787 // If first_child is not null, the top stack element points to the next
1788 // object. push() may or may not push a stackElement on the stack.
1789 if (first_child == NULL)
1792 // (c, cp, r) = (first_child, c, c_child_r)
1799 /* -----------------------------------------------------------------------------
1800 * Compute the retainer set for every object reachable from *tl.
1801 * -------------------------------------------------------------------------- */
1803 retainRoot( StgClosure **tl )
1807 // We no longer assume that only TSOs and WEAKs are roots; any closure can
1810 ASSERT(isEmptyRetainerStack());
1811 currentStackBoundary = stackTop;
1813 c = UNTAG_CLOSURE(*tl);
1814 if (c != &stg_END_TSO_QUEUE_closure && isRetainer(c)) {
1815 retainClosure(c, c, getRetainerFrom(c));
1817 retainClosure(c, c, CCS_SYSTEM);
1820 // NOT TRUE: ASSERT(isMember(getRetainerFrom(*tl), retainerSetOf(*tl)));
1821 // *tl might be a TSO which is ThreadComplete, in which
1822 // case we ignore it for the purposes of retainer profiling.
1825 /* -----------------------------------------------------------------------------
1826 * Compute the retainer set for each of the objects in the heap.
1827 * -------------------------------------------------------------------------- */
1829 computeRetainerSet( void )
1836 #ifdef DEBUG_RETAINER
1837 RetainerSet tmpRetainerSet;
1840 GetRoots(retainRoot); // for scheduler roots
1842 // This function is called after a major GC, when key, value, and finalizer
1843 // all are guaranteed to be valid, or reachable.
1845 // The following code assumes that WEAK objects are considered to be roots
1846 // for retainer profilng.
1847 for (weak = weak_ptr_list; weak != NULL; weak = weak->link)
1848 // retainRoot((StgClosure *)weak);
1849 retainRoot((StgClosure **)&weak);
1851 // Consider roots from the stable ptr table.
1852 markStablePtrTable(retainRoot);
1854 // The following code resets the rs field of each unvisited mutable
1855 // object (computing sumOfNewCostExtra and updating costArray[] when
1856 // debugging retainer profiler).
1857 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1858 // NOT TRUE: even G0 has a block on its mutable list
1859 // ASSERT(g != 0 || (generations[g].mut_list == NULL));
1861 // Traversing through mut_list is necessary
1862 // because we can find MUT_VAR objects which have not been
1863 // visited during retainer profiling.
1864 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
1865 for (ml = bd->start; ml < bd->free; ml++) {
1867 maybeInitRetainerSet((StgClosure *)*ml);
1868 rtl = retainerSetOf((StgClosure *)*ml);
1870 #ifdef DEBUG_RETAINER
1872 // first visit to *ml
1873 // This is a violation of the interface rule!
1874 RSET(ml) = (RetainerSet *)((StgWord)(&tmpRetainerSet) | flip);
1876 switch (get_itbl((StgClosure *)ml)->type) {
1880 case CONSTR_NOCAF_STATIC:
1884 barf("Invalid object in computeRetainerSet(): %d", get_itbl((StgClosure*)ml)->type);
1888 costArray[get_itbl((StgClosure *)ml)->type] += cost((StgClosure *)ml);
1889 sumOfNewCostExtra += cost((StgClosure *)ml);
1899 /* -----------------------------------------------------------------------------
1900 * Traverse all static objects for which we compute retainer sets,
1901 * and reset their rs fields to NULL, which is accomplished by
1902 * invoking maybeInitRetainerSet(). This function must be called
1903 * before zeroing all objects reachable from scavenged_static_objects
1904 * in the case of major gabage collections. See GarbageCollect() in
1907 * The mut_once_list of the oldest generation must also be traversed?
1908 * Why? Because if the evacuation of an object pointed to by a static
1909 * indirection object fails, it is put back to the mut_once_list of
1910 * the oldest generation.
1911 * However, this is not necessary because any static indirection objects
1912 * are just traversed through to reach dynamic objects. In other words,
1913 * they are not taken into consideration in computing retainer sets.
1914 * -------------------------------------------------------------------------- */
1916 resetStaticObjectForRetainerProfiling( void )
1918 #ifdef DEBUG_RETAINER
1923 #ifdef DEBUG_RETAINER
1926 p = scavenged_static_objects;
1927 while (p != END_OF_STATIC_LIST) {
1928 #ifdef DEBUG_RETAINER
1931 switch (get_itbl(p)->type) {
1933 // Since we do not compute the retainer set of any
1934 // IND_STATIC object, we don't have to reset its retainer
1936 p = (StgClosure*)*IND_STATIC_LINK(p);
1939 maybeInitRetainerSet(p);
1940 p = (StgClosure*)*THUNK_STATIC_LINK(p);
1943 maybeInitRetainerSet(p);
1944 p = (StgClosure*)*FUN_STATIC_LINK(p);
1947 maybeInitRetainerSet(p);
1948 p = (StgClosure*)*STATIC_LINK(get_itbl(p), p);
1951 barf("resetStaticObjectForRetainerProfiling: %p (%s)",
1952 p, get_itbl(p)->type);
1956 #ifdef DEBUG_RETAINER
1957 // debugBelch("count in scavenged_static_objects = %d\n", count);
1961 /* -----------------------------------------------------------------------------
1962 * Perform retainer profiling.
1963 * N is the oldest generation being profilied, where the generations are
1964 * numbered starting at 0.
1967 * This function should be called only immediately after major garbage
1969 * ------------------------------------------------------------------------- */
1971 retainerProfile(void)
1973 #ifdef DEBUG_RETAINER
1975 nat totalHeapSize; // total raw heap size (computed by linear scanning)
1978 #ifdef DEBUG_RETAINER
1979 debugBelch(" < retainerProfile() invoked : %d>\n", retainerGeneration);
1984 // We haven't flipped the bit yet.
1985 #ifdef DEBUG_RETAINER
1986 debugBelch("Before traversing:\n");
1987 sumOfCostLinear = 0;
1988 for (i = 0;i < N_CLOSURE_TYPES; i++)
1989 costArrayLinear[i] = 0;
1990 totalHeapSize = checkHeapSanityForRetainerProfiling();
1992 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
1994 debugBelch("costArrayLinear[] = ");
1995 for (i = 0;i < N_CLOSURE_TYPES; i++)
1996 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
2000 ASSERT(sumOfCostLinear == totalHeapSize);
2003 #define pcostArrayLinear(index) \
2004 if (costArrayLinear[index] > 0) \
2005 debugBelch("costArrayLinear[" #index "] = %u\n", costArrayLinear[index])
2006 pcostArrayLinear(THUNK_STATIC);
2007 pcostArrayLinear(FUN_STATIC);
2008 pcostArrayLinear(CONSTR_STATIC);
2009 pcostArrayLinear(CONSTR_NOCAF_STATIC);
2013 // Now we flips flip.
2016 #ifdef DEBUG_RETAINER
2022 numObjectVisited = 0;
2023 timesAnyObjectVisited = 0;
2025 #ifdef DEBUG_RETAINER
2026 debugBelch("During traversing:\n");
2028 sumOfNewCostExtra = 0;
2029 for (i = 0;i < N_CLOSURE_TYPES; i++)
2034 We initialize the traverse stack each time the retainer profiling is
2035 performed (because the traverse stack size varies on each retainer profiling
2036 and this operation is not costly anyhow). However, we just refresh the
2039 initializeTraverseStack();
2040 #ifdef DEBUG_RETAINER
2041 initializeAllRetainerSet();
2043 refreshAllRetainerSet();
2045 computeRetainerSet();
2047 #ifdef DEBUG_RETAINER
2048 debugBelch("After traversing:\n");
2049 sumOfCostLinear = 0;
2050 for (i = 0;i < N_CLOSURE_TYPES; i++)
2051 costArrayLinear[i] = 0;
2052 totalHeapSize = checkHeapSanityForRetainerProfiling();
2054 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
2055 ASSERT(sumOfCostLinear == totalHeapSize);
2057 // now, compare the two results
2060 costArray[] must be exactly the same as costArrayLinear[].
2062 1) Dead weak pointers, whose type is CONSTR. These objects are not
2063 reachable from any roots.
2065 debugBelch("Comparison:\n");
2066 debugBelch("\tcostArrayLinear[] (must be empty) = ");
2067 for (i = 0;i < N_CLOSURE_TYPES; i++)
2068 if (costArray[i] != costArrayLinear[i])
2069 // nothing should be printed except MUT_VAR after major GCs
2070 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
2073 debugBelch("\tsumOfNewCost = %u\n", sumOfNewCost);
2074 debugBelch("\tsumOfNewCostExtra = %u\n", sumOfNewCostExtra);
2075 debugBelch("\tcostArray[] (must be empty) = ");
2076 for (i = 0;i < N_CLOSURE_TYPES; i++)
2077 if (costArray[i] != costArrayLinear[i])
2078 // nothing should be printed except MUT_VAR after major GCs
2079 debugBelch("[%u:%u] ", i, costArray[i]);
2082 // only for major garbage collection
2083 ASSERT(sumOfNewCost + sumOfNewCostExtra == sumOfCostLinear);
2087 closeTraverseStack();
2088 #ifdef DEBUG_RETAINER
2089 closeAllRetainerSet();
2091 // Note that there is no post-processing for the retainer sets.
2093 retainerGeneration++;
2096 retainerGeneration - 1, // retainerGeneration has just been incremented!
2097 #ifdef DEBUG_RETAINER
2098 maxCStackSize, maxStackSize,
2100 (double)timesAnyObjectVisited / numObjectVisited);
2103 /* -----------------------------------------------------------------------------
2105 * -------------------------------------------------------------------------- */
2107 #ifdef DEBUG_RETAINER
2109 #define LOOKS_LIKE_PTR(r) ((LOOKS_LIKE_STATIC_CLOSURE(r) || \
2110 ((HEAP_ALLOCED(r) && ((Bdescr((P_)r)->flags & BF_FREE) == 0)))) && \
2111 ((StgWord)(*(StgPtr)r)!=0xaaaaaaaa))
2114 sanityCheckHeapClosure( StgClosure *c )
2118 ASSERT(LOOKS_LIKE_GHC_INFO(c->header.info));
2119 ASSERT(!closure_STATIC(c));
2120 ASSERT(LOOKS_LIKE_PTR(c));
2122 if ((((StgWord)RSET(c) & 1) ^ flip) != 0) {
2123 if (get_itbl(c)->type == CONSTR &&
2124 !strcmp(GET_PROF_TYPE(get_itbl(c)), "DEAD_WEAK") &&
2125 !strcmp(GET_PROF_DESC(get_itbl(c)), "DEAD_WEAK")) {
2126 debugBelch("\tUnvisited dead weak pointer object found: c = %p\n", c);
2127 costArray[get_itbl(c)->type] += cost(c);
2128 sumOfNewCost += cost(c);
2131 "Unvisited object: flip = %d, c = %p(%d, %s, %s), rs = %p\n",
2132 flip, c, get_itbl(c)->type,
2133 get_itbl(c)->prof.closure_type, GET_PROF_DESC(get_itbl(c)),
2136 // debugBelch("sanityCheckHeapClosure) S: flip = %d, c = %p(%d), rs = %p\n", flip, c, get_itbl(c)->type, RSET(c));
2139 return closure_sizeW(c);
2143 heapCheck( bdescr *bd )
2146 static nat costSum, size;
2149 while (bd != NULL) {
2151 while (p < bd->free) {
2152 size = sanityCheckHeapClosure((StgClosure *)p);
2153 sumOfCostLinear += size;
2154 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2156 // no need for slop check; I think slops are not used currently.
2158 ASSERT(p == bd->free);
2159 costSum += bd->free - bd->start;
2167 smallObjectPoolCheck(void)
2171 static nat costSum, size;
2181 while (p < alloc_Hp) {
2182 size = sanityCheckHeapClosure((StgClosure *)p);
2183 sumOfCostLinear += size;
2184 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2187 ASSERT(p == alloc_Hp);
2188 costSum += alloc_Hp - bd->start;
2191 while (bd != NULL) {
2193 while (p < bd->free) {
2194 size = sanityCheckHeapClosure((StgClosure *)p);
2195 sumOfCostLinear += size;
2196 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2199 ASSERT(p == bd->free);
2200 costSum += bd->free - bd->start;
2208 chainCheck(bdescr *bd)
2213 while (bd != NULL) {
2214 // bd->free - bd->start is not an accurate measurement of the
2215 // object size. Actually it is always zero, so we compute its
2217 size = sanityCheckHeapClosure((StgClosure *)bd->start);
2218 sumOfCostLinear += size;
2219 costArrayLinear[get_itbl((StgClosure *)bd->start)->type] += size;
2228 checkHeapSanityForRetainerProfiling( void )
2233 debugBelch("START: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2234 if (RtsFlags.GcFlags.generations == 1) {
2235 costSum += heapCheck(g0s0->to_blocks);
2236 debugBelch("heapCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2237 costSum += chainCheck(g0s0->large_objects);
2238 debugBelch("chainCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2240 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
2241 for (s = 0; s < generations[g].n_steps; s++) {
2243 After all live objects have been scavenged, the garbage
2244 collector may create some objects in
2245 scheduleFinalizers(). These objects are created throught
2246 allocate(), so the small object pool or the large object
2247 pool of the g0s0 may not be empty.
2249 if (g == 0 && s == 0) {
2250 costSum += smallObjectPoolCheck();
2251 debugBelch("smallObjectPoolCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2252 costSum += chainCheck(generations[g].steps[s].large_objects);
2253 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2255 costSum += heapCheck(generations[g].steps[s].blocks);
2256 debugBelch("heapCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2257 costSum += chainCheck(generations[g].steps[s].large_objects);
2258 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2267 findPointer(StgPtr p)
2273 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2274 for (s = 0; s < generations[g].n_steps; s++) {
2275 // if (g == 0 && s == 0) continue;
2276 bd = generations[g].steps[s].blocks;
2277 for (; bd; bd = bd->link) {
2278 for (q = bd->start; q < bd->free; q++) {
2279 if (*q == (StgWord)p) {
2281 while (!LOOKS_LIKE_GHC_INFO(*r)) r--;
2282 debugBelch("Found in gen[%d], step[%d]: q = %p, r = %p\n", g, s, q, r);
2287 bd = generations[g].steps[s].large_objects;
2288 for (; bd; bd = bd->link) {
2289 e = bd->start + cost((StgClosure *)bd->start);
2290 for (q = bd->start; q < e; q++) {
2291 if (*q == (StgWord)p) {
2293 while (*r == 0 || !LOOKS_LIKE_GHC_INFO(*r)) r--;
2294 debugBelch("Found in gen[%d], large_objects: %p\n", g, r);
2304 belongToHeap(StgPtr p)
2309 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2310 for (s = 0; s < generations[g].n_steps; s++) {
2311 // if (g == 0 && s == 0) continue;
2312 bd = generations[g].steps[s].blocks;
2313 for (; bd; bd = bd->link) {
2314 if (bd->start <= p && p < bd->free) {
2315 debugBelch("Belongs to gen[%d], step[%d]", g, s);
2319 bd = generations[g].steps[s].large_objects;
2320 for (; bd; bd = bd->link) {
2321 if (bd->start <= p && p < bd->start + getHeapClosureSize((StgClosure *)bd->start)) {
2322 debugBelch("Found in gen[%d], large_objects: %p\n", g, bd->start);
2329 #endif /* DEBUG_RETAINER */
2331 #endif /* PROFILING */