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"
29 #include "Profiling.h"
31 #include "BlockAlloc.h"
36 Note: what to change in order to plug-in a new retainer profiling scheme?
37 (1) type retainer in ../includes/StgRetainerProf.h
38 (2) retainer function R(), i.e., getRetainerFrom()
39 (3) the two hashing functions, hashKeySingleton() and hashKeyAddElement(),
40 in RetainerSet.h, if needed.
41 (4) printRetainer() and printRetainerSetShort() in RetainerSet.c.
44 /* -----------------------------------------------------------------------------
46 * -------------------------------------------------------------------------- */
48 static nat retainerGeneration; // generation
50 static nat numObjectVisited; // total number of objects visited
51 static nat timesAnyObjectVisited; // number of times any objects are visited
54 The rs field in the profile header of any object points to its retainer
55 set in an indirect way: if flip is 0, it points to the retainer set;
56 if flip is 1, it points to the next byte after the retainer set (even
57 for NULL pointers). Therefore, with flip 1, (rs ^ 1) is the actual
58 pointer. See retainerSetOf().
61 StgWord flip = 0; // flip bit
62 // must be 0 if DEBUG_RETAINER is on (for static closures)
64 #define setRetainerSetToNull(c) \
65 (c)->header.prof.hp.rs = (RetainerSet *)((StgWord)NULL | flip)
67 static void retainStack(StgClosure *, retainer, StgPtr, StgPtr);
68 static void retainClosure(StgClosure *, StgClosure *, retainer);
70 static void belongToHeap(StgPtr p);
75 cStackSize records how many times retainStack() has been invoked recursively,
76 that is, the number of activation records for retainStack() on the C stack.
77 maxCStackSize records its max value.
79 cStackSize <= maxCStackSize
81 static nat cStackSize, maxCStackSize;
83 static nat sumOfNewCost; // sum of the cost of each object, computed
84 // when the object is first visited
85 static nat sumOfNewCostExtra; // for those objects not visited during
86 // retainer profiling, e.g., MUT_VAR
87 static nat costArray[N_CLOSURE_TYPES];
89 nat sumOfCostLinear; // sum of the costs of all object, computed
90 // when linearly traversing the heap after
92 nat costArrayLinear[N_CLOSURE_TYPES];
95 /* -----------------------------------------------------------------------------
96 * Retainer stack - header
98 * Although the retainer stack implementation could be separated *
99 * from the retainer profiling engine, there does not seem to be
100 * any advantage in doing that; retainer stack is an integral part
101 * of retainer profiling engine and cannot be use elsewhere at
103 * -------------------------------------------------------------------------- */
113 // fixed layout or layout specified by a field in the closure
118 // See StgClosureInfo in InfoTables.h
119 #if SIZEOF_VOID_P == 8
156 firstStack points to the first block group.
157 currentStack points to the block group currently being used.
158 currentStack->free == stackLimit.
159 stackTop points to the topmost byte in the stack of currentStack.
160 Unless the whole stack is empty, stackTop must point to the topmost
161 object (or byte) in the whole stack. Thus, it is only when the whole stack
162 is empty that stackTop == stackLimit (not during the execution of push()
164 stackBottom == currentStack->start.
165 stackLimit == currentStack->start + BLOCK_SIZE_W * currentStack->blocks.
167 When a current stack becomes empty, stackTop is set to point to
168 the topmost element on the previous block group so as to satisfy
169 the invariants described above.
171 static bdescr *firstStack = NULL;
172 static bdescr *currentStack;
173 static stackElement *stackBottom, *stackTop, *stackLimit;
176 currentStackBoundary is used to mark the current stack chunk.
177 If stackTop == currentStackBoundary, it means that the current stack chunk
178 is empty. It is the responsibility of the user to keep currentStackBoundary
179 valid all the time if it is to be employed.
181 static stackElement *currentStackBoundary;
184 stackSize records the current size of the stack.
185 maxStackSize records its high water mark.
187 stackSize <= maxStackSize
189 stackSize is just an estimate measure of the depth of the graph. The reason
190 is that some heap objects have only a single child and may not result
191 in a new element being pushed onto the stack. Therefore, at the end of
192 retainer profiling, maxStackSize + maxCStackSize is some value no greater
193 than the actual depth of the graph.
195 #ifdef DEBUG_RETAINER
196 static int stackSize, maxStackSize;
199 // number of blocks allocated for one stack
200 #define BLOCKS_IN_STACK 1
202 /* -----------------------------------------------------------------------------
203 * Add a new block group to the stack.
205 * currentStack->link == s.
206 * -------------------------------------------------------------------------- */
208 newStackBlock( bdescr *bd )
211 stackTop = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
212 stackBottom = (stackElement *)bd->start;
213 stackLimit = (stackElement *)stackTop;
214 bd->free = (StgPtr)stackLimit;
217 /* -----------------------------------------------------------------------------
218 * Return to the previous block group.
220 * s->link == currentStack.
221 * -------------------------------------------------------------------------- */
223 returnToOldStack( bdescr *bd )
226 stackTop = (stackElement *)bd->free;
227 stackBottom = (stackElement *)bd->start;
228 stackLimit = (stackElement *)(bd->start + BLOCK_SIZE_W * bd->blocks);
229 bd->free = (StgPtr)stackLimit;
232 /* -----------------------------------------------------------------------------
233 * Initializes the traverse stack.
234 * -------------------------------------------------------------------------- */
236 initializeTraverseStack( void )
238 if (firstStack != NULL) {
239 freeChain(firstStack);
242 firstStack = allocGroup(BLOCKS_IN_STACK);
243 firstStack->link = NULL;
244 firstStack->u.back = NULL;
246 newStackBlock(firstStack);
249 /* -----------------------------------------------------------------------------
250 * Frees all the block groups in the traverse stack.
253 * -------------------------------------------------------------------------- */
255 closeTraverseStack( void )
257 freeChain(firstStack);
261 /* -----------------------------------------------------------------------------
262 * Returns rtsTrue if the whole stack is empty.
263 * -------------------------------------------------------------------------- */
264 static INLINE rtsBool
265 isEmptyRetainerStack( void )
267 return (firstStack == currentStack) && stackTop == stackLimit;
270 /* -----------------------------------------------------------------------------
271 * Returns size of stack
272 * -------------------------------------------------------------------------- */
275 retainerStackBlocks( void )
280 for (bd = firstStack; bd != NULL; bd = bd->link)
287 /* -----------------------------------------------------------------------------
288 * Returns rtsTrue if stackTop is at the stack boundary of the current stack,
289 * i.e., if the current stack chunk is empty.
290 * -------------------------------------------------------------------------- */
291 static INLINE rtsBool
294 return stackTop == currentStackBoundary;
297 /* -----------------------------------------------------------------------------
298 * Initializes *info from ptrs and payload.
300 * payload[] begins with ptrs pointers followed by non-pointers.
301 * -------------------------------------------------------------------------- */
303 init_ptrs( stackPos *info, nat ptrs, StgPtr payload )
305 info->type = posTypePtrs;
306 info->next.ptrs.pos = 0;
307 info->next.ptrs.ptrs = ptrs;
308 info->next.ptrs.payload = payload;
311 /* -----------------------------------------------------------------------------
312 * Find the next object from *info.
313 * -------------------------------------------------------------------------- */
314 static INLINE StgClosure *
315 find_ptrs( stackPos *info )
317 if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
318 return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
324 /* -----------------------------------------------------------------------------
325 * Initializes *info from SRT information stored in *infoTable.
326 * -------------------------------------------------------------------------- */
328 init_srt_fun( stackPos *info, StgFunInfoTable *infoTable )
330 if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
331 info->type = posTypeLargeSRT;
332 info->next.large_srt.srt = (StgLargeSRT *)GET_FUN_SRT(infoTable);
333 info->next.large_srt.offset = 0;
335 info->type = posTypeSRT;
336 info->next.srt.srt = (StgClosure **)GET_FUN_SRT(infoTable);
337 info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
342 init_srt_thunk( stackPos *info, StgThunkInfoTable *infoTable )
344 if (infoTable->i.srt_bitmap == (StgHalfWord)(-1)) {
345 info->type = posTypeLargeSRT;
346 info->next.large_srt.srt = (StgLargeSRT *)GET_SRT(infoTable);
347 info->next.large_srt.offset = 0;
349 info->type = posTypeSRT;
350 info->next.srt.srt = (StgClosure **)GET_SRT(infoTable);
351 info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
355 /* -----------------------------------------------------------------------------
356 * Find the next object from *info.
357 * -------------------------------------------------------------------------- */
358 static INLINE StgClosure *
359 find_srt( stackPos *info )
364 if (info->type == posTypeSRT) {
366 bitmap = info->next.srt.srt_bitmap;
367 while (bitmap != 0) {
368 if ((bitmap & 1) != 0) {
369 #ifdef ENABLE_WIN32_DLL_SUPPORT
371 if ((unsigned long)(*(info->next.srt.srt)) & 0x1)
372 c = (* (StgClosure **)((unsigned long)*(info->next.srt.srt)) & ~0x1);
374 c = *(info->next.srt.srt);
376 c = *(info->next.srt.srt);
378 bitmap = bitmap >> 1;
379 info->next.srt.srt++;
380 info->next.srt.srt_bitmap = bitmap;
383 bitmap = bitmap >> 1;
384 info->next.srt.srt++;
386 // bitmap is now zero...
391 nat i = info->next.large_srt.offset;
394 // Follow the pattern from GC.c:scavenge_large_srt_bitmap().
395 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
396 bitmap = bitmap >> (i % BITS_IN(StgWord));
397 while (i < info->next.large_srt.srt->l.size) {
398 if ((bitmap & 1) != 0) {
399 c = ((StgClosure **)info->next.large_srt.srt->srt)[i];
401 info->next.large_srt.offset = i;
405 if (i % BITS_IN(W_) == 0) {
406 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
408 bitmap = bitmap >> 1;
411 // reached the end of this bitmap.
412 info->next.large_srt.offset = i;
417 /* -----------------------------------------------------------------------------
418 * push() pushes a stackElement representing the next child of *c
419 * onto the traverse stack. If *c has no child, *first_child is set
420 * to NULL and nothing is pushed onto the stack. If *c has only one
421 * child, *c_chlid is set to that child and nothing is pushed onto
422 * the stack. If *c has more than two children, *first_child is set
423 * to the first child and a stackElement representing the second
424 * child is pushed onto the stack.
427 * *c_child_r is the most recent retainer of *c's children.
428 * *c is not any of TSO, AP, PAP, AP_STACK, which means that
429 * there cannot be any stack objects.
430 * Note: SRTs are considered to be children as well.
431 * -------------------------------------------------------------------------- */
433 push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
436 bdescr *nbd; // Next Block Descriptor
438 #ifdef DEBUG_RETAINER
439 // debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
442 ASSERT(get_itbl(c)->type != TSO);
443 ASSERT(get_itbl(c)->type != AP_STACK);
450 se.c_child_r = c_child_r;
453 switch (get_itbl(c)->type) {
460 case SE_CAF_BLACKHOLE:
465 // one child (fixed), no SRT
467 *first_child = ((StgMutVar *)c)->var;
470 *first_child = ((StgSelector *)c)->selectee;
473 case IND_OLDGEN_PERM:
475 *first_child = ((StgInd *)c)->indirectee;
479 *first_child = c->payload[0];
482 // For CONSTR_2_0 and MVAR, we use se.info.step to record the position
483 // of the next child. We do not write a separate initialization code.
484 // Also we do not have to initialize info.type;
486 // two children (fixed), no SRT
487 // need to push a stackElement, but nothing to store in se.info
489 *first_child = c->payload[0]; // return the first pointer
490 // se.info.type = posTypeStep;
491 // se.info.next.step = 2; // 2 = second
494 // three children (fixed), no SRT
495 // need to push a stackElement
497 // head must be TSO and the head of a linked list of TSOs.
498 // Shoule it be a child? Seems to be yes.
499 *first_child = (StgClosure *)((StgMVar *)c)->head;
500 // se.info.type = posTypeStep;
501 se.info.next.step = 2; // 2 = second
504 // three children (fixed), no SRT
506 *first_child = ((StgWeak *)c)->key;
507 // se.info.type = posTypeStep;
508 se.info.next.step = 2;
511 // layout.payload.ptrs, no SRT
517 init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs,
519 *first_child = find_ptrs(&se.info);
520 if (*first_child == NULL)
524 // StgMutArrPtr.ptrs, no SRT
526 case MUT_ARR_PTRS_FROZEN:
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, (StgPtr)c->payload);
547 *first_child = find_ptrs(&se.info);
548 if (*first_child == NULL)
549 // no child from ptrs, so check SRT
553 // 1 fixed child, SRT
556 *first_child = c->payload[0];
557 ASSERT(*first_child != NULL);
558 init_srt_fun(&se.info, get_fun_itbl(c));
563 *first_child = c->payload[0];
564 ASSERT(*first_child != NULL);
565 init_srt_thunk(&se.info, get_thunk_itbl(c));
568 case FUN_STATIC: // *c is a heap object.
569 ASSERT(get_itbl(c)->srt_bitmap != 0);
573 init_srt_fun(&se.info, get_fun_itbl(c));
574 *first_child = find_srt(&se.info);
575 if (*first_child == NULL)
581 ASSERT(get_itbl(c)->srt_bitmap != 0);
585 init_srt_thunk(&se.info, get_thunk_itbl(c));
586 *first_child = find_srt(&se.info);
587 if (*first_child == NULL)
598 case CONSTR_CHARLIKE:
599 case CONSTR_NOCAF_STATIC:
620 barf("Invalid object *c in push()");
624 if (stackTop - 1 < stackBottom) {
625 #ifdef DEBUG_RETAINER
626 // debugBelch("push() to the next stack.\n");
628 // currentStack->free is updated when the active stack is switched
629 // to the next stack.
630 currentStack->free = (StgPtr)stackTop;
632 if (currentStack->link == NULL) {
633 nbd = allocGroup(BLOCKS_IN_STACK);
635 nbd->u.back = currentStack;
636 currentStack->link = nbd;
638 nbd = currentStack->link;
643 // adjust stackTop (acutal push)
645 // If the size of stackElement was huge, we would better replace the
646 // following statement by either a memcpy() call or a switch statement
647 // on the type of the element. Currently, the size of stackElement is
648 // small enough (5 words) that this direct assignment seems to be enough.
651 #ifdef DEBUG_RETAINER
653 if (stackSize > maxStackSize) maxStackSize = stackSize;
654 // ASSERT(stackSize >= 0);
655 // debugBelch("stackSize = %d\n", stackSize);
659 /* -----------------------------------------------------------------------------
660 * popOff() and popOffReal(): Pop a stackElement off the traverse stack.
662 * stackTop cannot be equal to stackLimit unless the whole stack is
663 * empty, in which case popOff() is not allowed.
665 * You can think of popOffReal() as a part of popOff() which is
666 * executed at the end of popOff() in necessary. Since popOff() is
667 * likely to be executed quite often while popOffReal() is not, we
668 * separate popOffReal() from popOff(), which is declared as an
669 * INLINE function (for the sake of execution speed). popOffReal()
670 * is called only within popOff() and nowhere else.
671 * -------------------------------------------------------------------------- */
675 bdescr *pbd; // Previous Block Descriptor
677 #ifdef DEBUG_RETAINER
678 // debugBelch("pop() to the previous stack.\n");
681 ASSERT(stackTop + 1 == stackLimit);
682 ASSERT(stackBottom == (stackElement *)currentStack->start);
684 if (firstStack == currentStack) {
685 // The stack is completely empty.
687 ASSERT(stackTop == stackLimit);
688 #ifdef DEBUG_RETAINER
690 if (stackSize > maxStackSize) maxStackSize = stackSize;
692 ASSERT(stackSize >= 0);
693 debugBelch("stackSize = %d\n", stackSize);
699 // currentStack->free is updated when the active stack is switched back
700 // to the previous stack.
701 currentStack->free = (StgPtr)stackLimit;
703 // find the previous block descriptor
704 pbd = currentStack->u.back;
707 returnToOldStack(pbd);
709 #ifdef DEBUG_RETAINER
711 if (stackSize > maxStackSize) maxStackSize = stackSize;
713 ASSERT(stackSize >= 0);
714 debugBelch("stackSize = %d\n", stackSize);
721 #ifdef DEBUG_RETAINER
722 // debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
725 ASSERT(stackTop != stackLimit);
726 ASSERT(!isEmptyRetainerStack());
728 // <= (instead of <) is wrong!
729 if (stackTop + 1 < stackLimit) {
731 #ifdef DEBUG_RETAINER
733 if (stackSize > maxStackSize) maxStackSize = stackSize;
735 ASSERT(stackSize >= 0);
736 debugBelch("stackSize = %d\n", stackSize);
745 /* -----------------------------------------------------------------------------
746 * Finds the next object to be considered for retainer profiling and store
748 * Test if the topmost stack element indicates that more objects are left,
749 * and if so, retrieve the first object and store its pointer to *c. Also,
750 * set *cp and *r appropriately, both of which are stored in the stack element.
751 * The topmost stack element then is overwritten so as for it to now denote
753 * If the topmost stack element indicates no more objects are left, pop
754 * off the stack element until either an object can be retrieved or
755 * the current stack chunk becomes empty, indicated by rtsTrue returned by
756 * isOnBoundary(), in which case *c is set to NULL.
758 * It is okay to call this function even when the current stack chunk
760 * -------------------------------------------------------------------------- */
762 pop( StgClosure **c, StgClosure **cp, retainer *r )
766 #ifdef DEBUG_RETAINER
767 // debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
771 if (isOnBoundary()) { // if the current stack chunk is depleted
778 switch (get_itbl(se->c)->type) {
779 // two children (fixed), no SRT
780 // nothing in se.info
782 *c = se->c->payload[1];
788 // three children (fixed), no SRT
789 // need to push a stackElement
791 if (se->info.next.step == 2) {
792 *c = (StgClosure *)((StgMVar *)se->c)->tail;
793 se->info.next.step++; // move to the next step
796 *c = ((StgMVar *)se->c)->value;
803 // three children (fixed), no SRT
805 if (se->info.next.step == 2) {
806 *c = ((StgWeak *)se->c)->value;
807 se->info.next.step++;
810 *c = ((StgWeak *)se->c)->finalizer;
822 // StgMutArrPtr.ptrs, no SRT
824 case MUT_ARR_PTRS_FROZEN:
825 *c = find_ptrs(&se->info);
834 // layout.payload.ptrs, SRT
835 case FUN: // always a heap object
837 if (se->info.type == posTypePtrs) {
838 *c = find_ptrs(&se->info);
844 init_srt_fun(&se->info, get_fun_itbl(se->c));
850 if (se->info.type == posTypePtrs) {
851 *c = find_ptrs(&se->info);
857 init_srt_thunk(&se->info, get_thunk_itbl(se->c));
873 *c = find_srt(&se->info);
882 // no child (fixed), no SRT
888 case SE_CAF_BLACKHOLE:
890 // one child (fixed), no SRT
894 case IND_OLDGEN_PERM:
904 case CONSTR_CHARLIKE:
905 case CONSTR_NOCAF_STATIC:
926 barf("Invalid object *c in pop()");
932 /* -----------------------------------------------------------------------------
933 * RETAINER PROFILING ENGINE
934 * -------------------------------------------------------------------------- */
937 initRetainerProfiling( void )
939 initializeAllRetainerSet();
940 retainerGeneration = 0;
943 /* -----------------------------------------------------------------------------
944 * This function must be called before f-closing prof_file.
945 * -------------------------------------------------------------------------- */
947 endRetainerProfiling( void )
949 #ifdef SECOND_APPROACH
950 outputAllRetainerSet(prof_file);
954 /* -----------------------------------------------------------------------------
955 * Returns the actual pointer to the retainer set of the closure *c.
956 * It may adjust RSET(c) subject to flip.
958 * RSET(c) is initialized to NULL if its current value does not
961 * Even though this function has side effects, they CAN be ignored because
962 * subsequent calls to retainerSetOf() always result in the same return value
963 * and retainerSetOf() is the only way to retrieve retainerSet of a given
965 * We have to perform an XOR (^) operation each time a closure is examined.
966 * The reason is that we do not know when a closure is visited last.
967 * -------------------------------------------------------------------------- */
969 maybeInitRetainerSet( StgClosure *c )
971 if (!isRetainerSetFieldValid(c)) {
972 setRetainerSetToNull(c);
976 /* -----------------------------------------------------------------------------
977 * Returns rtsTrue if *c is a retainer.
978 * -------------------------------------------------------------------------- */
979 static INLINE rtsBool
980 isRetainer( StgClosure *c )
982 switch (get_itbl(c)->type) {
986 // TSOs MUST be retainers: they constitute the set of roots.
993 case MUT_ARR_PTRS_FROZEN:
995 // thunks are retainers.
1002 case THUNK_SELECTOR:
1006 // Static thunks, or CAFS, are obviously retainers.
1009 // WEAK objects are roots; there is separate code in which traversing
1010 // begins from WEAK objects.
1032 // partial applications
1038 case SE_CAF_BLACKHOLE:
1041 case IND_OLDGEN_PERM:
1056 // IND_STATIC cannot be *c, *cp, *r in the retainer profiling loop.
1058 // CONSTR_INTLIKE, CONSTR_CHARLIKE, and CONSTR_NOCAF_STATIC
1059 // cannot be *c, *cp, *r in the retainer profiling loop.
1060 case CONSTR_INTLIKE:
1061 case CONSTR_CHARLIKE:
1062 case CONSTR_NOCAF_STATIC:
1063 // Stack objects are invalid because they are never treated as
1064 // legal objects during retainer profiling.
1082 case INVALID_OBJECT:
1084 barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
1089 /* -----------------------------------------------------------------------------
1090 * Returns the retainer function value for the closure *c, i.e., R(*c).
1091 * This function does NOT return the retainer(s) of *c.
1093 * *c must be a retainer.
1095 * Depending on the definition of this function, the maintenance of retainer
1096 * sets can be made easier. If most retainer sets are likely to be created
1097 * again across garbage collections, refreshAllRetainerSet() in
1098 * RetainerSet.c can simply do nothing.
1099 * If this is not the case, we can free all the retainer sets and
1100 * re-initialize the hash table.
1101 * See refreshAllRetainerSet() in RetainerSet.c.
1102 * -------------------------------------------------------------------------- */
1103 static INLINE retainer
1104 getRetainerFrom( StgClosure *c )
1106 ASSERT(isRetainer(c));
1108 #if defined(RETAINER_SCHEME_INFO)
1109 // Retainer scheme 1: retainer = info table
1111 #elif defined(RETAINER_SCHEME_CCS)
1112 // Retainer scheme 2: retainer = cost centre stack
1113 return c->header.prof.ccs;
1114 #elif defined(RETAINER_SCHEME_CC)
1115 // Retainer scheme 3: retainer = cost centre
1116 return c->header.prof.ccs->cc;
1120 /* -----------------------------------------------------------------------------
1121 * Associates the retainer set *s with the closure *c, that is, *s becomes
1122 * the retainer set of *c.
1126 * -------------------------------------------------------------------------- */
1128 associate( StgClosure *c, RetainerSet *s )
1130 // StgWord has the same size as pointers, so the following type
1132 RSET(c) = (RetainerSet *)((StgWord)s | flip);
1135 /* -----------------------------------------------------------------------------
1136 Call retainClosure for each of the closures covered by a large bitmap.
1137 -------------------------------------------------------------------------- */
1140 retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, nat size,
1141 StgClosure *c, retainer c_child_r)
1147 bitmap = large_bitmap->bitmap[b];
1148 for (i = 0; i < size; ) {
1149 if ((bitmap & 1) == 0) {
1150 retainClosure((StgClosure *)*p, c, c_child_r);
1154 if (i % BITS_IN(W_) == 0) {
1156 bitmap = large_bitmap->bitmap[b];
1158 bitmap = bitmap >> 1;
1163 static INLINE StgPtr
1164 retain_small_bitmap (StgPtr p, nat size, StgWord bitmap,
1165 StgClosure *c, retainer c_child_r)
1168 if ((bitmap & 1) == 0) {
1169 retainClosure((StgClosure *)*p, c, c_child_r);
1172 bitmap = bitmap >> 1;
1178 /* -----------------------------------------------------------------------------
1179 * Call retainClosure for each of the closures in an SRT.
1180 * ------------------------------------------------------------------------- */
1183 retain_large_srt_bitmap (StgLargeSRT *srt, StgClosure *c, retainer c_child_r)
1190 p = (StgClosure **)srt->srt;
1192 bitmap = srt->l.bitmap[b];
1193 for (i = 0; i < size; ) {
1194 if ((bitmap & 1) != 0) {
1195 retainClosure((StgClosure *)*p, c, c_child_r);
1199 if (i % BITS_IN(W_) == 0) {
1201 bitmap = srt->l.bitmap[b];
1203 bitmap = bitmap >> 1;
1209 retainSRT (StgClosure **srt, nat srt_bitmap, StgClosure *c, retainer c_child_r)
1214 bitmap = srt_bitmap;
1217 if (bitmap == (StgHalfWord)(-1)) {
1218 retain_large_srt_bitmap( (StgLargeSRT *)srt, c, c_child_r );
1222 while (bitmap != 0) {
1223 if ((bitmap & 1) != 0) {
1224 #ifdef ENABLE_WIN32_DLL_SUPPORT
1225 if ( (unsigned long)(*srt) & 0x1 ) {
1226 retainClosure(*stgCast(StgClosure**,(stgCast(unsigned long, *srt) & ~0x1)),
1229 retainClosure(*srt,c,c_child_r);
1232 retainClosure(*srt,c,c_child_r);
1236 bitmap = bitmap >> 1;
1240 /* -----------------------------------------------------------------------------
1241 * Process all the objects in the stack chunk from stackStart to stackEnd
1242 * with *c and *c_child_r being their parent and their most recent retainer,
1243 * respectively. Treat stackOptionalFun as another child of *c if it is
1246 * *c is one of the following: TSO, AP_STACK.
1247 * If *c is TSO, c == c_child_r.
1248 * stackStart < stackEnd.
1249 * RSET(c) and RSET(c_child_r) are valid, i.e., their
1250 * interpretation conforms to the current value of flip (even when they
1251 * are interpreted to be NULL).
1252 * If *c is TSO, its state is not any of ThreadRelocated, ThreadComplete,
1253 * or ThreadKilled, which means that its stack is ready to process.
1255 * This code was almost plagiarzied from GC.c! For each pointer,
1256 * retainClosure() is invoked instead of evacuate().
1257 * -------------------------------------------------------------------------- */
1259 retainStack( StgClosure *c, retainer c_child_r,
1260 StgPtr stackStart, StgPtr stackEnd )
1262 stackElement *oldStackBoundary;
1264 StgRetInfoTable *info;
1268 #ifdef DEBUG_RETAINER
1270 if (cStackSize > maxCStackSize) maxCStackSize = cStackSize;
1274 Each invocation of retainStack() creates a new virtual
1275 stack. Since all such stacks share a single common stack, we
1276 record the current currentStackBoundary, which will be restored
1279 oldStackBoundary = currentStackBoundary;
1280 currentStackBoundary = stackTop;
1282 #ifdef DEBUG_RETAINER
1283 // debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n", oldStackBoundary, currentStackBoundary);
1286 ASSERT(get_itbl(c)->type != TSO ||
1287 (((StgTSO *)c)->what_next != ThreadRelocated &&
1288 ((StgTSO *)c)->what_next != ThreadComplete &&
1289 ((StgTSO *)c)->what_next != ThreadKilled));
1292 while (p < stackEnd) {
1293 info = get_ret_itbl((StgClosure *)p);
1295 switch(info->i.type) {
1298 retainClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
1299 p += sizeofW(StgUpdateFrame);
1306 bitmap = BITMAP_BITS(info->i.layout.bitmap);
1307 size = BITMAP_SIZE(info->i.layout.bitmap);
1309 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1312 retainSRT((StgClosure **)GET_SRT(info), info->i.srt_bitmap, c, c_child_r);
1319 retainClosure((StgClosure *)*p, c, c_child_r);
1322 size = BCO_BITMAP_SIZE(bco);
1323 retain_large_bitmap(p, BCO_BITMAP(bco), size, c, c_child_r);
1328 // large bitmap (> 32 entries, or > 64 on a 64-bit machine)
1331 size = GET_LARGE_BITMAP(&info->i)->size;
1333 retain_large_bitmap(p, GET_LARGE_BITMAP(&info->i),
1334 size, c, c_child_r);
1336 // and don't forget to follow the SRT
1339 // Dynamic bitmap: the mask is stored on the stack
1342 dyn = ((StgRetDyn *)p)->liveness;
1344 // traverse the bitmap first
1345 bitmap = RET_DYN_LIVENESS(dyn);
1346 p = (P_)&((StgRetDyn *)p)->payload[0];
1347 size = RET_DYN_BITMAP_SIZE;
1348 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1350 // skip over the non-ptr words
1351 p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
1353 // follow the ptr words
1354 for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
1355 retainClosure((StgClosure *)*p, c, c_child_r);
1362 StgRetFun *ret_fun = (StgRetFun *)p;
1363 StgFunInfoTable *fun_info;
1365 retainClosure(ret_fun->fun, c, c_child_r);
1366 fun_info = get_fun_itbl(ret_fun->fun);
1368 p = (P_)&ret_fun->payload;
1369 switch (fun_info->f.fun_type) {
1371 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1372 size = BITMAP_SIZE(fun_info->f.b.bitmap);
1373 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1376 size = GET_FUN_LARGE_BITMAP(fun_info)->size;
1377 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1378 size, c, c_child_r);
1382 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1383 size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
1384 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1391 barf("Invalid object found in retainStack(): %d",
1392 (int)(info->i.type));
1396 // restore currentStackBoundary
1397 currentStackBoundary = oldStackBoundary;
1398 #ifdef DEBUG_RETAINER
1399 // debugBelch("retainStack() finished: currentStackBoundary = 0x%x\n", currentStackBoundary);
1402 #ifdef DEBUG_RETAINER
1407 /* ----------------------------------------------------------------------------
1408 * Call retainClosure for each of the children of a PAP/AP
1409 * ------------------------------------------------------------------------- */
1411 static INLINE StgPtr
1412 retain_PAP (StgPAP *pap, retainer c_child_r)
1415 StgWord bitmap, size;
1416 StgFunInfoTable *fun_info;
1418 retainClosure(pap->fun, (StgClosure *)pap, c_child_r);
1419 fun_info = get_fun_itbl(pap->fun);
1420 ASSERT(fun_info->i.type != PAP);
1422 p = (StgPtr)pap->payload;
1425 switch (fun_info->f.fun_type) {
1427 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1428 p = retain_small_bitmap(p, pap->n_args, bitmap,
1429 (StgClosure *)pap, c_child_r);
1432 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1433 size, (StgClosure *)pap, c_child_r);
1437 retain_large_bitmap((StgPtr)pap->payload, BCO_BITMAP(pap->fun),
1438 size, (StgClosure *)pap, c_child_r);
1442 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1443 p = retain_small_bitmap(p, pap->n_args, bitmap,
1444 (StgClosure *)pap, c_child_r);
1450 /* -----------------------------------------------------------------------------
1451 * Compute the retainer set of *c0 and all its desecents by traversing.
1452 * *cp0 is the parent of *c0, and *r0 is the most recent retainer of *c0.
1454 * c0 = cp0 = r0 holds only for root objects.
1455 * RSET(cp0) and RSET(r0) are valid, i.e., their
1456 * interpretation conforms to the current value of flip (even when they
1457 * are interpreted to be NULL).
1458 * However, RSET(c0) may be corrupt, i.e., it may not conform to
1459 * the current value of flip. If it does not, during the execution
1460 * of this function, RSET(c0) must be initialized as well as all
1463 * stackTop must be the same at the beginning and the exit of this function.
1464 * *c0 can be TSO (as well as AP_STACK).
1465 * -------------------------------------------------------------------------- */
1467 retainClosure( StgClosure *c0, StgClosure *cp0, retainer r0 )
1469 // c = Current closure
1470 // cp = Current closure's Parent
1471 // r = current closures' most recent Retainer
1472 // c_child_r = current closure's children's most recent retainer
1473 // first_child = first child of c
1474 StgClosure *c, *cp, *first_child;
1475 RetainerSet *s, *retainerSetOfc;
1476 retainer r, c_child_r;
1479 #ifdef DEBUG_RETAINER
1480 // StgPtr oldStackTop;
1483 #ifdef DEBUG_RETAINER
1484 // oldStackTop = stackTop;
1485 // debugBelch("retainClosure() called: c0 = 0x%x, cp0 = 0x%x, r0 = 0x%x\n", c0, cp0, r0);
1488 // (c, cp, r) = (c0, cp0, r0)
1495 //debugBelch("loop");
1496 // pop to (c, cp, r);
1500 #ifdef DEBUG_RETAINER
1501 // debugBelch("retainClosure() ends: oldStackTop = 0x%x, stackTop = 0x%x\n", oldStackTop, stackTop);
1506 //debugBelch("inner_loop");
1509 // c = current closure under consideration,
1510 // cp = current closure's parent,
1511 // r = current closure's most recent retainer
1513 // Loop invariants (on the meaning of c, cp, r, and their retainer sets):
1514 // RSET(cp) and RSET(r) are valid.
1515 // RSET(c) is valid only if c has been visited before.
1517 // Loop invariants (on the relation between c, cp, and r)
1518 // if cp is not a retainer, r belongs to RSET(cp).
1519 // if cp is a retainer, r == cp.
1521 typeOfc = get_itbl(c)->type;
1523 #ifdef DEBUG_RETAINER
1526 case CONSTR_INTLIKE:
1527 case CONSTR_CHARLIKE:
1528 case CONSTR_NOCAF_STATIC:
1534 if (retainerSetOf(c) == NULL) { // first visit?
1535 costArray[typeOfc] += cost(c);
1536 sumOfNewCost += cost(c);
1545 if (((StgTSO *)c)->what_next == ThreadComplete ||
1546 ((StgTSO *)c)->what_next == ThreadKilled) {
1547 #ifdef DEBUG_RETAINER
1548 debugBelch("ThreadComplete or ThreadKilled encountered in retainClosure()\n");
1552 if (((StgTSO *)c)->what_next == ThreadRelocated) {
1553 #ifdef DEBUG_RETAINER
1554 debugBelch("ThreadRelocated encountered in retainClosure()\n");
1556 c = (StgClosure *)((StgTSO *)c)->link;
1562 // We just skip IND_STATIC, so its retainer set is never computed.
1563 c = ((StgIndStatic *)c)->indirectee;
1565 case CONSTR_INTLIKE:
1566 case CONSTR_CHARLIKE:
1567 // static objects with no pointers out, so goto loop.
1568 case CONSTR_NOCAF_STATIC:
1569 // It is not just enough not to compute the retainer set for *c; it is
1570 // mandatory because CONSTR_NOCAF_STATIC are not reachable from
1571 // scavenged_static_objects, the list from which is assumed to traverse
1572 // all static objects after major garbage collections.
1576 if (get_itbl(c)->srt_bitmap == 0) {
1577 // No need to compute the retainer set; no dynamic objects
1578 // are reachable from *c.
1580 // Static objects: if we traverse all the live closures,
1581 // including static closures, during each heap census then
1582 // we will observe that some static closures appear and
1583 // disappear. eg. a closure may contain a pointer to a
1584 // static function 'f' which is not otherwise reachable
1585 // (it doesn't indirectly point to any CAFs, so it doesn't
1586 // appear in any SRTs), so we would find 'f' during
1587 // traversal. However on the next sweep there may be no
1588 // closures pointing to 'f'.
1590 // We must therefore ignore static closures whose SRT is
1591 // empty, because these are exactly the closures that may
1592 // "appear". A closure with a non-empty SRT, and which is
1593 // still required, will always be reachable.
1595 // But what about CONSTR_STATIC? Surely these may be able
1596 // to appear, and they don't have SRTs, so we can't
1597 // check. So for now, we're calling
1598 // resetStaticObjectForRetainerProfiling() from the
1599 // garbage collector to reset the retainer sets in all the
1600 // reachable static objects.
1607 // The above objects are ignored in computing the average number of times
1608 // an object is visited.
1609 timesAnyObjectVisited++;
1611 // If this is the first visit to c, initialize its retainer set.
1612 maybeInitRetainerSet(c);
1613 retainerSetOfc = retainerSetOf(c);
1616 // isRetainer(cp) == rtsTrue => s == NULL
1617 // isRetainer(cp) == rtsFalse => s == cp.retainer
1621 s = retainerSetOf(cp);
1623 // (c, cp, r, s) is available.
1625 // (c, cp, r, s, R_r) is available, so compute the retainer set for *c.
1626 if (retainerSetOfc == NULL) {
1627 // This is the first visit to *c.
1631 associate(c, singleton(r));
1633 // s is actually the retainer set of *c!
1636 // compute c_child_r
1637 c_child_r = isRetainer(c) ? getRetainerFrom(c) : r;
1639 // This is not the first visit to *c.
1640 if (isMember(r, retainerSetOfc))
1641 goto loop; // no need to process child
1644 associate(c, addElement(r, retainerSetOfc));
1646 // s is not NULL and cp is not a retainer. This means that
1647 // each time *cp is visited, so is *c. Thus, if s has
1648 // exactly one more element in its retainer set than c, s
1649 // is also the new retainer set for *c.
1650 if (s->num == retainerSetOfc->num + 1) {
1653 // Otherwise, just add R_r to the current retainer set of *c.
1655 associate(c, addElement(r, retainerSetOfc));
1660 goto loop; // no need to process child
1662 // compute c_child_r
1666 // now, RSET() of all of *c, *cp, and *r is valid.
1667 // (c, c_child_r) are available.
1671 // Special case closures: we process these all in one go rather
1672 // than attempting to save the current position, because doing so
1676 retainStack(c, c_child_r,
1678 ((StgTSO *)c)->stack + ((StgTSO *)c)->stack_size);
1683 retain_PAP((StgPAP *)c, c_child_r);
1687 retainClosure(((StgAP_STACK *)c)->fun, c, c_child_r);
1688 retainStack(c, c_child_r,
1689 (StgPtr)((StgAP_STACK *)c)->payload,
1690 (StgPtr)((StgAP_STACK *)c)->payload +
1691 ((StgAP_STACK *)c)->size);
1695 push(c, c_child_r, &first_child);
1697 // If first_child is null, c has no child.
1698 // If first_child is not null, the top stack element points to the next
1699 // object. push() may or may not push a stackElement on the stack.
1700 if (first_child == NULL)
1703 // (c, cp, r) = (first_child, c, c_child_r)
1710 /* -----------------------------------------------------------------------------
1711 * Compute the retainer set for every object reachable from *tl.
1712 * -------------------------------------------------------------------------- */
1714 retainRoot( StgClosure **tl )
1716 // We no longer assume that only TSOs and WEAKs are roots; any closure can
1719 ASSERT(isEmptyRetainerStack());
1720 currentStackBoundary = stackTop;
1722 if (isRetainer(*tl)) {
1723 retainClosure(*tl, *tl, getRetainerFrom(*tl));
1725 retainClosure(*tl, *tl, CCS_SYSTEM);
1728 // NOT TRUE: ASSERT(isMember(getRetainerFrom(*tl), retainerSetOf(*tl)));
1729 // *tl might be a TSO which is ThreadComplete, in which
1730 // case we ignore it for the purposes of retainer profiling.
1733 /* -----------------------------------------------------------------------------
1734 * Compute the retainer set for each of the objects in the heap.
1735 * -------------------------------------------------------------------------- */
1737 computeRetainerSet( void )
1744 #ifdef DEBUG_RETAINER
1745 RetainerSet tmpRetainerSet;
1748 GetRoots(retainRoot); // for scheduler roots
1750 // This function is called after a major GC, when key, value, and finalizer
1751 // all are guaranteed to be valid, or reachable.
1753 // The following code assumes that WEAK objects are considered to be roots
1754 // for retainer profilng.
1755 for (weak = weak_ptr_list; weak != NULL; weak = weak->link)
1756 // retainRoot((StgClosure *)weak);
1757 retainRoot((StgClosure **)&weak);
1759 // Consider roots from the stable ptr table.
1760 markStablePtrTable(retainRoot);
1762 // The following code resets the rs field of each unvisited mutable
1763 // object (computing sumOfNewCostExtra and updating costArray[] when
1764 // debugging retainer profiler).
1765 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1766 ASSERT(g != 0 || (generations[g].mut_list == NULL));
1768 // Traversing through mut_list is necessary
1769 // because we can find MUT_VAR objects which have not been
1770 // visited during retainer profiling.
1771 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
1772 for (ml = bd->start; ml < bd->free; ml++) {
1774 maybeInitRetainerSet((StgClosure *)ml);
1775 rtl = retainerSetOf((StgClosure *)ml);
1777 #ifdef DEBUG_RETAINER
1779 // first visit to *ml
1780 // This is a violation of the interface rule!
1781 RSET(ml) = (RetainerSet *)((StgWord)(&tmpRetainerSet) | flip);
1783 switch (get_itbl((StgClosure *)ml)->type) {
1787 case CONSTR_INTLIKE:
1788 case CONSTR_CHARLIKE:
1789 case CONSTR_NOCAF_STATIC:
1793 barf("Invalid object in computeRetainerSet(): %d", get_itbl((StgClosure*)ml)->type);
1797 costArray[get_itbl((StgClosure *)ml)->type] += cost((StgClosure *)ml);
1798 sumOfNewCostExtra += cost((StgClosure *)ml);
1808 /* -----------------------------------------------------------------------------
1809 * Traverse all static objects for which we compute retainer sets,
1810 * and reset their rs fields to NULL, which is accomplished by
1811 * invoking maybeInitRetainerSet(). This function must be called
1812 * before zeroing all objects reachable from scavenged_static_objects
1813 * in the case of major gabage collections. See GarbageCollect() in
1816 * The mut_once_list of the oldest generation must also be traversed?
1817 * Why? Because if the evacuation of an object pointed to by a static
1818 * indirection object fails, it is put back to the mut_once_list of
1819 * the oldest generation.
1820 * However, this is not necessary because any static indirection objects
1821 * are just traversed through to reach dynamic objects. In other words,
1822 * they are not taken into consideration in computing retainer sets.
1823 * -------------------------------------------------------------------------- */
1825 resetStaticObjectForRetainerProfiling( void )
1827 #ifdef DEBUG_RETAINER
1832 #ifdef DEBUG_RETAINER
1835 p = scavenged_static_objects;
1836 while (p != END_OF_STATIC_LIST) {
1837 #ifdef DEBUG_RETAINER
1840 switch (get_itbl(p)->type) {
1842 // Since we do not compute the retainer set of any
1843 // IND_STATIC object, we don't have to reset its retainer
1845 p = IND_STATIC_LINK(p);
1848 maybeInitRetainerSet(p);
1849 p = THUNK_STATIC_LINK(p);
1852 maybeInitRetainerSet(p);
1853 p = FUN_STATIC_LINK(p);
1856 maybeInitRetainerSet(p);
1857 p = STATIC_LINK(get_itbl(p), p);
1860 barf("resetStaticObjectForRetainerProfiling: %p (%s)",
1861 p, get_itbl(p)->type);
1865 #ifdef DEBUG_RETAINER
1866 // debugBelch("count in scavenged_static_objects = %d\n", count);
1870 /* -----------------------------------------------------------------------------
1871 * Perform retainer profiling.
1872 * N is the oldest generation being profilied, where the generations are
1873 * numbered starting at 0.
1876 * This function should be called only immediately after major garbage
1878 * ------------------------------------------------------------------------- */
1880 retainerProfile(void)
1882 #ifdef DEBUG_RETAINER
1884 nat totalHeapSize; // total raw heap size (computed by linear scanning)
1887 #ifdef DEBUG_RETAINER
1888 debugBelch(" < retainerProfile() invoked : %d>\n", retainerGeneration);
1893 // We haven't flipped the bit yet.
1894 #ifdef DEBUG_RETAINER
1895 debugBelch("Before traversing:\n");
1896 sumOfCostLinear = 0;
1897 for (i = 0;i < N_CLOSURE_TYPES; i++)
1898 costArrayLinear[i] = 0;
1899 totalHeapSize = checkHeapSanityForRetainerProfiling();
1901 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
1903 debugBelch("costArrayLinear[] = ");
1904 for (i = 0;i < N_CLOSURE_TYPES; i++)
1905 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
1909 ASSERT(sumOfCostLinear == totalHeapSize);
1912 #define pcostArrayLinear(index) \
1913 if (costArrayLinear[index] > 0) \
1914 debugBelch("costArrayLinear[" #index "] = %u\n", costArrayLinear[index])
1915 pcostArrayLinear(THUNK_STATIC);
1916 pcostArrayLinear(FUN_STATIC);
1917 pcostArrayLinear(CONSTR_STATIC);
1918 pcostArrayLinear(CONSTR_NOCAF_STATIC);
1919 pcostArrayLinear(CONSTR_INTLIKE);
1920 pcostArrayLinear(CONSTR_CHARLIKE);
1924 // Now we flips flip.
1927 #ifdef DEBUG_RETAINER
1933 numObjectVisited = 0;
1934 timesAnyObjectVisited = 0;
1936 #ifdef DEBUG_RETAINER
1937 debugBelch("During traversing:\n");
1939 sumOfNewCostExtra = 0;
1940 for (i = 0;i < N_CLOSURE_TYPES; i++)
1945 We initialize the traverse stack each time the retainer profiling is
1946 performed (because the traverse stack size varies on each retainer profiling
1947 and this operation is not costly anyhow). However, we just refresh the
1950 initializeTraverseStack();
1951 #ifdef DEBUG_RETAINER
1952 initializeAllRetainerSet();
1954 refreshAllRetainerSet();
1956 computeRetainerSet();
1958 #ifdef DEBUG_RETAINER
1959 debugBelch("After traversing:\n");
1960 sumOfCostLinear = 0;
1961 for (i = 0;i < N_CLOSURE_TYPES; i++)
1962 costArrayLinear[i] = 0;
1963 totalHeapSize = checkHeapSanityForRetainerProfiling();
1965 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
1966 ASSERT(sumOfCostLinear == totalHeapSize);
1968 // now, compare the two results
1971 costArray[] must be exactly the same as costArrayLinear[].
1973 1) Dead weak pointers, whose type is CONSTR. These objects are not
1974 reachable from any roots.
1976 debugBelch("Comparison:\n");
1977 debugBelch("\tcostArrayLinear[] (must be empty) = ");
1978 for (i = 0;i < N_CLOSURE_TYPES; i++)
1979 if (costArray[i] != costArrayLinear[i])
1980 // nothing should be printed except MUT_VAR after major GCs
1981 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
1984 debugBelch("\tsumOfNewCost = %u\n", sumOfNewCost);
1985 debugBelch("\tsumOfNewCostExtra = %u\n", sumOfNewCostExtra);
1986 debugBelch("\tcostArray[] (must be empty) = ");
1987 for (i = 0;i < N_CLOSURE_TYPES; i++)
1988 if (costArray[i] != costArrayLinear[i])
1989 // nothing should be printed except MUT_VAR after major GCs
1990 debugBelch("[%u:%u] ", i, costArray[i]);
1993 // only for major garbage collection
1994 ASSERT(sumOfNewCost + sumOfNewCostExtra == sumOfCostLinear);
1998 closeTraverseStack();
1999 #ifdef DEBUG_RETAINER
2000 closeAllRetainerSet();
2002 // Note that there is no post-processing for the retainer sets.
2004 retainerGeneration++;
2007 retainerGeneration - 1, // retainerGeneration has just been incremented!
2008 #ifdef DEBUG_RETAINER
2009 maxCStackSize, maxStackSize,
2011 (double)timesAnyObjectVisited / numObjectVisited);
2014 /* -----------------------------------------------------------------------------
2016 * -------------------------------------------------------------------------- */
2018 #ifdef DEBUG_RETAINER
2020 #define LOOKS_LIKE_PTR(r) ((LOOKS_LIKE_STATIC_CLOSURE(r) || \
2021 ((HEAP_ALLOCED(r) && Bdescr((P_)r)->free != (void *)-1))) && \
2022 ((StgWord)(*(StgPtr)r)!=0xaaaaaaaa))
2025 sanityCheckHeapClosure( StgClosure *c )
2029 ASSERT(LOOKS_LIKE_GHC_INFO(c->header.info));
2030 ASSERT(!closure_STATIC(c));
2031 ASSERT(LOOKS_LIKE_PTR(c));
2033 if ((((StgWord)RSET(c) & 1) ^ flip) != 0) {
2034 if (get_itbl(c)->type == CONSTR &&
2035 !strcmp(get_itbl(c)->prof.closure_type, "DEAD_WEAK") &&
2036 !strcmp(get_itbl(c)->prof.closure_desc, "DEAD_WEAK")) {
2037 debugBelch("\tUnvisited dead weak pointer object found: c = %p\n", c);
2038 costArray[get_itbl(c)->type] += cost(c);
2039 sumOfNewCost += cost(c);
2042 "Unvisited object: flip = %d, c = %p(%d, %s, %s), rs = %p\n",
2043 flip, c, get_itbl(c)->type,
2044 get_itbl(c)->prof.closure_type, get_itbl(c)->prof.closure_desc,
2047 // debugBelch("sanityCheckHeapClosure) S: flip = %d, c = %p(%d), rs = %p\n", flip, c, get_itbl(c)->type, RSET(c));
2051 switch (info->type) {
2053 return tso_sizeW((StgTSO *)c);
2061 return stg_max(sizeW_fromITBL(info), sizeofW(StgHeader) + MIN_UPD_SIZE);
2064 return sizeofW(StgMVar);
2067 case MUT_ARR_PTRS_FROZEN:
2068 return mut_arr_ptrs_sizeW((StgMutArrPtrs *)c);
2072 return pap_sizeW((StgPAP *)c);
2075 return ap_stack_sizeW((StgAP_STACK *)c);
2078 return arr_words_sizeW((StgArrWords *)c);
2097 case SE_CAF_BLACKHOLE:
2100 case IND_OLDGEN_PERM:
2104 return sizeW_fromITBL(info);
2106 case THUNK_SELECTOR:
2107 return sizeofW(StgHeader) + MIN_UPD_SIZE;
2116 case CONSTR_INTLIKE:
2117 case CONSTR_CHARLIKE:
2118 case CONSTR_NOCAF_STATIC:
2135 case INVALID_OBJECT:
2137 barf("Invalid object in sanityCheckHeapClosure(): %d",
2144 heapCheck( bdescr *bd )
2147 static nat costSum, size;
2150 while (bd != NULL) {
2152 while (p < bd->free) {
2153 size = sanityCheckHeapClosure((StgClosure *)p);
2154 sumOfCostLinear += size;
2155 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2157 // no need for slop check; I think slops are not used currently.
2159 ASSERT(p == bd->free);
2160 costSum += bd->free - bd->start;
2168 smallObjectPoolCheck(void)
2172 static nat costSum, size;
2174 bd = small_alloc_list;
2182 while (p < alloc_Hp) {
2183 size = sanityCheckHeapClosure((StgClosure *)p);
2184 sumOfCostLinear += size;
2185 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2188 ASSERT(p == alloc_Hp);
2189 costSum += alloc_Hp - bd->start;
2192 while (bd != NULL) {
2194 while (p < bd->free) {
2195 size = sanityCheckHeapClosure((StgClosure *)p);
2196 sumOfCostLinear += size;
2197 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2200 ASSERT(p == bd->free);
2201 costSum += bd->free - bd->start;
2209 chainCheck(bdescr *bd)
2214 while (bd != NULL) {
2215 // bd->free - bd->start is not an accurate measurement of the
2216 // object size. Actually it is always zero, so we compute its
2218 size = sanityCheckHeapClosure((StgClosure *)bd->start);
2219 sumOfCostLinear += size;
2220 costArrayLinear[get_itbl((StgClosure *)bd->start)->type] += size;
2229 checkHeapSanityForRetainerProfiling( void )
2234 debugBelch("START: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2235 if (RtsFlags.GcFlags.generations == 1) {
2236 costSum += heapCheck(g0s0->to_blocks);
2237 debugBelch("heapCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2238 costSum += chainCheck(g0s0->large_objects);
2239 debugBelch("chainCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2241 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
2242 for (s = 0; s < generations[g].n_steps; s++) {
2244 After all live objects have been scavenged, the garbage
2245 collector may create some objects in
2246 scheduleFinalizers(). These objects are created throught
2247 allocate(), so the small object pool or the large object
2248 pool of the g0s0 may not be empty.
2250 if (g == 0 && s == 0) {
2251 costSum += smallObjectPoolCheck();
2252 debugBelch("smallObjectPoolCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2253 costSum += chainCheck(generations[g].steps[s].large_objects);
2254 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2256 costSum += heapCheck(generations[g].steps[s].blocks);
2257 debugBelch("heapCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2258 costSum += chainCheck(generations[g].steps[s].large_objects);
2259 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2268 findPointer(StgPtr p)
2274 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2275 for (s = 0; s < generations[g].n_steps; s++) {
2276 // if (g == 0 && s == 0) continue;
2277 bd = generations[g].steps[s].blocks;
2278 for (; bd; bd = bd->link) {
2279 for (q = bd->start; q < bd->free; q++) {
2280 if (*q == (StgWord)p) {
2282 while (!LOOKS_LIKE_GHC_INFO(*r)) r--;
2283 debugBelch("Found in gen[%d], step[%d]: q = %p, r = %p\n", g, s, q, r);
2288 bd = generations[g].steps[s].large_objects;
2289 for (; bd; bd = bd->link) {
2290 e = bd->start + cost((StgClosure *)bd->start);
2291 for (q = bd->start; q < e; q++) {
2292 if (*q == (StgWord)p) {
2294 while (*r == 0 || !LOOKS_LIKE_GHC_INFO(*r)) r--;
2295 debugBelch("Found in gen[%d], large_objects: %p\n", g, r);
2305 belongToHeap(StgPtr p)
2310 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2311 for (s = 0; s < generations[g].n_steps; s++) {
2312 // if (g == 0 && s == 0) continue;
2313 bd = generations[g].steps[s].blocks;
2314 for (; bd; bd = bd->link) {
2315 if (bd->start <= p && p < bd->free) {
2316 debugBelch("Belongs to gen[%d], step[%d]", g, s);
2320 bd = generations[g].steps[s].large_objects;
2321 for (; bd; bd = bd->link) {
2322 if (bd->start <= p && p < bd->start + getHeapClosureSize((StgClosure *)bd->start)) {
2323 debugBelch("Found in gen[%d], large_objects: %p\n", g, bd->start);
2330 #endif /* DEBUG_RETAINER */
2332 #endif /* PROFILING */