Remove the IND_OLDGEN and IND_OLDGEN_PERM closure types
[ghc.git] / rts / RetainerProfile.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 2001
4 * Author: Sungwoo Park
5 *
6 * Retainer profiling.
7 *
8 * ---------------------------------------------------------------------------*/
9
10 #ifdef PROFILING
11
12 // Turn off inlining when debugging - it obfuscates things
13 #ifdef DEBUG
14 #define INLINE
15 #else
16 #define INLINE inline
17 #endif
18
19 #include "PosixSource.h"
20 #include "Rts.h"
21
22 #include "RtsUtils.h"
23 #include "RetainerProfile.h"
24 #include "RetainerSet.h"
25 #include "Schedule.h"
26 #include "Printer.h"
27 #include "Weak.h"
28 #include "sm/Sanity.h"
29 #include "Profiling.h"
30 #include "Stats.h"
31 #include "ProfHeap.h"
32 #include "Apply.h"
33 #include "sm/Storage.h" // for END_OF_STATIC_LIST
34
35 /*
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.
42 */
43
44 /* -----------------------------------------------------------------------------
45 * Declarations...
46 * -------------------------------------------------------------------------- */
47
48 static nat retainerGeneration; // generation
49
50 static nat numObjectVisited; // total number of objects visited
51 static nat timesAnyObjectVisited; // number of times any objects are visited
52
53 /*
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().
59 */
60
61 StgWord flip = 0; // flip bit
62 // must be 0 if DEBUG_RETAINER is on (for static closures)
63
64 #define setRetainerSetToNull(c) \
65 (c)->header.prof.hp.rs = (RetainerSet *)((StgWord)NULL | flip)
66
67 static void retainStack(StgClosure *, retainer, StgPtr, StgPtr);
68 static void retainClosure(StgClosure *, StgClosure *, retainer);
69 #ifdef DEBUG_RETAINER
70 static void belongToHeap(StgPtr p);
71 #endif
72
73 #ifdef DEBUG_RETAINER
74 /*
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.
78 Invariants:
79 cStackSize <= maxCStackSize
80 */
81 static nat cStackSize, maxCStackSize;
82
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];
88
89 nat sumOfCostLinear; // sum of the costs of all object, computed
90 // when linearly traversing the heap after
91 // retainer profiling
92 nat costArrayLinear[N_CLOSURE_TYPES];
93 #endif
94
95 /* -----------------------------------------------------------------------------
96 * Retainer stack - header
97 * Note:
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
102 * all.
103 * -------------------------------------------------------------------------- */
104
105 typedef enum {
106 posTypeStep,
107 posTypePtrs,
108 posTypeSRT,
109 posTypeLargeSRT,
110 } nextPosType;
111
112 typedef union {
113 // fixed layout or layout specified by a field in the closure
114 StgWord step;
115
116 // layout.payload
117 struct {
118 // See StgClosureInfo in InfoTables.h
119 #if SIZEOF_VOID_P == 8
120 StgWord32 pos;
121 StgWord32 ptrs;
122 #else
123 StgWord16 pos;
124 StgWord16 ptrs;
125 #endif
126 StgPtr payload;
127 } ptrs;
128
129 // SRT
130 struct {
131 StgClosure **srt;
132 StgWord srt_bitmap;
133 } srt;
134
135 // Large SRT
136 struct {
137 StgLargeSRT *srt;
138 StgWord offset;
139 } large_srt;
140
141 } nextPos;
142
143 typedef struct {
144 nextPosType type;
145 nextPos next;
146 } stackPos;
147
148 typedef struct {
149 StgClosure *c;
150 retainer c_child_r;
151 stackPos info;
152 } stackElement;
153
154 /*
155 Invariants:
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()
163 and pop()).
164 stackBottom == currentStack->start.
165 stackLimit == currentStack->start + BLOCK_SIZE_W * currentStack->blocks.
166 Note:
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.
170 */
171 static bdescr *firstStack = NULL;
172 static bdescr *currentStack;
173 static stackElement *stackBottom, *stackTop, *stackLimit;
174
175 /*
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.
180 */
181 static stackElement *currentStackBoundary;
182
183 /*
184 stackSize records the current size of the stack.
185 maxStackSize records its high water mark.
186 Invariants:
187 stackSize <= maxStackSize
188 Note:
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.
194 */
195 #ifdef DEBUG_RETAINER
196 static int stackSize, maxStackSize;
197 #endif
198
199 // number of blocks allocated for one stack
200 #define BLOCKS_IN_STACK 1
201
202 /* -----------------------------------------------------------------------------
203 * Add a new block group to the stack.
204 * Invariants:
205 * currentStack->link == s.
206 * -------------------------------------------------------------------------- */
207 static INLINE void
208 newStackBlock( bdescr *bd )
209 {
210 currentStack = 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;
215 }
216
217 /* -----------------------------------------------------------------------------
218 * Return to the previous block group.
219 * Invariants:
220 * s->link == currentStack.
221 * -------------------------------------------------------------------------- */
222 static INLINE void
223 returnToOldStack( bdescr *bd )
224 {
225 currentStack = 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;
230 }
231
232 /* -----------------------------------------------------------------------------
233 * Initializes the traverse stack.
234 * -------------------------------------------------------------------------- */
235 static void
236 initializeTraverseStack( void )
237 {
238 if (firstStack != NULL) {
239 freeChain(firstStack);
240 }
241
242 firstStack = allocGroup(BLOCKS_IN_STACK);
243 firstStack->link = NULL;
244 firstStack->u.back = NULL;
245
246 newStackBlock(firstStack);
247 }
248
249 /* -----------------------------------------------------------------------------
250 * Frees all the block groups in the traverse stack.
251 * Invariants:
252 * firstStack != NULL
253 * -------------------------------------------------------------------------- */
254 static void
255 closeTraverseStack( void )
256 {
257 freeChain(firstStack);
258 firstStack = NULL;
259 }
260
261 /* -----------------------------------------------------------------------------
262 * Returns rtsTrue if the whole stack is empty.
263 * -------------------------------------------------------------------------- */
264 static INLINE rtsBool
265 isEmptyRetainerStack( void )
266 {
267 return (firstStack == currentStack) && stackTop == stackLimit;
268 }
269
270 /* -----------------------------------------------------------------------------
271 * Returns size of stack
272 * -------------------------------------------------------------------------- */
273 #ifdef DEBUG
274 lnat
275 retainerStackBlocks( void )
276 {
277 bdescr* bd;
278 lnat res = 0;
279
280 for (bd = firstStack; bd != NULL; bd = bd->link)
281 res += bd->blocks;
282
283 return res;
284 }
285 #endif
286
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
292 isOnBoundary( void )
293 {
294 return stackTop == currentStackBoundary;
295 }
296
297 /* -----------------------------------------------------------------------------
298 * Initializes *info from ptrs and payload.
299 * Invariants:
300 * payload[] begins with ptrs pointers followed by non-pointers.
301 * -------------------------------------------------------------------------- */
302 static INLINE void
303 init_ptrs( stackPos *info, nat ptrs, StgPtr payload )
304 {
305 info->type = posTypePtrs;
306 info->next.ptrs.pos = 0;
307 info->next.ptrs.ptrs = ptrs;
308 info->next.ptrs.payload = payload;
309 }
310
311 /* -----------------------------------------------------------------------------
312 * Find the next object from *info.
313 * -------------------------------------------------------------------------- */
314 static INLINE StgClosure *
315 find_ptrs( stackPos *info )
316 {
317 if (info->next.ptrs.pos < info->next.ptrs.ptrs) {
318 return (StgClosure *)info->next.ptrs.payload[info->next.ptrs.pos++];
319 } else {
320 return NULL;
321 }
322 }
323
324 /* -----------------------------------------------------------------------------
325 * Initializes *info from SRT information stored in *infoTable.
326 * -------------------------------------------------------------------------- */
327 static INLINE void
328 init_srt_fun( stackPos *info, StgFunInfoTable *infoTable )
329 {
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;
334 } else {
335 info->type = posTypeSRT;
336 info->next.srt.srt = (StgClosure **)GET_FUN_SRT(infoTable);
337 info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
338 }
339 }
340
341 static INLINE void
342 init_srt_thunk( stackPos *info, StgThunkInfoTable *infoTable )
343 {
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;
348 } else {
349 info->type = posTypeSRT;
350 info->next.srt.srt = (StgClosure **)GET_SRT(infoTable);
351 info->next.srt.srt_bitmap = infoTable->i.srt_bitmap;
352 }
353 }
354
355 /* -----------------------------------------------------------------------------
356 * Find the next object from *info.
357 * -------------------------------------------------------------------------- */
358 static INLINE StgClosure *
359 find_srt( stackPos *info )
360 {
361 StgClosure *c;
362 StgWord bitmap;
363
364 if (info->type == posTypeSRT) {
365 // Small SRT bitmap
366 bitmap = info->next.srt.srt_bitmap;
367 while (bitmap != 0) {
368 if ((bitmap & 1) != 0) {
369 #if defined(__PIC__) && defined(mingw32_TARGET_OS)
370 if ((unsigned long)(*(info->next.srt.srt)) & 0x1)
371 c = (* (StgClosure **)((unsigned long)*(info->next.srt.srt)) & ~0x1);
372 else
373 c = *(info->next.srt.srt);
374 #else
375 c = *(info->next.srt.srt);
376 #endif
377 bitmap = bitmap >> 1;
378 info->next.srt.srt++;
379 info->next.srt.srt_bitmap = bitmap;
380 return c;
381 }
382 bitmap = bitmap >> 1;
383 info->next.srt.srt++;
384 }
385 // bitmap is now zero...
386 return NULL;
387 }
388 else {
389 // Large SRT bitmap
390 nat i = info->next.large_srt.offset;
391 StgWord bitmap;
392
393 // Follow the pattern from GC.c:scavenge_large_srt_bitmap().
394 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
395 bitmap = bitmap >> (i % BITS_IN(StgWord));
396 while (i < info->next.large_srt.srt->l.size) {
397 if ((bitmap & 1) != 0) {
398 c = ((StgClosure **)info->next.large_srt.srt->srt)[i];
399 i++;
400 info->next.large_srt.offset = i;
401 return c;
402 }
403 i++;
404 if (i % BITS_IN(W_) == 0) {
405 bitmap = info->next.large_srt.srt->l.bitmap[i / BITS_IN(W_)];
406 } else {
407 bitmap = bitmap >> 1;
408 }
409 }
410 // reached the end of this bitmap.
411 info->next.large_srt.offset = i;
412 return NULL;
413 }
414 }
415
416 /* -----------------------------------------------------------------------------
417 * push() pushes a stackElement representing the next child of *c
418 * onto the traverse stack. If *c has no child, *first_child is set
419 * to NULL and nothing is pushed onto the stack. If *c has only one
420 * child, *c_chlid is set to that child and nothing is pushed onto
421 * the stack. If *c has more than two children, *first_child is set
422 * to the first child and a stackElement representing the second
423 * child is pushed onto the stack.
424
425 * Invariants:
426 * *c_child_r is the most recent retainer of *c's children.
427 * *c is not any of TSO, AP, PAP, AP_STACK, which means that
428 * there cannot be any stack objects.
429 * Note: SRTs are considered to be children as well.
430 * -------------------------------------------------------------------------- */
431 static INLINE void
432 push( StgClosure *c, retainer c_child_r, StgClosure **first_child )
433 {
434 stackElement se;
435 bdescr *nbd; // Next Block Descriptor
436
437 #ifdef DEBUG_RETAINER
438 // debugBelch("push(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
439 #endif
440
441 ASSERT(get_itbl(c)->type != TSO);
442 ASSERT(get_itbl(c)->type != AP_STACK);
443
444 //
445 // fill in se
446 //
447
448 se.c = c;
449 se.c_child_r = c_child_r;
450
451 // fill in se.info
452 switch (get_itbl(c)->type) {
453 // no child, no SRT
454 case CONSTR_0_1:
455 case CONSTR_0_2:
456 case ARR_WORDS:
457 *first_child = NULL;
458 return;
459
460 // one child (fixed), no SRT
461 case MUT_VAR_CLEAN:
462 case MUT_VAR_DIRTY:
463 *first_child = ((StgMutVar *)c)->var;
464 return;
465 case THUNK_SELECTOR:
466 *first_child = ((StgSelector *)c)->selectee;
467 return;
468 case IND_PERM:
469 case BLACKHOLE:
470 *first_child = ((StgInd *)c)->indirectee;
471 return;
472 case CONSTR_1_0:
473 case CONSTR_1_1:
474 *first_child = c->payload[0];
475 return;
476
477 // For CONSTR_2_0 and MVAR, we use se.info.step to record the position
478 // of the next child. We do not write a separate initialization code.
479 // Also we do not have to initialize info.type;
480
481 // two children (fixed), no SRT
482 // need to push a stackElement, but nothing to store in se.info
483 case CONSTR_2_0:
484 *first_child = c->payload[0]; // return the first pointer
485 // se.info.type = posTypeStep;
486 // se.info.next.step = 2; // 2 = second
487 break;
488
489 // three children (fixed), no SRT
490 // need to push a stackElement
491 case MVAR_CLEAN:
492 case MVAR_DIRTY:
493 // head must be TSO and the head of a linked list of TSOs.
494 // Shoule it be a child? Seems to be yes.
495 *first_child = (StgClosure *)((StgMVar *)c)->head;
496 // se.info.type = posTypeStep;
497 se.info.next.step = 2; // 2 = second
498 break;
499
500 // three children (fixed), no SRT
501 case WEAK:
502 *first_child = ((StgWeak *)c)->key;
503 // se.info.type = posTypeStep;
504 se.info.next.step = 2;
505 break;
506
507 // layout.payload.ptrs, no SRT
508 case CONSTR:
509 case PRIM:
510 case MUT_PRIM:
511 case BCO:
512 case CONSTR_STATIC:
513 init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs,
514 (StgPtr)c->payload);
515 *first_child = find_ptrs(&se.info);
516 if (*first_child == NULL)
517 return; // no child
518 break;
519
520 // StgMutArrPtr.ptrs, no SRT
521 case MUT_ARR_PTRS_CLEAN:
522 case MUT_ARR_PTRS_DIRTY:
523 case MUT_ARR_PTRS_FROZEN:
524 case MUT_ARR_PTRS_FROZEN0:
525 init_ptrs(&se.info, ((StgMutArrPtrs *)c)->ptrs,
526 (StgPtr)(((StgMutArrPtrs *)c)->payload));
527 *first_child = find_ptrs(&se.info);
528 if (*first_child == NULL)
529 return;
530 break;
531
532 // layout.payload.ptrs, SRT
533 case FUN: // *c is a heap object.
534 case FUN_2_0:
535 init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs, (StgPtr)c->payload);
536 *first_child = find_ptrs(&se.info);
537 if (*first_child == NULL)
538 // no child from ptrs, so check SRT
539 goto fun_srt_only;
540 break;
541
542 case THUNK:
543 case THUNK_2_0:
544 init_ptrs(&se.info, get_itbl(c)->layout.payload.ptrs,
545 (StgPtr)((StgThunk *)c)->payload);
546 *first_child = find_ptrs(&se.info);
547 if (*first_child == NULL)
548 // no child from ptrs, so check SRT
549 goto thunk_srt_only;
550 break;
551
552 // 1 fixed child, SRT
553 case FUN_1_0:
554 case FUN_1_1:
555 *first_child = c->payload[0];
556 ASSERT(*first_child != NULL);
557 init_srt_fun(&se.info, get_fun_itbl(c));
558 break;
559
560 case THUNK_1_0:
561 case THUNK_1_1:
562 *first_child = ((StgThunk *)c)->payload[0];
563 ASSERT(*first_child != NULL);
564 init_srt_thunk(&se.info, get_thunk_itbl(c));
565 break;
566
567 case FUN_STATIC: // *c is a heap object.
568 ASSERT(get_itbl(c)->srt_bitmap != 0);
569 case FUN_0_1:
570 case FUN_0_2:
571 fun_srt_only:
572 init_srt_fun(&se.info, get_fun_itbl(c));
573 *first_child = find_srt(&se.info);
574 if (*first_child == NULL)
575 return; // no child
576 break;
577
578 // SRT only
579 case THUNK_STATIC:
580 ASSERT(get_itbl(c)->srt_bitmap != 0);
581 case THUNK_0_1:
582 case THUNK_0_2:
583 thunk_srt_only:
584 init_srt_thunk(&se.info, get_thunk_itbl(c));
585 *first_child = find_srt(&se.info);
586 if (*first_child == NULL)
587 return; // no child
588 break;
589
590 case TREC_CHUNK:
591 *first_child = (StgClosure *)((StgTRecChunk *)c)->prev_chunk;
592 se.info.next.step = 0; // entry no.
593 break;
594
595 // cannot appear
596 case PAP:
597 case AP:
598 case AP_STACK:
599 case TSO:
600 case IND_STATIC:
601 case CONSTR_NOCAF_STATIC:
602 // stack objects
603 case UPDATE_FRAME:
604 case CATCH_FRAME:
605 case STOP_FRAME:
606 case RET_DYN:
607 case RET_BCO:
608 case RET_SMALL:
609 case RET_BIG:
610 // invalid objects
611 case IND:
612 case INVALID_OBJECT:
613 default:
614 barf("Invalid object *c in push()");
615 return;
616 }
617
618 if (stackTop - 1 < stackBottom) {
619 #ifdef DEBUG_RETAINER
620 // debugBelch("push() to the next stack.\n");
621 #endif
622 // currentStack->free is updated when the active stack is switched
623 // to the next stack.
624 currentStack->free = (StgPtr)stackTop;
625
626 if (currentStack->link == NULL) {
627 nbd = allocGroup(BLOCKS_IN_STACK);
628 nbd->link = NULL;
629 nbd->u.back = currentStack;
630 currentStack->link = nbd;
631 } else
632 nbd = currentStack->link;
633
634 newStackBlock(nbd);
635 }
636
637 // adjust stackTop (acutal push)
638 stackTop--;
639 // If the size of stackElement was huge, we would better replace the
640 // following statement by either a memcpy() call or a switch statement
641 // on the type of the element. Currently, the size of stackElement is
642 // small enough (5 words) that this direct assignment seems to be enough.
643
644 // ToDo: The line below leads to the warning:
645 // warning: 'se.info.type' may be used uninitialized in this function
646 // This is caused by the fact that there are execution paths through the
647 // large switch statement above where some cases do not initialize this
648 // field. Is this really harmless? Can we avoid the warning?
649 *stackTop = se;
650
651 #ifdef DEBUG_RETAINER
652 stackSize++;
653 if (stackSize > maxStackSize) maxStackSize = stackSize;
654 // ASSERT(stackSize >= 0);
655 // debugBelch("stackSize = %d\n", stackSize);
656 #endif
657 }
658
659 /* -----------------------------------------------------------------------------
660 * popOff() and popOffReal(): Pop a stackElement off the traverse stack.
661 * Invariants:
662 * stackTop cannot be equal to stackLimit unless the whole stack is
663 * empty, in which case popOff() is not allowed.
664 * Note:
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 * -------------------------------------------------------------------------- */
672 static void
673 popOffReal(void)
674 {
675 bdescr *pbd; // Previous Block Descriptor
676
677 #ifdef DEBUG_RETAINER
678 // debugBelch("pop() to the previous stack.\n");
679 #endif
680
681 ASSERT(stackTop + 1 == stackLimit);
682 ASSERT(stackBottom == (stackElement *)currentStack->start);
683
684 if (firstStack == currentStack) {
685 // The stack is completely empty.
686 stackTop++;
687 ASSERT(stackTop == stackLimit);
688 #ifdef DEBUG_RETAINER
689 stackSize--;
690 if (stackSize > maxStackSize) maxStackSize = stackSize;
691 /*
692 ASSERT(stackSize >= 0);
693 debugBelch("stackSize = %d\n", stackSize);
694 */
695 #endif
696 return;
697 }
698
699 // currentStack->free is updated when the active stack is switched back
700 // to the previous stack.
701 currentStack->free = (StgPtr)stackLimit;
702
703 // find the previous block descriptor
704 pbd = currentStack->u.back;
705 ASSERT(pbd != NULL);
706
707 returnToOldStack(pbd);
708
709 #ifdef DEBUG_RETAINER
710 stackSize--;
711 if (stackSize > maxStackSize) maxStackSize = stackSize;
712 /*
713 ASSERT(stackSize >= 0);
714 debugBelch("stackSize = %d\n", stackSize);
715 */
716 #endif
717 }
718
719 static INLINE void
720 popOff(void) {
721 #ifdef DEBUG_RETAINER
722 // debugBelch("\tpopOff(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
723 #endif
724
725 ASSERT(stackTop != stackLimit);
726 ASSERT(!isEmptyRetainerStack());
727
728 // <= (instead of <) is wrong!
729 if (stackTop + 1 < stackLimit) {
730 stackTop++;
731 #ifdef DEBUG_RETAINER
732 stackSize--;
733 if (stackSize > maxStackSize) maxStackSize = stackSize;
734 /*
735 ASSERT(stackSize >= 0);
736 debugBelch("stackSize = %d\n", stackSize);
737 */
738 #endif
739 return;
740 }
741
742 popOffReal();
743 }
744
745 /* -----------------------------------------------------------------------------
746 * Finds the next object to be considered for retainer profiling and store
747 * its pointer to *c.
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
752 * the next object.
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.
757 * Note:
758 * It is okay to call this function even when the current stack chunk
759 * is empty.
760 * -------------------------------------------------------------------------- */
761 static INLINE void
762 pop( StgClosure **c, StgClosure **cp, retainer *r )
763 {
764 stackElement *se;
765
766 #ifdef DEBUG_RETAINER
767 // debugBelch("pop(): stackTop = 0x%x, currentStackBoundary = 0x%x\n", stackTop, currentStackBoundary);
768 #endif
769
770 do {
771 if (isOnBoundary()) { // if the current stack chunk is depleted
772 *c = NULL;
773 return;
774 }
775
776 se = stackTop;
777
778 switch (get_itbl(se->c)->type) {
779 // two children (fixed), no SRT
780 // nothing in se.info
781 case CONSTR_2_0:
782 *c = se->c->payload[1];
783 *cp = se->c;
784 *r = se->c_child_r;
785 popOff();
786 return;
787
788 // three children (fixed), no SRT
789 // need to push a stackElement
790 case MVAR_CLEAN:
791 case MVAR_DIRTY:
792 if (se->info.next.step == 2) {
793 *c = (StgClosure *)((StgMVar *)se->c)->tail;
794 se->info.next.step++; // move to the next step
795 // no popOff
796 } else {
797 *c = ((StgMVar *)se->c)->value;
798 popOff();
799 }
800 *cp = se->c;
801 *r = se->c_child_r;
802 return;
803
804 // three children (fixed), no SRT
805 case WEAK:
806 if (se->info.next.step == 2) {
807 *c = ((StgWeak *)se->c)->value;
808 se->info.next.step++;
809 // no popOff
810 } else {
811 *c = ((StgWeak *)se->c)->finalizer;
812 popOff();
813 }
814 *cp = se->c;
815 *r = se->c_child_r;
816 return;
817
818 case TREC_CHUNK: {
819 // These are pretty complicated: we have N entries, each
820 // of which contains 3 fields that we want to follow. So
821 // we divide the step counter: the 2 low bits indicate
822 // which field, and the rest of the bits indicate the
823 // entry number (starting from zero).
824 TRecEntry *entry;
825 nat entry_no = se->info.next.step >> 2;
826 nat field_no = se->info.next.step & 3;
827 if (entry_no == ((StgTRecChunk *)se->c)->next_entry_idx) {
828 *c = NULL;
829 popOff();
830 return;
831 }
832 entry = &((StgTRecChunk *)se->c)->entries[entry_no];
833 if (field_no == 0) {
834 *c = (StgClosure *)entry->tvar;
835 } else if (field_no == 1) {
836 *c = entry->expected_value;
837 } else {
838 *c = entry->new_value;
839 }
840 *cp = se->c;
841 *r = se->c_child_r;
842 se->info.next.step++;
843 return;
844 }
845
846 case CONSTR:
847 case PRIM:
848 case MUT_PRIM:
849 case BCO:
850 case CONSTR_STATIC:
851 // StgMutArrPtr.ptrs, no SRT
852 case MUT_ARR_PTRS_CLEAN:
853 case MUT_ARR_PTRS_DIRTY:
854 case MUT_ARR_PTRS_FROZEN:
855 case MUT_ARR_PTRS_FROZEN0:
856 *c = find_ptrs(&se->info);
857 if (*c == NULL) {
858 popOff();
859 break;
860 }
861 *cp = se->c;
862 *r = se->c_child_r;
863 return;
864
865 // layout.payload.ptrs, SRT
866 case FUN: // always a heap object
867 case FUN_2_0:
868 if (se->info.type == posTypePtrs) {
869 *c = find_ptrs(&se->info);
870 if (*c != NULL) {
871 *cp = se->c;
872 *r = se->c_child_r;
873 return;
874 }
875 init_srt_fun(&se->info, get_fun_itbl(se->c));
876 }
877 goto do_srt;
878
879 case THUNK:
880 case THUNK_2_0:
881 if (se->info.type == posTypePtrs) {
882 *c = find_ptrs(&se->info);
883 if (*c != NULL) {
884 *cp = se->c;
885 *r = se->c_child_r;
886 return;
887 }
888 init_srt_thunk(&se->info, get_thunk_itbl(se->c));
889 }
890 goto do_srt;
891
892 // SRT
893 do_srt:
894 case THUNK_STATIC:
895 case FUN_STATIC:
896 case FUN_0_1:
897 case FUN_0_2:
898 case THUNK_0_1:
899 case THUNK_0_2:
900 case FUN_1_0:
901 case FUN_1_1:
902 case THUNK_1_0:
903 case THUNK_1_1:
904 *c = find_srt(&se->info);
905 if (*c != NULL) {
906 *cp = se->c;
907 *r = se->c_child_r;
908 return;
909 }
910 popOff();
911 break;
912
913 // no child (fixed), no SRT
914 case CONSTR_0_1:
915 case CONSTR_0_2:
916 case ARR_WORDS:
917 // one child (fixed), no SRT
918 case MUT_VAR_CLEAN:
919 case MUT_VAR_DIRTY:
920 case THUNK_SELECTOR:
921 case IND_PERM:
922 case CONSTR_1_1:
923 // cannot appear
924 case PAP:
925 case AP:
926 case AP_STACK:
927 case TSO:
928 case IND_STATIC:
929 case CONSTR_NOCAF_STATIC:
930 // stack objects
931 case RET_DYN:
932 case UPDATE_FRAME:
933 case CATCH_FRAME:
934 case STOP_FRAME:
935 case RET_BCO:
936 case RET_SMALL:
937 case RET_BIG:
938 // invalid objects
939 case IND:
940 case INVALID_OBJECT:
941 default:
942 barf("Invalid object *c in pop()");
943 return;
944 }
945 } while (rtsTrue);
946 }
947
948 /* -----------------------------------------------------------------------------
949 * RETAINER PROFILING ENGINE
950 * -------------------------------------------------------------------------- */
951
952 void
953 initRetainerProfiling( void )
954 {
955 initializeAllRetainerSet();
956 retainerGeneration = 0;
957 }
958
959 /* -----------------------------------------------------------------------------
960 * This function must be called before f-closing prof_file.
961 * -------------------------------------------------------------------------- */
962 void
963 endRetainerProfiling( void )
964 {
965 #ifdef SECOND_APPROACH
966 outputAllRetainerSet(prof_file);
967 #endif
968 }
969
970 /* -----------------------------------------------------------------------------
971 * Returns the actual pointer to the retainer set of the closure *c.
972 * It may adjust RSET(c) subject to flip.
973 * Side effects:
974 * RSET(c) is initialized to NULL if its current value does not
975 * conform to flip.
976 * Note:
977 * Even though this function has side effects, they CAN be ignored because
978 * subsequent calls to retainerSetOf() always result in the same return value
979 * and retainerSetOf() is the only way to retrieve retainerSet of a given
980 * closure.
981 * We have to perform an XOR (^) operation each time a closure is examined.
982 * The reason is that we do not know when a closure is visited last.
983 * -------------------------------------------------------------------------- */
984 static INLINE void
985 maybeInitRetainerSet( StgClosure *c )
986 {
987 if (!isRetainerSetFieldValid(c)) {
988 setRetainerSetToNull(c);
989 }
990 }
991
992 /* -----------------------------------------------------------------------------
993 * Returns rtsTrue if *c is a retainer.
994 * -------------------------------------------------------------------------- */
995 static INLINE rtsBool
996 isRetainer( StgClosure *c )
997 {
998 switch (get_itbl(c)->type) {
999 //
1000 // True case
1001 //
1002 // TSOs MUST be retainers: they constitute the set of roots.
1003 case TSO:
1004
1005 // mutable objects
1006 case MUT_PRIM:
1007 case MVAR_CLEAN:
1008 case MVAR_DIRTY:
1009 case MUT_VAR_CLEAN:
1010 case MUT_VAR_DIRTY:
1011 case MUT_ARR_PTRS_CLEAN:
1012 case MUT_ARR_PTRS_DIRTY:
1013 case MUT_ARR_PTRS_FROZEN:
1014 case MUT_ARR_PTRS_FROZEN0:
1015
1016 // thunks are retainers.
1017 case THUNK:
1018 case THUNK_1_0:
1019 case THUNK_0_1:
1020 case THUNK_2_0:
1021 case THUNK_1_1:
1022 case THUNK_0_2:
1023 case THUNK_SELECTOR:
1024 case AP:
1025 case AP_STACK:
1026
1027 // Static thunks, or CAFS, are obviously retainers.
1028 case THUNK_STATIC:
1029
1030 // WEAK objects are roots; there is separate code in which traversing
1031 // begins from WEAK objects.
1032 case WEAK:
1033 return rtsTrue;
1034
1035 //
1036 // False case
1037 //
1038
1039 // constructors
1040 case CONSTR:
1041 case CONSTR_1_0:
1042 case CONSTR_0_1:
1043 case CONSTR_2_0:
1044 case CONSTR_1_1:
1045 case CONSTR_0_2:
1046 // functions
1047 case FUN:
1048 case FUN_1_0:
1049 case FUN_0_1:
1050 case FUN_2_0:
1051 case FUN_1_1:
1052 case FUN_0_2:
1053 // partial applications
1054 case PAP:
1055 // indirection
1056 case IND_PERM:
1057 case BLACKHOLE:
1058 // static objects
1059 case CONSTR_STATIC:
1060 case FUN_STATIC:
1061 // misc
1062 case PRIM:
1063 case BCO:
1064 case ARR_WORDS:
1065 // STM
1066 case TREC_CHUNK:
1067 return rtsFalse;
1068
1069 //
1070 // Error case
1071 //
1072 // IND_STATIC cannot be *c, *cp, *r in the retainer profiling loop.
1073 case IND_STATIC:
1074 // CONSTR_NOCAF_STATIC
1075 // cannot be *c, *cp, *r in the retainer profiling loop.
1076 case CONSTR_NOCAF_STATIC:
1077 // Stack objects are invalid because they are never treated as
1078 // legal objects during retainer profiling.
1079 case UPDATE_FRAME:
1080 case CATCH_FRAME:
1081 case STOP_FRAME:
1082 case RET_DYN:
1083 case RET_BCO:
1084 case RET_SMALL:
1085 case RET_BIG:
1086 // other cases
1087 case IND:
1088 case INVALID_OBJECT:
1089 default:
1090 barf("Invalid object in isRetainer(): %d", get_itbl(c)->type);
1091 return rtsFalse;
1092 }
1093 }
1094
1095 /* -----------------------------------------------------------------------------
1096 * Returns the retainer function value for the closure *c, i.e., R(*c).
1097 * This function does NOT return the retainer(s) of *c.
1098 * Invariants:
1099 * *c must be a retainer.
1100 * Note:
1101 * Depending on the definition of this function, the maintenance of retainer
1102 * sets can be made easier. If most retainer sets are likely to be created
1103 * again across garbage collections, refreshAllRetainerSet() in
1104 * RetainerSet.c can simply do nothing.
1105 * If this is not the case, we can free all the retainer sets and
1106 * re-initialize the hash table.
1107 * See refreshAllRetainerSet() in RetainerSet.c.
1108 * -------------------------------------------------------------------------- */
1109 static INLINE retainer
1110 getRetainerFrom( StgClosure *c )
1111 {
1112 ASSERT(isRetainer(c));
1113
1114 #if defined(RETAINER_SCHEME_INFO)
1115 // Retainer scheme 1: retainer = info table
1116 return get_itbl(c);
1117 #elif defined(RETAINER_SCHEME_CCS)
1118 // Retainer scheme 2: retainer = cost centre stack
1119 return c->header.prof.ccs;
1120 #elif defined(RETAINER_SCHEME_CC)
1121 // Retainer scheme 3: retainer = cost centre
1122 return c->header.prof.ccs->cc;
1123 #endif
1124 }
1125
1126 /* -----------------------------------------------------------------------------
1127 * Associates the retainer set *s with the closure *c, that is, *s becomes
1128 * the retainer set of *c.
1129 * Invariants:
1130 * c != NULL
1131 * s != NULL
1132 * -------------------------------------------------------------------------- */
1133 static INLINE void
1134 associate( StgClosure *c, RetainerSet *s )
1135 {
1136 // StgWord has the same size as pointers, so the following type
1137 // casting is okay.
1138 RSET(c) = (RetainerSet *)((StgWord)s | flip);
1139 }
1140
1141 /* -----------------------------------------------------------------------------
1142 Call retainClosure for each of the closures covered by a large bitmap.
1143 -------------------------------------------------------------------------- */
1144
1145 static void
1146 retain_large_bitmap (StgPtr p, StgLargeBitmap *large_bitmap, nat size,
1147 StgClosure *c, retainer c_child_r)
1148 {
1149 nat i, b;
1150 StgWord bitmap;
1151
1152 b = 0;
1153 bitmap = large_bitmap->bitmap[b];
1154 for (i = 0; i < size; ) {
1155 if ((bitmap & 1) == 0) {
1156 retainClosure((StgClosure *)*p, c, c_child_r);
1157 }
1158 i++;
1159 p++;
1160 if (i % BITS_IN(W_) == 0) {
1161 b++;
1162 bitmap = large_bitmap->bitmap[b];
1163 } else {
1164 bitmap = bitmap >> 1;
1165 }
1166 }
1167 }
1168
1169 static INLINE StgPtr
1170 retain_small_bitmap (StgPtr p, nat size, StgWord bitmap,
1171 StgClosure *c, retainer c_child_r)
1172 {
1173 while (size > 0) {
1174 if ((bitmap & 1) == 0) {
1175 retainClosure((StgClosure *)*p, c, c_child_r);
1176 }
1177 p++;
1178 bitmap = bitmap >> 1;
1179 size--;
1180 }
1181 return p;
1182 }
1183
1184 /* -----------------------------------------------------------------------------
1185 * Call retainClosure for each of the closures in an SRT.
1186 * ------------------------------------------------------------------------- */
1187
1188 static void
1189 retain_large_srt_bitmap (StgLargeSRT *srt, StgClosure *c, retainer c_child_r)
1190 {
1191 nat i, b, size;
1192 StgWord bitmap;
1193 StgClosure **p;
1194
1195 b = 0;
1196 p = (StgClosure **)srt->srt;
1197 size = srt->l.size;
1198 bitmap = srt->l.bitmap[b];
1199 for (i = 0; i < size; ) {
1200 if ((bitmap & 1) != 0) {
1201 retainClosure((StgClosure *)*p, c, c_child_r);
1202 }
1203 i++;
1204 p++;
1205 if (i % BITS_IN(W_) == 0) {
1206 b++;
1207 bitmap = srt->l.bitmap[b];
1208 } else {
1209 bitmap = bitmap >> 1;
1210 }
1211 }
1212 }
1213
1214 static INLINE void
1215 retainSRT (StgClosure **srt, nat srt_bitmap, StgClosure *c, retainer c_child_r)
1216 {
1217 nat bitmap;
1218 StgClosure **p;
1219
1220 bitmap = srt_bitmap;
1221 p = srt;
1222
1223 if (bitmap == (StgHalfWord)(-1)) {
1224 retain_large_srt_bitmap( (StgLargeSRT *)srt, c, c_child_r );
1225 return;
1226 }
1227
1228 while (bitmap != 0) {
1229 if ((bitmap & 1) != 0) {
1230 #if defined(__PIC__) && defined(mingw32_TARGET_OS)
1231 if ( (unsigned long)(*srt) & 0x1 ) {
1232 retainClosure(* (StgClosure**) ((unsigned long) (*srt) & ~0x1),
1233 c, c_child_r);
1234 } else {
1235 retainClosure(*srt,c,c_child_r);
1236 }
1237 #else
1238 retainClosure(*srt,c,c_child_r);
1239 #endif
1240 }
1241 p++;
1242 bitmap = bitmap >> 1;
1243 }
1244 }
1245
1246 /* -----------------------------------------------------------------------------
1247 * Process all the objects in the stack chunk from stackStart to stackEnd
1248 * with *c and *c_child_r being their parent and their most recent retainer,
1249 * respectively. Treat stackOptionalFun as another child of *c if it is
1250 * not NULL.
1251 * Invariants:
1252 * *c is one of the following: TSO, AP_STACK.
1253 * If *c is TSO, c == c_child_r.
1254 * stackStart < stackEnd.
1255 * RSET(c) and RSET(c_child_r) are valid, i.e., their
1256 * interpretation conforms to the current value of flip (even when they
1257 * are interpreted to be NULL).
1258 * If *c is TSO, its state is not any of ThreadRelocated, ThreadComplete,
1259 * or ThreadKilled, which means that its stack is ready to process.
1260 * Note:
1261 * This code was almost plagiarzied from GC.c! For each pointer,
1262 * retainClosure() is invoked instead of evacuate().
1263 * -------------------------------------------------------------------------- */
1264 static void
1265 retainStack( StgClosure *c, retainer c_child_r,
1266 StgPtr stackStart, StgPtr stackEnd )
1267 {
1268 stackElement *oldStackBoundary;
1269 StgPtr p;
1270 StgRetInfoTable *info;
1271 StgWord32 bitmap;
1272 nat size;
1273
1274 #ifdef DEBUG_RETAINER
1275 cStackSize++;
1276 if (cStackSize > maxCStackSize) maxCStackSize = cStackSize;
1277 #endif
1278
1279 /*
1280 Each invocation of retainStack() creates a new virtual
1281 stack. Since all such stacks share a single common stack, we
1282 record the current currentStackBoundary, which will be restored
1283 at the exit.
1284 */
1285 oldStackBoundary = currentStackBoundary;
1286 currentStackBoundary = stackTop;
1287
1288 #ifdef DEBUG_RETAINER
1289 // debugBelch("retainStack() called: oldStackBoundary = 0x%x, currentStackBoundary = 0x%x\n", oldStackBoundary, currentStackBoundary);
1290 #endif
1291
1292 ASSERT(get_itbl(c)->type != TSO ||
1293 (((StgTSO *)c)->what_next != ThreadRelocated &&
1294 ((StgTSO *)c)->what_next != ThreadComplete &&
1295 ((StgTSO *)c)->what_next != ThreadKilled));
1296
1297 p = stackStart;
1298 while (p < stackEnd) {
1299 info = get_ret_itbl((StgClosure *)p);
1300
1301 switch(info->i.type) {
1302
1303 case UPDATE_FRAME:
1304 retainClosure(((StgUpdateFrame *)p)->updatee, c, c_child_r);
1305 p += sizeofW(StgUpdateFrame);
1306 continue;
1307
1308 case STOP_FRAME:
1309 case CATCH_FRAME:
1310 case CATCH_STM_FRAME:
1311 case CATCH_RETRY_FRAME:
1312 case ATOMICALLY_FRAME:
1313 case RET_SMALL:
1314 bitmap = BITMAP_BITS(info->i.layout.bitmap);
1315 size = BITMAP_SIZE(info->i.layout.bitmap);
1316 p++;
1317 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1318
1319 follow_srt:
1320 retainSRT((StgClosure **)GET_SRT(info), info->i.srt_bitmap, c, c_child_r);
1321 continue;
1322
1323 case RET_BCO: {
1324 StgBCO *bco;
1325
1326 p++;
1327 retainClosure((StgClosure *)*p, c, c_child_r);
1328 bco = (StgBCO *)*p;
1329 p++;
1330 size = BCO_BITMAP_SIZE(bco);
1331 retain_large_bitmap(p, BCO_BITMAP(bco), size, c, c_child_r);
1332 p += size;
1333 continue;
1334 }
1335
1336 // large bitmap (> 32 entries, or > 64 on a 64-bit machine)
1337 case RET_BIG:
1338 size = GET_LARGE_BITMAP(&info->i)->size;
1339 p++;
1340 retain_large_bitmap(p, GET_LARGE_BITMAP(&info->i),
1341 size, c, c_child_r);
1342 p += size;
1343 // and don't forget to follow the SRT
1344 goto follow_srt;
1345
1346 // Dynamic bitmap: the mask is stored on the stack
1347 case RET_DYN: {
1348 StgWord dyn;
1349 dyn = ((StgRetDyn *)p)->liveness;
1350
1351 // traverse the bitmap first
1352 bitmap = RET_DYN_LIVENESS(dyn);
1353 p = (P_)&((StgRetDyn *)p)->payload[0];
1354 size = RET_DYN_BITMAP_SIZE;
1355 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1356
1357 // skip over the non-ptr words
1358 p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
1359
1360 // follow the ptr words
1361 for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
1362 retainClosure((StgClosure *)*p, c, c_child_r);
1363 p++;
1364 }
1365 continue;
1366 }
1367
1368 case RET_FUN: {
1369 StgRetFun *ret_fun = (StgRetFun *)p;
1370 StgFunInfoTable *fun_info;
1371
1372 retainClosure(ret_fun->fun, c, c_child_r);
1373 fun_info = get_fun_itbl(UNTAG_CLOSURE(ret_fun->fun));
1374
1375 p = (P_)&ret_fun->payload;
1376 switch (fun_info->f.fun_type) {
1377 case ARG_GEN:
1378 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1379 size = BITMAP_SIZE(fun_info->f.b.bitmap);
1380 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1381 break;
1382 case ARG_GEN_BIG:
1383 size = GET_FUN_LARGE_BITMAP(fun_info)->size;
1384 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1385 size, c, c_child_r);
1386 p += size;
1387 break;
1388 default:
1389 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1390 size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
1391 p = retain_small_bitmap(p, size, bitmap, c, c_child_r);
1392 break;
1393 }
1394 goto follow_srt;
1395 }
1396
1397 default:
1398 barf("Invalid object found in retainStack(): %d",
1399 (int)(info->i.type));
1400 }
1401 }
1402
1403 // restore currentStackBoundary
1404 currentStackBoundary = oldStackBoundary;
1405 #ifdef DEBUG_RETAINER
1406 // debugBelch("retainStack() finished: currentStackBoundary = 0x%x\n", currentStackBoundary);
1407 #endif
1408
1409 #ifdef DEBUG_RETAINER
1410 cStackSize--;
1411 #endif
1412 }
1413
1414 /* ----------------------------------------------------------------------------
1415 * Call retainClosure for each of the children of a PAP/AP
1416 * ------------------------------------------------------------------------- */
1417
1418 static INLINE StgPtr
1419 retain_PAP_payload (StgClosure *pap, /* NOT tagged */
1420 retainer c_child_r, /* NOT tagged */
1421 StgClosure *fun, /* tagged */
1422 StgClosure** payload, StgWord n_args)
1423 {
1424 StgPtr p;
1425 StgWord bitmap;
1426 StgFunInfoTable *fun_info;
1427
1428 retainClosure(fun, pap, c_child_r);
1429 fun = UNTAG_CLOSURE(fun);
1430 fun_info = get_fun_itbl(fun);
1431 ASSERT(fun_info->i.type != PAP);
1432
1433 p = (StgPtr)payload;
1434
1435 switch (fun_info->f.fun_type) {
1436 case ARG_GEN:
1437 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
1438 p = retain_small_bitmap(p, n_args, bitmap,
1439 pap, c_child_r);
1440 break;
1441 case ARG_GEN_BIG:
1442 retain_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info),
1443 n_args, pap, c_child_r);
1444 p += n_args;
1445 break;
1446 case ARG_BCO:
1447 retain_large_bitmap((StgPtr)payload, BCO_BITMAP(fun),
1448 n_args, pap, c_child_r);
1449 p += n_args;
1450 break;
1451 default:
1452 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
1453 p = retain_small_bitmap(p, n_args, bitmap, pap, c_child_r);
1454 break;
1455 }
1456 return p;
1457 }
1458
1459 /* -----------------------------------------------------------------------------
1460 * Compute the retainer set of *c0 and all its desecents by traversing.
1461 * *cp0 is the parent of *c0, and *r0 is the most recent retainer of *c0.
1462 * Invariants:
1463 * c0 = cp0 = r0 holds only for root objects.
1464 * RSET(cp0) and RSET(r0) are valid, i.e., their
1465 * interpretation conforms to the current value of flip (even when they
1466 * are interpreted to be NULL).
1467 * However, RSET(c0) may be corrupt, i.e., it may not conform to
1468 * the current value of flip. If it does not, during the execution
1469 * of this function, RSET(c0) must be initialized as well as all
1470 * its descendants.
1471 * Note:
1472 * stackTop must be the same at the beginning and the exit of this function.
1473 * *c0 can be TSO (as well as AP_STACK).
1474 * -------------------------------------------------------------------------- */
1475 static void
1476 retainClosure( StgClosure *c0, StgClosure *cp0, retainer r0 )
1477 {
1478 // c = Current closure (possibly tagged)
1479 // cp = Current closure's Parent (NOT tagged)
1480 // r = current closures' most recent Retainer (NOT tagged)
1481 // c_child_r = current closure's children's most recent retainer
1482 // first_child = first child of c
1483 StgClosure *c, *cp, *first_child;
1484 RetainerSet *s, *retainerSetOfc;
1485 retainer r, c_child_r;
1486 StgWord typeOfc;
1487
1488 #ifdef DEBUG_RETAINER
1489 // StgPtr oldStackTop;
1490 #endif
1491
1492 #ifdef DEBUG_RETAINER
1493 // oldStackTop = stackTop;
1494 // debugBelch("retainClosure() called: c0 = 0x%x, cp0 = 0x%x, r0 = 0x%x\n", c0, cp0, r0);
1495 #endif
1496
1497 // (c, cp, r) = (c0, cp0, r0)
1498 c = c0;
1499 cp = cp0;
1500 r = r0;
1501 goto inner_loop;
1502
1503 loop:
1504 //debugBelch("loop");
1505 // pop to (c, cp, r);
1506 pop(&c, &cp, &r);
1507
1508 if (c == NULL) {
1509 #ifdef DEBUG_RETAINER
1510 // debugBelch("retainClosure() ends: oldStackTop = 0x%x, stackTop = 0x%x\n", oldStackTop, stackTop);
1511 #endif
1512 return;
1513 }
1514
1515 //debugBelch("inner_loop");
1516
1517 inner_loop:
1518 c = UNTAG_CLOSURE(c);
1519
1520 // c = current closure under consideration,
1521 // cp = current closure's parent,
1522 // r = current closure's most recent retainer
1523 //
1524 // Loop invariants (on the meaning of c, cp, r, and their retainer sets):
1525 // RSET(cp) and RSET(r) are valid.
1526 // RSET(c) is valid only if c has been visited before.
1527 //
1528 // Loop invariants (on the relation between c, cp, and r)
1529 // if cp is not a retainer, r belongs to RSET(cp).
1530 // if cp is a retainer, r == cp.
1531
1532 typeOfc = get_itbl(c)->type;
1533
1534 #ifdef DEBUG_RETAINER
1535 switch (typeOfc) {
1536 case IND_STATIC:
1537 case CONSTR_NOCAF_STATIC:
1538 case CONSTR_STATIC:
1539 case THUNK_STATIC:
1540 case FUN_STATIC:
1541 break;
1542 default:
1543 if (retainerSetOf(c) == NULL) { // first visit?
1544 costArray[typeOfc] += cost(c);
1545 sumOfNewCost += cost(c);
1546 }
1547 break;
1548 }
1549 #endif
1550
1551 // special cases
1552 switch (typeOfc) {
1553 case TSO:
1554 if (((StgTSO *)c)->what_next == ThreadComplete ||
1555 ((StgTSO *)c)->what_next == ThreadKilled) {
1556 #ifdef DEBUG_RETAINER
1557 debugBelch("ThreadComplete or ThreadKilled encountered in retainClosure()\n");
1558 #endif
1559 goto loop;
1560 }
1561 if (((StgTSO *)c)->what_next == ThreadRelocated) {
1562 #ifdef DEBUG_RETAINER
1563 debugBelch("ThreadRelocated encountered in retainClosure()\n");
1564 #endif
1565 c = (StgClosure *)((StgTSO *)c)->_link;
1566 goto inner_loop;
1567 }
1568 break;
1569
1570 case IND_STATIC:
1571 // We just skip IND_STATIC, so its retainer set is never computed.
1572 c = ((StgIndStatic *)c)->indirectee;
1573 goto inner_loop;
1574 // static objects with no pointers out, so goto loop.
1575 case CONSTR_NOCAF_STATIC:
1576 // It is not just enough not to compute the retainer set for *c; it is
1577 // mandatory because CONSTR_NOCAF_STATIC are not reachable from
1578 // scavenged_static_objects, the list from which is assumed to traverse
1579 // all static objects after major garbage collections.
1580 goto loop;
1581 case THUNK_STATIC:
1582 case FUN_STATIC:
1583 if (get_itbl(c)->srt_bitmap == 0) {
1584 // No need to compute the retainer set; no dynamic objects
1585 // are reachable from *c.
1586 //
1587 // Static objects: if we traverse all the live closures,
1588 // including static closures, during each heap census then
1589 // we will observe that some static closures appear and
1590 // disappear. eg. a closure may contain a pointer to a
1591 // static function 'f' which is not otherwise reachable
1592 // (it doesn't indirectly point to any CAFs, so it doesn't
1593 // appear in any SRTs), so we would find 'f' during
1594 // traversal. However on the next sweep there may be no
1595 // closures pointing to 'f'.
1596 //
1597 // We must therefore ignore static closures whose SRT is
1598 // empty, because these are exactly the closures that may
1599 // "appear". A closure with a non-empty SRT, and which is
1600 // still required, will always be reachable.
1601 //
1602 // But what about CONSTR_STATIC? Surely these may be able
1603 // to appear, and they don't have SRTs, so we can't
1604 // check. So for now, we're calling
1605 // resetStaticObjectForRetainerProfiling() from the
1606 // garbage collector to reset the retainer sets in all the
1607 // reachable static objects.
1608 goto loop;
1609 }
1610 default:
1611 break;
1612 }
1613
1614 // The above objects are ignored in computing the average number of times
1615 // an object is visited.
1616 timesAnyObjectVisited++;
1617
1618 // If this is the first visit to c, initialize its retainer set.
1619 maybeInitRetainerSet(c);
1620 retainerSetOfc = retainerSetOf(c);
1621
1622 // Now compute s:
1623 // isRetainer(cp) == rtsTrue => s == NULL
1624 // isRetainer(cp) == rtsFalse => s == cp.retainer
1625 if (isRetainer(cp))
1626 s = NULL;
1627 else
1628 s = retainerSetOf(cp);
1629
1630 // (c, cp, r, s) is available.
1631
1632 // (c, cp, r, s, R_r) is available, so compute the retainer set for *c.
1633 if (retainerSetOfc == NULL) {
1634 // This is the first visit to *c.
1635 numObjectVisited++;
1636
1637 if (s == NULL)
1638 associate(c, singleton(r));
1639 else
1640 // s is actually the retainer set of *c!
1641 associate(c, s);
1642
1643 // compute c_child_r
1644 c_child_r = isRetainer(c) ? getRetainerFrom(c) : r;
1645 } else {
1646 // This is not the first visit to *c.
1647 if (isMember(r, retainerSetOfc))
1648 goto loop; // no need to process child
1649
1650 if (s == NULL)
1651 associate(c, addElement(r, retainerSetOfc));
1652 else {
1653 // s is not NULL and cp is not a retainer. This means that
1654 // each time *cp is visited, so is *c. Thus, if s has
1655 // exactly one more element in its retainer set than c, s
1656 // is also the new retainer set for *c.
1657 if (s->num == retainerSetOfc->num + 1) {
1658 associate(c, s);
1659 }
1660 // Otherwise, just add R_r to the current retainer set of *c.
1661 else {
1662 associate(c, addElement(r, retainerSetOfc));
1663 }
1664 }
1665
1666 if (isRetainer(c))
1667 goto loop; // no need to process child
1668
1669 // compute c_child_r
1670 c_child_r = r;
1671 }
1672
1673 // now, RSET() of all of *c, *cp, and *r is valid.
1674 // (c, c_child_r) are available.
1675
1676 // process child
1677
1678 // Special case closures: we process these all in one go rather
1679 // than attempting to save the current position, because doing so
1680 // would be hard.
1681 switch (typeOfc) {
1682 case TSO:
1683 retainStack(c, c_child_r,
1684 ((StgTSO *)c)->sp,
1685 ((StgTSO *)c)->stack + ((StgTSO *)c)->stack_size);
1686 goto loop;
1687
1688 case PAP:
1689 {
1690 StgPAP *pap = (StgPAP *)c;
1691 retain_PAP_payload(c, c_child_r, pap->fun, pap->payload, pap->n_args);
1692 goto loop;
1693 }
1694
1695 case AP:
1696 {
1697 StgAP *ap = (StgAP *)c;
1698 retain_PAP_payload(c, c_child_r, ap->fun, ap->payload, ap->n_args);
1699 goto loop;
1700 }
1701
1702 case AP_STACK:
1703 retainClosure(((StgAP_STACK *)c)->fun, c, c_child_r);
1704 retainStack(c, c_child_r,
1705 (StgPtr)((StgAP_STACK *)c)->payload,
1706 (StgPtr)((StgAP_STACK *)c)->payload +
1707 ((StgAP_STACK *)c)->size);
1708 goto loop;
1709 }
1710
1711 push(c, c_child_r, &first_child);
1712
1713 // If first_child is null, c has no child.
1714 // If first_child is not null, the top stack element points to the next
1715 // object. push() may or may not push a stackElement on the stack.
1716 if (first_child == NULL)
1717 goto loop;
1718
1719 // (c, cp, r) = (first_child, c, c_child_r)
1720 r = c_child_r;
1721 cp = c;
1722 c = first_child;
1723 goto inner_loop;
1724 }
1725
1726 /* -----------------------------------------------------------------------------
1727 * Compute the retainer set for every object reachable from *tl.
1728 * -------------------------------------------------------------------------- */
1729 static void
1730 retainRoot(void *user STG_UNUSED, StgClosure **tl)
1731 {
1732 StgClosure *c;
1733
1734 // We no longer assume that only TSOs and WEAKs are roots; any closure can
1735 // be a root.
1736
1737 ASSERT(isEmptyRetainerStack());
1738 currentStackBoundary = stackTop;
1739
1740 c = UNTAG_CLOSURE(*tl);
1741 if (c != &stg_END_TSO_QUEUE_closure && isRetainer(c)) {
1742 retainClosure(c, c, getRetainerFrom(c));
1743 } else {
1744 retainClosure(c, c, CCS_SYSTEM);
1745 }
1746
1747 // NOT TRUE: ASSERT(isMember(getRetainerFrom(*tl), retainerSetOf(*tl)));
1748 // *tl might be a TSO which is ThreadComplete, in which
1749 // case we ignore it for the purposes of retainer profiling.
1750 }
1751
1752 /* -----------------------------------------------------------------------------
1753 * Compute the retainer set for each of the objects in the heap.
1754 * -------------------------------------------------------------------------- */
1755 static void
1756 computeRetainerSet( void )
1757 {
1758 StgWeak *weak;
1759 RetainerSet *rtl;
1760 nat g;
1761 StgPtr ml;
1762 bdescr *bd;
1763 #ifdef DEBUG_RETAINER
1764 RetainerSet tmpRetainerSet;
1765 #endif
1766
1767 markCapabilities(retainRoot, NULL); // for scheduler roots
1768
1769 // This function is called after a major GC, when key, value, and finalizer
1770 // all are guaranteed to be valid, or reachable.
1771 //
1772 // The following code assumes that WEAK objects are considered to be roots
1773 // for retainer profilng.
1774 for (weak = weak_ptr_list; weak != NULL; weak = weak->link)
1775 // retainRoot((StgClosure *)weak);
1776 retainRoot(NULL, (StgClosure **)&weak);
1777
1778 // Consider roots from the stable ptr table.
1779 markStablePtrTable(retainRoot, NULL);
1780
1781 // The following code resets the rs field of each unvisited mutable
1782 // object (computing sumOfNewCostExtra and updating costArray[] when
1783 // debugging retainer profiler).
1784 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1785 // NOT TRUE: even G0 has a block on its mutable list
1786 // ASSERT(g != 0 || (generations[g].mut_list == NULL));
1787
1788 // Traversing through mut_list is necessary
1789 // because we can find MUT_VAR objects which have not been
1790 // visited during retainer profiling.
1791 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
1792 for (ml = bd->start; ml < bd->free; ml++) {
1793
1794 maybeInitRetainerSet((StgClosure *)*ml);
1795 rtl = retainerSetOf((StgClosure *)*ml);
1796
1797 #ifdef DEBUG_RETAINER
1798 if (rtl == NULL) {
1799 // first visit to *ml
1800 // This is a violation of the interface rule!
1801 RSET(ml) = (RetainerSet *)((StgWord)(&tmpRetainerSet) | flip);
1802
1803 switch (get_itbl((StgClosure *)ml)->type) {
1804 case IND_STATIC:
1805 // no cost involved
1806 break;
1807 case CONSTR_NOCAF_STATIC:
1808 case CONSTR_STATIC:
1809 case THUNK_STATIC:
1810 case FUN_STATIC:
1811 barf("Invalid object in computeRetainerSet(): %d", get_itbl((StgClosure*)ml)->type);
1812 break;
1813 default:
1814 // dynamic objects
1815 costArray[get_itbl((StgClosure *)ml)->type] += cost((StgClosure *)ml);
1816 sumOfNewCostExtra += cost((StgClosure *)ml);
1817 break;
1818 }
1819 }
1820 #endif
1821 }
1822 }
1823 }
1824 }
1825
1826 /* -----------------------------------------------------------------------------
1827 * Traverse all static objects for which we compute retainer sets,
1828 * and reset their rs fields to NULL, which is accomplished by
1829 * invoking maybeInitRetainerSet(). This function must be called
1830 * before zeroing all objects reachable from scavenged_static_objects
1831 * in the case of major gabage collections. See GarbageCollect() in
1832 * GC.c.
1833 * Note:
1834 * The mut_once_list of the oldest generation must also be traversed?
1835 * Why? Because if the evacuation of an object pointed to by a static
1836 * indirection object fails, it is put back to the mut_once_list of
1837 * the oldest generation.
1838 * However, this is not necessary because any static indirection objects
1839 * are just traversed through to reach dynamic objects. In other words,
1840 * they are not taken into consideration in computing retainer sets.
1841 * -------------------------------------------------------------------------- */
1842 void
1843 resetStaticObjectForRetainerProfiling( StgClosure *static_objects )
1844 {
1845 #ifdef DEBUG_RETAINER
1846 nat count;
1847 #endif
1848 StgClosure *p;
1849
1850 #ifdef DEBUG_RETAINER
1851 count = 0;
1852 #endif
1853 p = static_objects;
1854 while (p != END_OF_STATIC_LIST) {
1855 #ifdef DEBUG_RETAINER
1856 count++;
1857 #endif
1858 switch (get_itbl(p)->type) {
1859 case IND_STATIC:
1860 // Since we do not compute the retainer set of any
1861 // IND_STATIC object, we don't have to reset its retainer
1862 // field.
1863 p = (StgClosure*)*IND_STATIC_LINK(p);
1864 break;
1865 case THUNK_STATIC:
1866 maybeInitRetainerSet(p);
1867 p = (StgClosure*)*THUNK_STATIC_LINK(p);
1868 break;
1869 case FUN_STATIC:
1870 maybeInitRetainerSet(p);
1871 p = (StgClosure*)*FUN_STATIC_LINK(p);
1872 break;
1873 case CONSTR_STATIC:
1874 maybeInitRetainerSet(p);
1875 p = (StgClosure*)*STATIC_LINK(get_itbl(p), p);
1876 break;
1877 default:
1878 barf("resetStaticObjectForRetainerProfiling: %p (%s)",
1879 p, get_itbl(p)->type);
1880 break;
1881 }
1882 }
1883 #ifdef DEBUG_RETAINER
1884 // debugBelch("count in scavenged_static_objects = %d\n", count);
1885 #endif
1886 }
1887
1888 /* -----------------------------------------------------------------------------
1889 * Perform retainer profiling.
1890 * N is the oldest generation being profilied, where the generations are
1891 * numbered starting at 0.
1892 * Invariants:
1893 * Note:
1894 * This function should be called only immediately after major garbage
1895 * collection.
1896 * ------------------------------------------------------------------------- */
1897 void
1898 retainerProfile(void)
1899 {
1900 #ifdef DEBUG_RETAINER
1901 nat i;
1902 nat totalHeapSize; // total raw heap size (computed by linear scanning)
1903 #endif
1904
1905 #ifdef DEBUG_RETAINER
1906 debugBelch(" < retainerProfile() invoked : %d>\n", retainerGeneration);
1907 #endif
1908
1909 stat_startRP();
1910
1911 // We haven't flipped the bit yet.
1912 #ifdef DEBUG_RETAINER
1913 debugBelch("Before traversing:\n");
1914 sumOfCostLinear = 0;
1915 for (i = 0;i < N_CLOSURE_TYPES; i++)
1916 costArrayLinear[i] = 0;
1917 totalHeapSize = checkHeapSanityForRetainerProfiling();
1918
1919 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
1920 /*
1921 debugBelch("costArrayLinear[] = ");
1922 for (i = 0;i < N_CLOSURE_TYPES; i++)
1923 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
1924 debugBelch("\n");
1925 */
1926
1927 ASSERT(sumOfCostLinear == totalHeapSize);
1928
1929 /*
1930 #define pcostArrayLinear(index) \
1931 if (costArrayLinear[index] > 0) \
1932 debugBelch("costArrayLinear[" #index "] = %u\n", costArrayLinear[index])
1933 pcostArrayLinear(THUNK_STATIC);
1934 pcostArrayLinear(FUN_STATIC);
1935 pcostArrayLinear(CONSTR_STATIC);
1936 pcostArrayLinear(CONSTR_NOCAF_STATIC);
1937 */
1938 #endif
1939
1940 // Now we flips flip.
1941 flip = flip ^ 1;
1942
1943 #ifdef DEBUG_RETAINER
1944 stackSize = 0;
1945 maxStackSize = 0;
1946 cStackSize = 0;
1947 maxCStackSize = 0;
1948 #endif
1949 numObjectVisited = 0;
1950 timesAnyObjectVisited = 0;
1951
1952 #ifdef DEBUG_RETAINER
1953 debugBelch("During traversing:\n");
1954 sumOfNewCost = 0;
1955 sumOfNewCostExtra = 0;
1956 for (i = 0;i < N_CLOSURE_TYPES; i++)
1957 costArray[i] = 0;
1958 #endif
1959
1960 /*
1961 We initialize the traverse stack each time the retainer profiling is
1962 performed (because the traverse stack size varies on each retainer profiling
1963 and this operation is not costly anyhow). However, we just refresh the
1964 retainer sets.
1965 */
1966 initializeTraverseStack();
1967 #ifdef DEBUG_RETAINER
1968 initializeAllRetainerSet();
1969 #else
1970 refreshAllRetainerSet();
1971 #endif
1972 computeRetainerSet();
1973
1974 #ifdef DEBUG_RETAINER
1975 debugBelch("After traversing:\n");
1976 sumOfCostLinear = 0;
1977 for (i = 0;i < N_CLOSURE_TYPES; i++)
1978 costArrayLinear[i] = 0;
1979 totalHeapSize = checkHeapSanityForRetainerProfiling();
1980
1981 debugBelch("\tsumOfCostLinear = %d, totalHeapSize = %d\n", sumOfCostLinear, totalHeapSize);
1982 ASSERT(sumOfCostLinear == totalHeapSize);
1983
1984 // now, compare the two results
1985 /*
1986 Note:
1987 costArray[] must be exactly the same as costArrayLinear[].
1988 Known exceptions:
1989 1) Dead weak pointers, whose type is CONSTR. These objects are not
1990 reachable from any roots.
1991 */
1992 debugBelch("Comparison:\n");
1993 debugBelch("\tcostArrayLinear[] (must be empty) = ");
1994 for (i = 0;i < N_CLOSURE_TYPES; i++)
1995 if (costArray[i] != costArrayLinear[i])
1996 // nothing should be printed except MUT_VAR after major GCs
1997 debugBelch("[%u:%u] ", i, costArrayLinear[i]);
1998 debugBelch("\n");
1999
2000 debugBelch("\tsumOfNewCost = %u\n", sumOfNewCost);
2001 debugBelch("\tsumOfNewCostExtra = %u\n", sumOfNewCostExtra);
2002 debugBelch("\tcostArray[] (must be empty) = ");
2003 for (i = 0;i < N_CLOSURE_TYPES; i++)
2004 if (costArray[i] != costArrayLinear[i])
2005 // nothing should be printed except MUT_VAR after major GCs
2006 debugBelch("[%u:%u] ", i, costArray[i]);
2007 debugBelch("\n");
2008
2009 // only for major garbage collection
2010 ASSERT(sumOfNewCost + sumOfNewCostExtra == sumOfCostLinear);
2011 #endif
2012
2013 // post-processing
2014 closeTraverseStack();
2015 #ifdef DEBUG_RETAINER
2016 closeAllRetainerSet();
2017 #else
2018 // Note that there is no post-processing for the retainer sets.
2019 #endif
2020 retainerGeneration++;
2021
2022 stat_endRP(
2023 retainerGeneration - 1, // retainerGeneration has just been incremented!
2024 #ifdef DEBUG_RETAINER
2025 maxCStackSize, maxStackSize,
2026 #endif
2027 (double)timesAnyObjectVisited / numObjectVisited);
2028 }
2029
2030 /* -----------------------------------------------------------------------------
2031 * DEBUGGING CODE
2032 * -------------------------------------------------------------------------- */
2033
2034 #ifdef DEBUG_RETAINER
2035
2036 #define LOOKS_LIKE_PTR(r) ((LOOKS_LIKE_STATIC_CLOSURE(r) || \
2037 ((HEAP_ALLOCED(r) && ((Bdescr((P_)r)->flags & BF_FREE) == 0)))) && \
2038 ((StgWord)(*(StgPtr)r)!=0xaaaaaaaa))
2039
2040 static nat
2041 sanityCheckHeapClosure( StgClosure *c )
2042 {
2043 StgInfoTable *info;
2044
2045 ASSERT(LOOKS_LIKE_GHC_INFO(c->header.info));
2046 ASSERT(!closure_STATIC(c));
2047 ASSERT(LOOKS_LIKE_PTR(c));
2048
2049 if ((((StgWord)RSET(c) & 1) ^ flip) != 0) {
2050 if (get_itbl(c)->type == CONSTR &&
2051 !strcmp(GET_PROF_TYPE(get_itbl(c)), "DEAD_WEAK") &&
2052 !strcmp(GET_PROF_DESC(get_itbl(c)), "DEAD_WEAK")) {
2053 debugBelch("\tUnvisited dead weak pointer object found: c = %p\n", c);
2054 costArray[get_itbl(c)->type] += cost(c);
2055 sumOfNewCost += cost(c);
2056 } else
2057 debugBelch(
2058 "Unvisited object: flip = %d, c = %p(%d, %s, %s), rs = %p\n",
2059 flip, c, get_itbl(c)->type,
2060 get_itbl(c)->prof.closure_type, GET_PROF_DESC(get_itbl(c)),
2061 RSET(c));
2062 } else {
2063 // debugBelch("sanityCheckHeapClosure) S: flip = %d, c = %p(%d), rs = %p\n", flip, c, get_itbl(c)->type, RSET(c));
2064 }
2065
2066 return closure_sizeW(c);
2067 }
2068
2069 static nat
2070 heapCheck( bdescr *bd )
2071 {
2072 StgPtr p;
2073 static nat costSum, size;
2074
2075 costSum = 0;
2076 while (bd != NULL) {
2077 p = bd->start;
2078 while (p < bd->free) {
2079 size = sanityCheckHeapClosure((StgClosure *)p);
2080 sumOfCostLinear += size;
2081 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2082 p += size;
2083 // no need for slop check; I think slops are not used currently.
2084 }
2085 ASSERT(p == bd->free);
2086 costSum += bd->free - bd->start;
2087 bd = bd->link;
2088 }
2089
2090 return costSum;
2091 }
2092
2093 static nat
2094 smallObjectPoolCheck(void)
2095 {
2096 bdescr *bd;
2097 StgPtr p;
2098 static nat costSum, size;
2099
2100 bd = g0s0->blocks;
2101 costSum = 0;
2102
2103 // first block
2104 if (bd == NULL)
2105 return costSum;
2106
2107 p = bd->start;
2108 while (p < alloc_Hp) {
2109 size = sanityCheckHeapClosure((StgClosure *)p);
2110 sumOfCostLinear += size;
2111 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2112 p += size;
2113 }
2114 ASSERT(p == alloc_Hp);
2115 costSum += alloc_Hp - bd->start;
2116
2117 bd = bd->link;
2118 while (bd != NULL) {
2119 p = bd->start;
2120 while (p < bd->free) {
2121 size = sanityCheckHeapClosure((StgClosure *)p);
2122 sumOfCostLinear += size;
2123 costArrayLinear[get_itbl((StgClosure *)p)->type] += size;
2124 p += size;
2125 }
2126 ASSERT(p == bd->free);
2127 costSum += bd->free - bd->start;
2128 bd = bd->link;
2129 }
2130
2131 return costSum;
2132 }
2133
2134 static nat
2135 chainCheck(bdescr *bd)
2136 {
2137 nat costSum, size;
2138
2139 costSum = 0;
2140 while (bd != NULL) {
2141 // bd->free - bd->start is not an accurate measurement of the
2142 // object size. Actually it is always zero, so we compute its
2143 // size explicitly.
2144 size = sanityCheckHeapClosure((StgClosure *)bd->start);
2145 sumOfCostLinear += size;
2146 costArrayLinear[get_itbl((StgClosure *)bd->start)->type] += size;
2147 costSum += size;
2148 bd = bd->link;
2149 }
2150
2151 return costSum;
2152 }
2153
2154 static nat
2155 checkHeapSanityForRetainerProfiling( void )
2156 {
2157 nat costSum, g, s;
2158
2159 costSum = 0;
2160 debugBelch("START: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2161 if (RtsFlags.GcFlags.generations == 1) {
2162 costSum += heapCheck(g0s0->to_blocks);
2163 debugBelch("heapCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2164 costSum += chainCheck(g0s0->large_objects);
2165 debugBelch("chainCheck: sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2166 } else {
2167 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
2168 for (s = 0; s < generations[g].n_steps; s++) {
2169 /*
2170 After all live objects have been scavenged, the garbage
2171 collector may create some objects in
2172 scheduleFinalizers(). These objects are created throught
2173 allocate(), so the small object pool or the large object
2174 pool of the g0s0 may not be empty.
2175 */
2176 if (g == 0 && s == 0) {
2177 costSum += smallObjectPoolCheck();
2178 debugBelch("smallObjectPoolCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2179 costSum += chainCheck(generations[g].steps[s].large_objects);
2180 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2181 } else {
2182 costSum += heapCheck(generations[g].steps[s].blocks);
2183 debugBelch("heapCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2184 costSum += chainCheck(generations[g].steps[s].large_objects);
2185 debugBelch("chainCheck(): sumOfCostLinear = %d, costSum = %d\n", sumOfCostLinear, costSum);
2186 }
2187 }
2188 }
2189
2190 return costSum;
2191 }
2192
2193 void
2194 findPointer(StgPtr p)
2195 {
2196 StgPtr q, r, e;
2197 bdescr *bd;
2198 nat g, s;
2199
2200 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2201 for (s = 0; s < generations[g].n_steps; s++) {
2202 // if (g == 0 && s == 0) continue;
2203 bd = generations[g].steps[s].blocks;
2204 for (; bd; bd = bd->link) {
2205 for (q = bd->start; q < bd->free; q++) {
2206 if (*q == (StgWord)p) {
2207 r = q;
2208 while (!LOOKS_LIKE_GHC_INFO(*r)) r--;
2209 debugBelch("Found in gen[%d], step[%d]: q = %p, r = %p\n", g, s, q, r);
2210 // return;
2211 }
2212 }
2213 }
2214 bd = generations[g].steps[s].large_objects;
2215 for (; bd; bd = bd->link) {
2216 e = bd->start + cost((StgClosure *)bd->start);
2217 for (q = bd->start; q < e; q++) {
2218 if (*q == (StgWord)p) {
2219 r = q;
2220 while (*r == 0 || !LOOKS_LIKE_GHC_INFO(*r)) r--;
2221 debugBelch("Found in gen[%d], large_objects: %p\n", g, r);
2222 // return;
2223 }
2224 }
2225 }
2226 }
2227 }
2228 }
2229
2230 static void
2231 belongToHeap(StgPtr p)
2232 {
2233 bdescr *bd;
2234 nat g, s;
2235
2236 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
2237 for (s = 0; s < generations[g].n_steps; s++) {
2238 // if (g == 0 && s == 0) continue;
2239 bd = generations[g].steps[s].blocks;
2240 for (; bd; bd = bd->link) {
2241 if (bd->start <= p && p < bd->free) {
2242 debugBelch("Belongs to gen[%d], step[%d]", g, s);
2243 return;
2244 }
2245 }
2246 bd = generations[g].steps[s].large_objects;
2247 for (; bd; bd = bd->link) {
2248 if (bd->start <= p && p < bd->start + getHeapClosureSize((StgClosure *)bd->start)) {
2249 debugBelch("Found in gen[%d], large_objects: %p\n", g, bd->start);
2250 return;
2251 }
2252 }
2253 }
2254 }
2255 }
2256 #endif /* DEBUG_RETAINER */
2257
2258 #endif /* PROFILING */