1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 2001
5 * Compacting garbage collector
7 * ---------------------------------------------------------------------------*/
9 #include "PosixSource.h"
13 #include "OSThreads.h"
15 #include "BlockAlloc.h"
17 #include "GCCompact.h"
21 // Turn off inlining when debugging - it obfuscates things
24 # define STATIC_INLINE static
27 /* -----------------------------------------------------------------------------
28 Threading / unthreading pointers.
30 The basic idea here is to chain together all the fields pointing at
31 a particular object, with the root of the chain in the object's
32 info table field. The original contents of the info pointer goes
33 at the end of the chain.
35 Adding a new field to the chain is a matter of swapping the
36 contents of the field with the contents of the object's info table
39 To unthread the chain, we walk down it updating all the fields on
40 the chain with the new location of the object. We stop when we
41 reach the info pointer at the end.
43 We use a trick to identify the info pointer: when swapping pointers
44 for threading, we set the low bit of the original pointer, with the
45 result that all the pointers in the chain have their low bits set
46 except for the info pointer.
47 -------------------------------------------------------------------------- */
52 StgPtr q = (StgPtr)*p;
55 // It doesn't look like a closure at the moment, because the info
56 // ptr is possibly threaded:
57 // ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
59 if (HEAP_ALLOCED(q)) {
61 // a handy way to discover whether the ptr is into the
62 // compacted area of the old gen, is that the EVACUATED flag
63 // is zero (it's non-zero for all the other areas of live
65 if ((bd->flags & BF_EVACUATED) == 0) {
67 *q = (StgWord)p + 1; // set the low bit
73 unthread( StgPtr p, StgPtr free )
77 while ((q & 1) != 0) {
78 q -= 1; // unset the low bit again
80 *((StgPtr)q) = (StgWord)free;
86 STATIC_INLINE StgInfoTable *
87 get_threaded_info( StgPtr p )
89 StgPtr q = (P_)GET_INFO((StgClosure *)p);
91 while (((StgWord)q & 1) != 0) {
92 q = (P_)*((StgPtr)((StgWord)q-1));
95 ASSERT(LOOKS_LIKE_INFO_PTR(q));
96 return INFO_PTR_TO_STRUCT((StgInfoTable *)q);
99 // A word-aligned memmove will be faster for small objects than libc's or gcc's.
100 // Remember, the two regions *might* overlap, but: to <= from.
102 move(StgPtr to, StgPtr from, nat size)
104 for(; size > 0; --size) {
110 obj_sizeW( StgClosure *p, StgInfoTable *info )
112 switch (info->type) {
115 return sizeofW(StgThunk) + 1;
120 return sizeofW(StgHeader) + 1;
124 return sizeofW(StgThunk) + 2;
131 return sizeofW(StgHeader) + 2;
133 return THUNK_SELECTOR_sizeW();
135 return ap_stack_sizeW((StgAP_STACK *)p);
138 return pap_sizeW((StgPAP *)p);
140 return arr_words_sizeW((StgArrWords *)p);
142 case MUT_ARR_PTRS_FROZEN:
143 case MUT_ARR_PTRS_FROZEN0:
144 return mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
146 return tso_sizeW((StgTSO *)p);
148 return bco_sizeW((StgBCO *)p);
149 case TVAR_WAIT_QUEUE:
150 return sizeofW(StgTVarWaitQueue);
152 return sizeofW(StgTVar);
154 return sizeofW(StgTRecChunk);
156 return sizeofW(StgTRecHeader);
158 return sizeW_fromITBL(info);
163 thread_static( StgClosure* p )
165 const StgInfoTable *info;
167 // keep going until we've threaded all the objects on the linked
169 while (p != END_OF_STATIC_LIST) {
172 switch (info->type) {
175 thread((StgPtr)&((StgInd *)p)->indirectee);
176 p = *IND_STATIC_LINK(p);
180 p = *THUNK_STATIC_LINK(p);
183 p = *FUN_STATIC_LINK(p);
186 p = *STATIC_LINK(info,p);
190 barf("thread_static: strange closure %d", (int)(info->type));
197 thread_large_bitmap( StgPtr p, StgLargeBitmap *large_bitmap, nat size )
203 bitmap = large_bitmap->bitmap[b];
204 for (i = 0; i < size; ) {
205 if ((bitmap & 1) == 0) {
210 if (i % BITS_IN(W_) == 0) {
212 bitmap = large_bitmap->bitmap[b];
214 bitmap = bitmap >> 1;
220 thread_arg_block (StgFunInfoTable *fun_info, StgClosure **args)
227 switch (fun_info->f.fun_type) {
229 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
230 size = BITMAP_SIZE(fun_info->f.b.bitmap);
233 size = GET_FUN_LARGE_BITMAP(fun_info)->size;
234 thread_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info), size);
238 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
239 size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
242 if ((bitmap & 1) == 0) {
246 bitmap = bitmap >> 1;
255 thread_stack(StgPtr p, StgPtr stack_end)
257 const StgRetInfoTable* info;
261 // highly similar to scavenge_stack, but we do pointer threading here.
263 while (p < stack_end) {
265 // *p must be the info pointer of an activation
266 // record. All activation records have 'bitmap' style layout
269 info = get_ret_itbl((StgClosure *)p);
271 switch (info->i.type) {
273 // Dynamic bitmap: the mask is stored on the stack
277 dyn = ((StgRetDyn *)p)->liveness;
279 // traverse the bitmap first
280 bitmap = RET_DYN_LIVENESS(dyn);
281 p = (P_)&((StgRetDyn *)p)->payload[0];
282 size = RET_DYN_BITMAP_SIZE;
284 if ((bitmap & 1) == 0) {
288 bitmap = bitmap >> 1;
292 // skip over the non-ptr words
293 p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
295 // follow the ptr words
296 for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
303 // small bitmap (<= 32 entries, or 64 on a 64-bit machine)
304 case CATCH_RETRY_FRAME:
305 case CATCH_STM_FRAME:
306 case ATOMICALLY_FRAME:
312 bitmap = BITMAP_BITS(info->i.layout.bitmap);
313 size = BITMAP_SIZE(info->i.layout.bitmap);
315 // NOTE: the payload starts immediately after the info-ptr, we
316 // don't have an StgHeader in the same sense as a heap closure.
318 if ((bitmap & 1) == 0) {
322 bitmap = bitmap >> 1;
335 size = BCO_BITMAP_SIZE(bco);
336 thread_large_bitmap(p, BCO_BITMAP(bco), size);
341 // large bitmap (> 32 entries, or 64 on a 64-bit machine)
345 size = GET_LARGE_BITMAP(&info->i)->size;
346 thread_large_bitmap(p, GET_LARGE_BITMAP(&info->i), size);
352 StgRetFun *ret_fun = (StgRetFun *)p;
353 StgFunInfoTable *fun_info;
355 fun_info = itbl_to_fun_itbl(
356 get_threaded_info((StgPtr)ret_fun->fun));
357 // *before* threading it!
358 thread((StgPtr)&ret_fun->fun);
359 p = thread_arg_block(fun_info, ret_fun->payload);
364 barf("thread_stack: weird activation record found on stack: %d",
365 (int)(info->i.type));
371 thread_PAP_payload (StgClosure *fun, StgClosure **payload, StgWord size)
375 StgFunInfoTable *fun_info;
377 fun_info = itbl_to_fun_itbl(get_threaded_info((StgPtr)fun));
378 ASSERT(fun_info->i.type != PAP);
382 switch (fun_info->f.fun_type) {
384 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
387 thread_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info), size);
391 thread_large_bitmap((StgPtr)payload, BCO_BITMAP(fun), size);
395 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
398 if ((bitmap & 1) == 0) {
402 bitmap = bitmap >> 1;
412 thread_PAP (StgPAP *pap)
415 p = thread_PAP_payload(pap->fun, pap->payload, pap->n_args);
416 thread((StgPtr)&pap->fun);
421 thread_AP (StgAP *ap)
424 p = thread_PAP_payload(ap->fun, ap->payload, ap->n_args);
425 thread((StgPtr)&ap->fun);
430 thread_AP_STACK (StgAP_STACK *ap)
432 thread((StgPtr)&ap->fun);
433 thread_stack((P_)ap->payload, (P_)ap->payload + ap->size);
434 return (P_)ap + sizeofW(StgAP_STACK) + ap->size;
438 thread_TSO (StgTSO *tso)
440 thread((StgPtr)&tso->link);
441 thread((StgPtr)&tso->global_link);
443 if ( tso->why_blocked == BlockedOnMVar
444 || tso->why_blocked == BlockedOnBlackHole
445 || tso->why_blocked == BlockedOnException
447 || tso->why_blocked == BlockedOnGA
448 || tso->why_blocked == BlockedOnGA_NoSend
451 thread((StgPtr)&tso->block_info.closure);
453 if ( tso->blocked_exceptions != NULL ) {
454 thread((StgPtr)&tso->blocked_exceptions);
457 thread((StgPtr)&tso->trec);
459 thread_stack(tso->sp, &(tso->stack[tso->stack_size]));
460 return (StgPtr)tso + tso_sizeW(tso);
465 update_fwd_large( bdescr *bd )
468 const StgInfoTable* info;
470 for (; bd != NULL; bd = bd->link) {
473 info = get_itbl((StgClosure *)p);
475 switch (info->type) {
482 case MUT_ARR_PTRS_FROZEN:
483 case MUT_ARR_PTRS_FROZEN0:
488 next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
489 for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
496 thread_TSO((StgTSO *)p);
500 thread_AP_STACK((StgAP_STACK *)p);
504 thread_PAP((StgPAP *)p);
508 barf("update_fwd_large: unknown/strange object %d", (int)(info->type));
514 thread_obj (StgInfoTable *info, StgPtr p)
516 switch (info->type) {
518 return p + sizeofW(StgThunk) + 1;
522 return p + sizeofW(StgHeader) + 1;
526 thread((StgPtr)&((StgClosure *)p)->payload[0]);
527 return p + sizeofW(StgHeader) + 1;
530 thread((StgPtr)&((StgThunk *)p)->payload[0]);
531 return p + sizeofW(StgThunk) + 1;
534 return p + sizeofW(StgThunk) + 2;
538 return p + sizeofW(StgHeader) + 2;
541 thread((StgPtr)&((StgThunk *)p)->payload[0]);
542 return p + sizeofW(StgThunk) + 2;
546 thread((StgPtr)&((StgClosure *)p)->payload[0]);
547 return p + sizeofW(StgHeader) + 2;
550 thread((StgPtr)&((StgThunk *)p)->payload[0]);
551 thread((StgPtr)&((StgThunk *)p)->payload[1]);
552 return p + sizeofW(StgThunk) + 2;
556 thread((StgPtr)&((StgClosure *)p)->payload[0]);
557 thread((StgPtr)&((StgClosure *)p)->payload[1]);
558 return p + sizeofW(StgHeader) + 2;
561 StgBCO *bco = (StgBCO *)p;
562 thread((StgPtr)&bco->instrs);
563 thread((StgPtr)&bco->literals);
564 thread((StgPtr)&bco->ptrs);
565 thread((StgPtr)&bco->itbls);
566 return p + bco_sizeW(bco);
573 end = (P_)((StgThunk *)p)->payload +
574 info->layout.payload.ptrs;
575 for (p = (P_)((StgThunk *)p)->payload; p < end; p++) {
578 return p + info->layout.payload.nptrs;
588 case SE_CAF_BLACKHOLE:
594 end = (P_)((StgClosure *)p)->payload +
595 info->layout.payload.ptrs;
596 for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
599 return p + info->layout.payload.nptrs;
604 StgWeak *w = (StgWeak *)p;
605 thread((StgPtr)&w->key);
606 thread((StgPtr)&w->value);
607 thread((StgPtr)&w->finalizer);
608 if (w->link != NULL) {
609 thread((StgPtr)&w->link);
611 return p + sizeofW(StgWeak);
616 StgMVar *mvar = (StgMVar *)p;
617 thread((StgPtr)&mvar->head);
618 thread((StgPtr)&mvar->tail);
619 thread((StgPtr)&mvar->value);
620 return p + sizeofW(StgMVar);
624 case IND_OLDGEN_PERM:
625 thread((StgPtr)&((StgInd *)p)->indirectee);
626 return p + sizeofW(StgInd);
630 StgSelector *s = (StgSelector *)p;
631 thread((StgPtr)&s->selectee);
632 return p + THUNK_SELECTOR_sizeW();
636 return thread_AP_STACK((StgAP_STACK *)p);
639 return thread_PAP((StgPAP *)p);
642 return thread_AP((StgAP *)p);
645 return p + arr_words_sizeW((StgArrWords *)p);
648 case MUT_ARR_PTRS_FROZEN:
649 case MUT_ARR_PTRS_FROZEN0:
654 next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
655 for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
662 return thread_TSO((StgTSO *)p);
664 case TVAR_WAIT_QUEUE:
666 StgTVarWaitQueue *wq = (StgTVarWaitQueue *)p;
667 thread((StgPtr)&wq->waiting_tso);
668 thread((StgPtr)&wq->next_queue_entry);
669 thread((StgPtr)&wq->prev_queue_entry);
670 return p + sizeofW(StgTVarWaitQueue);
675 StgTVar *tvar = (StgTVar *)p;
676 thread((StgPtr)&tvar->current_value);
677 thread((StgPtr)&tvar->first_wait_queue_entry);
679 thread((StgPtr)&tvar->last_update_by);
681 return p + sizeofW(StgTVar);
686 StgTRecHeader *trec = (StgTRecHeader *)p;
687 thread((StgPtr)&trec->enclosing_trec);
688 thread((StgPtr)&trec->current_chunk);
689 return p + sizeofW(StgTRecHeader);
695 StgTRecChunk *tc = (StgTRecChunk *)p;
696 TRecEntry *e = &(tc -> entries[0]);
697 thread((StgPtr)&tc->prev_chunk);
698 for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
699 thread((StgPtr)&e->tvar);
700 thread((StgPtr)&e->expected_value);
701 thread((StgPtr)&e->new_value);
703 return p + sizeofW(StgTRecChunk);
707 barf("update_fwd: unknown/strange object %d", (int)(info->type));
713 update_fwd( bdescr *blocks )
722 barf("update_fwd: ToDo");
725 // cycle through all the blocks in the step
726 for (; bd != NULL; bd = bd->link) {
729 // linearly scan the objects in this block
730 while (p < bd->free) {
731 ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
732 info = get_itbl((StgClosure *)p);
733 p = thread_obj(info, p);
739 update_fwd_compact( bdescr *blocks )
745 bdescr *bd, *free_bd;
751 free = free_bd->start;
754 barf("update_fwd: ToDo");
757 // cycle through all the blocks in the step
758 for (; bd != NULL; bd = bd->link) {
761 while (p < bd->free ) {
763 while ( p < bd->free && !is_marked(p,bd) ) {
772 m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
773 m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));
775 while ( p < bd->free ) {
780 if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
788 // Problem: we need to know the destination for this cell
789 // in order to unthread its info pointer. But we can't
790 // know the destination without the size, because we may
791 // spill into the next block. So we have to run down the
792 // threaded list and get the info ptr first.
793 info = get_threaded_info(p);
797 p = thread_obj(info, p);
800 if (free + size > free_bd->start + BLOCK_SIZE_W) {
801 // unset the next bit in the bitmap to indicate that
802 // this object needs to be pushed into the next
803 // block. This saves us having to run down the
804 // threaded info pointer list twice during the next pass.
806 free_bd = free_bd->link;
807 free = free_bd->start;
809 ASSERT(is_marked(q+1,bd));
822 update_bkwd_compact( step *stp )
828 bdescr *bd, *free_bd;
830 nat size, free_blocks;
832 bd = free_bd = stp->blocks;
833 free = free_bd->start;
837 barf("update_bkwd: ToDo");
840 // cycle through all the blocks in the step
841 for (; bd != NULL; bd = bd->link) {
844 while (p < bd->free ) {
846 while ( p < bd->free && !is_marked(p,bd) ) {
855 m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
856 m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));
858 while ( p < bd->free ) {
863 if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
871 if (!is_marked(p+1,bd)) {
872 // don't forget to update the free ptr in the block desc.
873 free_bd->free = free;
874 free_bd = free_bd->link;
875 free = free_bd->start;
880 ASSERT(LOOKS_LIKE_INFO_PTR(((StgClosure *)p)->header.info));
881 info = get_itbl((StgClosure *)p);
882 size = obj_sizeW((StgClosure *)p,info);
889 if (info->type == TSO) {
890 move_TSO((StgTSO *)p, (StgTSO *)free);
901 // free the remaining blocks and count what's left.
902 free_bd->free = free;
903 if (free_bd->link != NULL) {
904 freeChain(free_bd->link);
905 free_bd->link = NULL;
907 stp->n_blocks = free_blocks;
913 compact( void (*get_roots)(evac_fn) )
918 // 1. thread the roots
919 get_roots((evac_fn)thread);
921 // the weak pointer lists...
922 if (weak_ptr_list != NULL) {
923 thread((StgPtr)&weak_ptr_list);
925 if (old_weak_ptr_list != NULL) {
926 thread((StgPtr)&old_weak_ptr_list); // tmp
930 for (g = 1; g < RtsFlags.GcFlags.generations; g++) {
933 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
934 for (p = bd->start; p < bd->free; p++) {
940 // the global thread list
941 thread((StgPtr)&all_threads);
943 // any threads resurrected during this GC
944 thread((StgPtr)&resurrected_threads);
946 // the main threads list
949 for (m = main_threads; m != NULL; m = m->link) {
950 thread((StgPtr)&m->tso);
954 // the static objects
955 thread_static(scavenged_static_objects);
957 // the stable pointer table
958 threadStablePtrTable((evac_fn)thread);
960 // the CAF list (used by GHCi)
961 markCAFs((evac_fn)thread);
963 // 2. update forward ptrs
964 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
965 for (s = 0; s < generations[g].n_steps; s++) {
966 stp = &generations[g].steps[s];
967 IF_DEBUG(gc, debugBelch("update_fwd: %d.%d\n", stp->gen->no, stp->no););
969 update_fwd(stp->to_blocks);
970 update_fwd_large(stp->scavenged_large_objects);
971 if (g == RtsFlags.GcFlags.generations-1 && stp->blocks != NULL) {
972 IF_DEBUG(gc, debugBelch("update_fwd: %d.%d (compact)\n", stp->gen->no, stp->no););
973 update_fwd_compact(stp->blocks);
978 // 3. update backward ptrs
979 stp = &oldest_gen->steps[0];
980 if (stp->blocks != NULL) {
981 blocks = update_bkwd_compact(stp);
982 IF_DEBUG(gc, debugBelch("update_bkwd: %d.%d (compact, old: %d blocks, now %d blocks)\n",
983 stp->gen->no, stp->no,
984 stp->n_blocks, blocks););
985 stp->n_blocks = blocks;