3212ce5852737cbf8e5703630294786c8ac86a14
[ghc.git] / rts / sm / Evac.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team 1998-2008
4 *
5 * Generational garbage collector: evacuation functions
6 *
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
9 *
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
11 *
12 * ---------------------------------------------------------------------------*/
13
14 #include "PosixSource.h"
15 #include "Rts.h"
16
17 #include "Evac.h"
18 #include "Storage.h"
19 #include "GC.h"
20 #include "GCThread.h"
21 #include "GCUtils.h"
22 #include "Compact.h"
23 #include "MarkStack.h"
24 #include "Prelude.h"
25 #include "Trace.h"
26 #include "LdvProfile.h"
27
28 #if defined(PROF_SPIN) && defined(THREADED_RTS) && defined(PARALLEL_GC)
29 StgWord64 whitehole_spin = 0;
30 #endif
31
32 #if defined(THREADED_RTS) && !defined(PARALLEL_GC)
33 #define evacuate(p) evacuate1(p)
34 #define HEAP_ALLOCED_GC(p) HEAP_ALLOCED(p)
35 #endif
36
37 #if !defined(PARALLEL_GC)
38 #define copy_tag_nolock(p, info, src, size, stp, tag) \
39 copy_tag(p, info, src, size, stp, tag)
40 #endif
41
42 /* Used to avoid long recursion due to selector thunks
43 */
44 #define MAX_THUNK_SELECTOR_DEPTH 16
45
46 static void eval_thunk_selector (StgClosure **q, StgSelector * p, rtsBool);
47 STATIC_INLINE void evacuate_large(StgPtr p);
48
49 /* -----------------------------------------------------------------------------
50 Allocate some space in which to copy an object.
51 -------------------------------------------------------------------------- */
52
53 STATIC_INLINE StgPtr
54 alloc_for_copy (nat size, step *stp)
55 {
56 StgPtr to;
57 step_workspace *ws;
58
59 /* Find out where we're going, using the handy "to" pointer in
60 * the step of the source object. If it turns out we need to
61 * evacuate to an older generation, adjust it here (see comment
62 * by evacuate()).
63 */
64 if (stp < gct->evac_step) {
65 if (gct->eager_promotion) {
66 stp = gct->evac_step;
67 } else {
68 gct->failed_to_evac = rtsTrue;
69 }
70 }
71
72 ws = &gct->steps[stp->abs_no];
73 // this compiles to a single mem access to stp->abs_no only
74
75 /* chain a new block onto the to-space for the destination step if
76 * necessary.
77 */
78 to = ws->todo_free;
79 ws->todo_free += size;
80 if (ws->todo_free > ws->todo_lim) {
81 to = todo_block_full(size, ws);
82 }
83 ASSERT(ws->todo_free >= ws->todo_bd->free && ws->todo_free <= ws->todo_lim);
84
85 return to;
86 }
87
88 /* -----------------------------------------------------------------------------
89 The evacuate() code
90 -------------------------------------------------------------------------- */
91
92 STATIC_INLINE GNUC_ATTR_HOT void
93 copy_tag(StgClosure **p, const StgInfoTable *info,
94 StgClosure *src, nat size, step *stp, StgWord tag)
95 {
96 StgPtr to, from;
97 nat i;
98
99 to = alloc_for_copy(size,stp);
100
101 from = (StgPtr)src;
102 to[0] = (W_)info;
103 for (i = 1; i < size; i++) { // unroll for small i
104 to[i] = from[i];
105 }
106
107 // if (to+size+2 < bd->start + BLOCK_SIZE_W) {
108 // __builtin_prefetch(to + size + 2, 1);
109 // }
110
111 #if defined(PARALLEL_GC)
112 {
113 const StgInfoTable *new_info;
114 new_info = (const StgInfoTable *)cas((StgPtr)&src->header.info, (W_)info, MK_FORWARDING_PTR(to));
115 if (new_info != info) {
116 return evacuate(p); // does the failed_to_evac stuff
117 } else {
118 *p = TAG_CLOSURE(tag,(StgClosure*)to);
119 }
120 }
121 #else
122 src->header.info = (const StgInfoTable *)MK_FORWARDING_PTR(to);
123 *p = TAG_CLOSURE(tag,(StgClosure*)to);
124 #endif
125
126 #ifdef PROFILING
127 // We store the size of the just evacuated object in the LDV word so that
128 // the profiler can guess the position of the next object later.
129 SET_EVACUAEE_FOR_LDV(from, size);
130 #endif
131 }
132
133 #if defined(PARALLEL_GC)
134 STATIC_INLINE void
135 copy_tag_nolock(StgClosure **p, const StgInfoTable *info,
136 StgClosure *src, nat size, step *stp, StgWord tag)
137 {
138 StgPtr to, from;
139 nat i;
140
141 to = alloc_for_copy(size,stp);
142 *p = TAG_CLOSURE(tag,(StgClosure*)to);
143 src->header.info = (const StgInfoTable *)MK_FORWARDING_PTR(to);
144
145 from = (StgPtr)src;
146 to[0] = (W_)info;
147 for (i = 1; i < size; i++) { // unroll for small i
148 to[i] = from[i];
149 }
150
151 // if (to+size+2 < bd->start + BLOCK_SIZE_W) {
152 // __builtin_prefetch(to + size + 2, 1);
153 // }
154
155 #ifdef PROFILING
156 // We store the size of the just evacuated object in the LDV word so that
157 // the profiler can guess the position of the next object later.
158 SET_EVACUAEE_FOR_LDV(from, size);
159 #endif
160 }
161 #endif
162
163 /* Special version of copy() for when we only want to copy the info
164 * pointer of an object, but reserve some padding after it. This is
165 * used to optimise evacuation of BLACKHOLEs.
166 */
167 static rtsBool
168 copyPart(StgClosure **p, StgClosure *src, nat size_to_reserve, nat size_to_copy, step *stp)
169 {
170 StgPtr to, from;
171 nat i;
172 StgWord info;
173
174 #if defined(PARALLEL_GC)
175 spin:
176 info = xchg((StgPtr)&src->header.info, (W_)&stg_WHITEHOLE_info);
177 if (info == (W_)&stg_WHITEHOLE_info) {
178 #ifdef PROF_SPIN
179 whitehole_spin++;
180 #endif
181 goto spin;
182 }
183 if (IS_FORWARDING_PTR(info)) {
184 src->header.info = (const StgInfoTable *)info;
185 evacuate(p); // does the failed_to_evac stuff
186 return rtsFalse;
187 }
188 #else
189 info = (W_)src->header.info;
190 #endif
191
192 to = alloc_for_copy(size_to_reserve, stp);
193 *p = (StgClosure *)to;
194
195 from = (StgPtr)src;
196 to[0] = info;
197 for (i = 1; i < size_to_copy; i++) { // unroll for small i
198 to[i] = from[i];
199 }
200
201 #if defined(PARALLEL_GC)
202 write_barrier();
203 #endif
204 src->header.info = (const StgInfoTable*)MK_FORWARDING_PTR(to);
205
206 #ifdef PROFILING
207 // We store the size of the just evacuated object in the LDV word so that
208 // the profiler can guess the position of the next object later.
209 SET_EVACUAEE_FOR_LDV(from, size_to_reserve);
210 // fill the slop
211 if (size_to_reserve - size_to_copy > 0)
212 LDV_FILL_SLOP(to + size_to_copy, (int)(size_to_reserve - size_to_copy));
213 #endif
214
215 return rtsTrue;
216 }
217
218
219 /* Copy wrappers that don't tag the closure after copying */
220 STATIC_INLINE GNUC_ATTR_HOT void
221 copy(StgClosure **p, const StgInfoTable *info,
222 StgClosure *src, nat size, step *stp)
223 {
224 copy_tag(p,info,src,size,stp,0);
225 }
226
227 /* -----------------------------------------------------------------------------
228 Evacuate a large object
229
230 This just consists of removing the object from the (doubly-linked)
231 step->large_objects list, and linking it on to the (singly-linked)
232 step->new_large_objects list, from where it will be scavenged later.
233
234 Convention: bd->flags has BF_EVACUATED set for a large object
235 that has been evacuated, or unset otherwise.
236 -------------------------------------------------------------------------- */
237
238 STATIC_INLINE void
239 evacuate_large(StgPtr p)
240 {
241 bdescr *bd = Bdescr(p);
242 step *stp, *new_stp;
243 step_workspace *ws;
244
245 stp = bd->step;
246 ACQUIRE_SPIN_LOCK(&stp->sync_large_objects);
247
248 // already evacuated?
249 if (bd->flags & BF_EVACUATED) {
250 /* Don't forget to set the gct->failed_to_evac flag if we didn't get
251 * the desired destination (see comments in evacuate()).
252 */
253 if (stp < gct->evac_step) {
254 gct->failed_to_evac = rtsTrue;
255 TICK_GC_FAILED_PROMOTION();
256 }
257 RELEASE_SPIN_LOCK(&stp->sync_large_objects);
258 return;
259 }
260
261 // remove from large_object list
262 if (bd->u.back) {
263 bd->u.back->link = bd->link;
264 } else { // first object in the list
265 stp->large_objects = bd->link;
266 }
267 if (bd->link) {
268 bd->link->u.back = bd->u.back;
269 }
270
271 /* link it on to the evacuated large object list of the destination step
272 */
273 new_stp = stp->to;
274 if (new_stp < gct->evac_step) {
275 if (gct->eager_promotion) {
276 new_stp = gct->evac_step;
277 } else {
278 gct->failed_to_evac = rtsTrue;
279 }
280 }
281
282 ws = &gct->steps[new_stp->abs_no];
283
284 bd->flags |= BF_EVACUATED;
285 bd->step = new_stp;
286 bd->gen_no = new_stp->gen_no;
287
288 // If this is a block of pinned objects, we don't have to scan
289 // these objects, because they aren't allowed to contain any
290 // pointers. For these blocks, we skip the scavenge stage and put
291 // them straight on the scavenged_large_objects list.
292 if (bd->flags & BF_PINNED) {
293 ASSERT(get_itbl((StgClosure *)p)->type == ARR_WORDS);
294 if (new_stp != stp) { ACQUIRE_SPIN_LOCK(&new_stp->sync_large_objects); }
295 dbl_link_onto(bd, &new_stp->scavenged_large_objects);
296 new_stp->n_scavenged_large_blocks += bd->blocks;
297 if (new_stp != stp) { RELEASE_SPIN_LOCK(&new_stp->sync_large_objects); }
298 } else {
299 bd->link = ws->todo_large_objects;
300 ws->todo_large_objects = bd;
301 }
302
303 RELEASE_SPIN_LOCK(&stp->sync_large_objects);
304 }
305
306 /* ----------------------------------------------------------------------------
307 Evacuate
308
309 This is called (eventually) for every live object in the system.
310
311 The caller to evacuate specifies a desired generation in the
312 gct->evac_step thread-local variable. The following conditions apply to
313 evacuating an object which resides in generation M when we're
314 collecting up to generation N
315
316 if M >= gct->evac_step
317 if M > N do nothing
318 else evac to step->to
319
320 if M < gct->evac_step evac to gct->evac_step, step 0
321
322 if the object is already evacuated, then we check which generation
323 it now resides in.
324
325 if M >= gct->evac_step do nothing
326 if M < gct->evac_step set gct->failed_to_evac flag to indicate that we
327 didn't manage to evacuate this object into gct->evac_step.
328
329
330 OPTIMISATION NOTES:
331
332 evacuate() is the single most important function performance-wise
333 in the GC. Various things have been tried to speed it up, but as
334 far as I can tell the code generated by gcc 3.2 with -O2 is about
335 as good as it's going to get. We pass the argument to evacuate()
336 in a register using the 'regparm' attribute (see the prototype for
337 evacuate() near the top of this file).
338
339 Changing evacuate() to take an (StgClosure **) rather than
340 returning the new pointer seems attractive, because we can avoid
341 writing back the pointer when it hasn't changed (eg. for a static
342 object, or an object in a generation > N). However, I tried it and
343 it doesn't help. One reason is that the (StgClosure **) pointer
344 gets spilled to the stack inside evacuate(), resulting in far more
345 extra reads/writes than we save.
346 ------------------------------------------------------------------------- */
347
348 REGPARM1 GNUC_ATTR_HOT void
349 evacuate(StgClosure **p)
350 {
351 bdescr *bd = NULL;
352 step *stp;
353 StgClosure *q;
354 const StgInfoTable *info;
355 StgWord tag;
356
357 q = *p;
358
359 loop:
360 /* The tag and the pointer are split, to be merged after evacing */
361 tag = GET_CLOSURE_TAG(q);
362 q = UNTAG_CLOSURE(q);
363
364 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
365
366 if (!HEAP_ALLOCED_GC(q)) {
367
368 if (!major_gc) return;
369
370 info = get_itbl(q);
371 switch (info->type) {
372
373 case THUNK_STATIC:
374 if (info->srt_bitmap != 0) {
375 if (*THUNK_STATIC_LINK((StgClosure *)q) == NULL) {
376 #ifndef THREADED_RTS
377 *THUNK_STATIC_LINK((StgClosure *)q) = gct->static_objects;
378 gct->static_objects = (StgClosure *)q;
379 #else
380 StgPtr link;
381 link = (StgPtr)cas((StgPtr)THUNK_STATIC_LINK((StgClosure *)q),
382 (StgWord)NULL,
383 (StgWord)gct->static_objects);
384 if (link == NULL) {
385 gct->static_objects = (StgClosure *)q;
386 }
387 #endif
388 }
389 }
390 return;
391
392 case FUN_STATIC:
393 if (info->srt_bitmap != 0 &&
394 *FUN_STATIC_LINK((StgClosure *)q) == NULL) {
395 #ifndef THREADED_RTS
396 *FUN_STATIC_LINK((StgClosure *)q) = gct->static_objects;
397 gct->static_objects = (StgClosure *)q;
398 #else
399 StgPtr link;
400 link = (StgPtr)cas((StgPtr)FUN_STATIC_LINK((StgClosure *)q),
401 (StgWord)NULL,
402 (StgWord)gct->static_objects);
403 if (link == NULL) {
404 gct->static_objects = (StgClosure *)q;
405 }
406 #endif
407 }
408 return;
409
410 case IND_STATIC:
411 /* If q->saved_info != NULL, then it's a revertible CAF - it'll be
412 * on the CAF list, so don't do anything with it here (we'll
413 * scavenge it later).
414 */
415 if (((StgIndStatic *)q)->saved_info == NULL) {
416 if (*IND_STATIC_LINK((StgClosure *)q) == NULL) {
417 #ifndef THREADED_RTS
418 *IND_STATIC_LINK((StgClosure *)q) = gct->static_objects;
419 gct->static_objects = (StgClosure *)q;
420 #else
421 StgPtr link;
422 link = (StgPtr)cas((StgPtr)IND_STATIC_LINK((StgClosure *)q),
423 (StgWord)NULL,
424 (StgWord)gct->static_objects);
425 if (link == NULL) {
426 gct->static_objects = (StgClosure *)q;
427 }
428 #endif
429 }
430 }
431 return;
432
433 case CONSTR_STATIC:
434 if (*STATIC_LINK(info,(StgClosure *)q) == NULL) {
435 #ifndef THREADED_RTS
436 *STATIC_LINK(info,(StgClosure *)q) = gct->static_objects;
437 gct->static_objects = (StgClosure *)q;
438 #else
439 StgPtr link;
440 link = (StgPtr)cas((StgPtr)STATIC_LINK(info,(StgClosure *)q),
441 (StgWord)NULL,
442 (StgWord)gct->static_objects);
443 if (link == NULL) {
444 gct->static_objects = (StgClosure *)q;
445 }
446 #endif
447 }
448 /* I am assuming that static_objects pointers are not
449 * written to other objects, and thus, no need to retag. */
450 return;
451
452 case CONSTR_NOCAF_STATIC:
453 /* no need to put these on the static linked list, they don't need
454 * to be scavenged.
455 */
456 return;
457
458 default:
459 barf("evacuate(static): strange closure type %d", (int)(info->type));
460 }
461 }
462
463 bd = Bdescr((P_)q);
464
465 if ((bd->flags & (BF_LARGE | BF_MARKED | BF_EVACUATED)) != 0) {
466
467 // pointer into to-space: just return it. It might be a pointer
468 // into a generation that we aren't collecting (> N), or it
469 // might just be a pointer into to-space. The latter doesn't
470 // happen often, but allowing it makes certain things a bit
471 // easier; e.g. scavenging an object is idempotent, so it's OK to
472 // have an object on the mutable list multiple times.
473 if (bd->flags & BF_EVACUATED) {
474 // We aren't copying this object, so we have to check
475 // whether it is already in the target generation. (this is
476 // the write barrier).
477 if (bd->step < gct->evac_step) {
478 gct->failed_to_evac = rtsTrue;
479 TICK_GC_FAILED_PROMOTION();
480 }
481 return;
482 }
483
484 /* evacuate large objects by re-linking them onto a different list.
485 */
486 if (bd->flags & BF_LARGE) {
487 info = get_itbl(q);
488 if (info->type == TSO &&
489 ((StgTSO *)q)->what_next == ThreadRelocated) {
490 q = (StgClosure *)((StgTSO *)q)->_link;
491 *p = q;
492 goto loop;
493 }
494 evacuate_large((P_)q);
495 return;
496 }
497
498 /* If the object is in a step that we're compacting, then we
499 * need to use an alternative evacuate procedure.
500 */
501 if (!is_marked((P_)q,bd)) {
502 mark((P_)q,bd);
503 push_mark_stack((P_)q);
504 }
505 return;
506 }
507
508 stp = bd->step->to;
509
510 info = q->header.info;
511 if (IS_FORWARDING_PTR(info))
512 {
513 /* Already evacuated, just return the forwarding address.
514 * HOWEVER: if the requested destination generation (gct->evac_step) is
515 * older than the actual generation (because the object was
516 * already evacuated to a younger generation) then we have to
517 * set the gct->failed_to_evac flag to indicate that we couldn't
518 * manage to promote the object to the desired generation.
519 */
520 /*
521 * Optimisation: the check is fairly expensive, but we can often
522 * shortcut it if either the required generation is 0, or the
523 * current object (the EVACUATED) is in a high enough generation.
524 * We know that an EVACUATED always points to an object in the
525 * same or an older generation. stp is the lowest step that the
526 * current object would be evacuated to, so we only do the full
527 * check if stp is too low.
528 */
529 StgClosure *e = (StgClosure*)UN_FORWARDING_PTR(info);
530 *p = TAG_CLOSURE(tag,e);
531 if (stp < gct->evac_step) { // optimisation
532 if (Bdescr((P_)e)->step < gct->evac_step) {
533 gct->failed_to_evac = rtsTrue;
534 TICK_GC_FAILED_PROMOTION();
535 }
536 }
537 return;
538 }
539
540 switch (INFO_PTR_TO_STRUCT(info)->type) {
541
542 case WHITEHOLE:
543 goto loop;
544
545 case MUT_VAR_CLEAN:
546 case MUT_VAR_DIRTY:
547 case MVAR_CLEAN:
548 case MVAR_DIRTY:
549 copy(p,info,q,sizeW_fromITBL(INFO_PTR_TO_STRUCT(info)),stp);
550 return;
551
552 // For ints and chars of low value, save space by replacing references to
553 // these with closures with references to common, shared ones in the RTS.
554 //
555 // * Except when compiling into Windows DLLs which don't support cross-package
556 // data references very well.
557 //
558 case CONSTR_0_1:
559 {
560 #if defined(__PIC__) && defined(mingw32_HOST_OS)
561 copy_tag_nolock(p,info,q,sizeofW(StgHeader)+1,stp,tag);
562 #else
563 StgWord w = (StgWord)q->payload[0];
564 if (info == Czh_con_info &&
565 // unsigned, so always true: (StgChar)w >= MIN_CHARLIKE &&
566 (StgChar)w <= MAX_CHARLIKE) {
567 *p = TAG_CLOSURE(tag,
568 (StgClosure *)CHARLIKE_CLOSURE((StgChar)w)
569 );
570 }
571 else if (info == Izh_con_info &&
572 (StgInt)w >= MIN_INTLIKE && (StgInt)w <= MAX_INTLIKE) {
573 *p = TAG_CLOSURE(tag,
574 (StgClosure *)INTLIKE_CLOSURE((StgInt)w)
575 );
576 }
577 else {
578 copy_tag_nolock(p,info,q,sizeofW(StgHeader)+1,stp,tag);
579 }
580 #endif
581 return;
582 }
583
584 case FUN_0_1:
585 case FUN_1_0:
586 case CONSTR_1_0:
587 copy_tag_nolock(p,info,q,sizeofW(StgHeader)+1,stp,tag);
588 return;
589
590 case THUNK_1_0:
591 case THUNK_0_1:
592 copy(p,info,q,sizeofW(StgThunk)+1,stp);
593 return;
594
595 case THUNK_1_1:
596 case THUNK_2_0:
597 case THUNK_0_2:
598 #ifdef NO_PROMOTE_THUNKS
599 if (bd->gen_no == 0 &&
600 bd->step->no != 0 &&
601 bd->step->no == generations[bd->gen_no].n_steps-1) {
602 stp = bd->step;
603 }
604 #endif
605 copy(p,info,q,sizeofW(StgThunk)+2,stp);
606 return;
607
608 case FUN_1_1:
609 case FUN_2_0:
610 case FUN_0_2:
611 case CONSTR_1_1:
612 case CONSTR_2_0:
613 copy_tag_nolock(p,info,q,sizeofW(StgHeader)+2,stp,tag);
614 return;
615
616 case CONSTR_0_2:
617 copy_tag_nolock(p,info,q,sizeofW(StgHeader)+2,stp,tag);
618 return;
619
620 case THUNK:
621 copy(p,info,q,thunk_sizeW_fromITBL(INFO_PTR_TO_STRUCT(info)),stp);
622 return;
623
624 case FUN:
625 case IND_PERM:
626 case IND_OLDGEN_PERM:
627 case CONSTR:
628 copy_tag_nolock(p,info,q,sizeW_fromITBL(INFO_PTR_TO_STRUCT(info)),stp,tag);
629 return;
630
631 case WEAK:
632 case STABLE_NAME:
633 copy_tag(p,info,q,sizeW_fromITBL(INFO_PTR_TO_STRUCT(info)),stp,tag);
634 return;
635
636 case BCO:
637 copy(p,info,q,bco_sizeW((StgBCO *)q),stp);
638 return;
639
640 case CAF_BLACKHOLE:
641 case BLACKHOLE:
642 copyPart(p,q,BLACKHOLE_sizeW(),sizeofW(StgHeader),stp);
643 return;
644
645 case THUNK_SELECTOR:
646 eval_thunk_selector(p, (StgSelector *)q, rtsTrue);
647 return;
648
649 case IND:
650 case IND_OLDGEN:
651 // follow chains of indirections, don't evacuate them
652 q = ((StgInd*)q)->indirectee;
653 *p = q;
654 goto loop;
655
656 case RET_BCO:
657 case RET_SMALL:
658 case RET_BIG:
659 case RET_DYN:
660 case UPDATE_FRAME:
661 case STOP_FRAME:
662 case CATCH_FRAME:
663 case CATCH_STM_FRAME:
664 case CATCH_RETRY_FRAME:
665 case ATOMICALLY_FRAME:
666 // shouldn't see these
667 barf("evacuate: stack frame at %p\n", q);
668
669 case PAP:
670 copy(p,info,q,pap_sizeW((StgPAP*)q),stp);
671 return;
672
673 case AP:
674 copy(p,info,q,ap_sizeW((StgAP*)q),stp);
675 return;
676
677 case AP_STACK:
678 copy(p,info,q,ap_stack_sizeW((StgAP_STACK*)q),stp);
679 return;
680
681 case ARR_WORDS:
682 // just copy the block
683 copy(p,info,q,arr_words_sizeW((StgArrWords *)q),stp);
684 return;
685
686 case MUT_ARR_PTRS_CLEAN:
687 case MUT_ARR_PTRS_DIRTY:
688 case MUT_ARR_PTRS_FROZEN:
689 case MUT_ARR_PTRS_FROZEN0:
690 // just copy the block
691 copy(p,info,q,mut_arr_ptrs_sizeW((StgMutArrPtrs *)q),stp);
692 return;
693
694 case TSO:
695 {
696 StgTSO *tso = (StgTSO *)q;
697
698 /* Deal with redirected TSOs (a TSO that's had its stack enlarged).
699 */
700 if (tso->what_next == ThreadRelocated) {
701 q = (StgClosure *)tso->_link;
702 *p = q;
703 goto loop;
704 }
705
706 /* To evacuate a small TSO, we need to adjust the stack pointer
707 */
708 {
709 StgTSO *new_tso;
710 StgPtr r, s;
711 rtsBool mine;
712
713 mine = copyPart(p,(StgClosure *)tso, tso_sizeW(tso),
714 sizeofW(StgTSO), stp);
715 if (mine) {
716 new_tso = (StgTSO *)*p;
717 move_TSO(tso, new_tso);
718 for (r = tso->sp, s = new_tso->sp;
719 r < tso->stack+tso->stack_size;) {
720 *s++ = *r++;
721 }
722 }
723 return;
724 }
725 }
726
727 case TREC_HEADER:
728 copy(p,info,q,sizeofW(StgTRecHeader),stp);
729 return;
730
731 case TVAR_WATCH_QUEUE:
732 copy(p,info,q,sizeofW(StgTVarWatchQueue),stp);
733 return;
734
735 case TVAR:
736 copy(p,info,q,sizeofW(StgTVar),stp);
737 return;
738
739 case TREC_CHUNK:
740 copy(p,info,q,sizeofW(StgTRecChunk),stp);
741 return;
742
743 case ATOMIC_INVARIANT:
744 copy(p,info,q,sizeofW(StgAtomicInvariant),stp);
745 return;
746
747 case INVARIANT_CHECK_QUEUE:
748 copy(p,info,q,sizeofW(StgInvariantCheckQueue),stp);
749 return;
750
751 default:
752 barf("evacuate: strange closure type %d", (int)(INFO_PTR_TO_STRUCT(info)->type));
753 }
754
755 barf("evacuate");
756 }
757
758 /* -----------------------------------------------------------------------------
759 Evaluate a THUNK_SELECTOR if possible.
760
761 p points to a THUNK_SELECTOR that we want to evaluate. The
762 result of "evaluating" it will be evacuated and a pointer to the
763 to-space closure will be returned.
764
765 If the THUNK_SELECTOR could not be evaluated (its selectee is still
766 a THUNK, for example), then the THUNK_SELECTOR itself will be
767 evacuated.
768 -------------------------------------------------------------------------- */
769 static void
770 unchain_thunk_selectors(StgSelector *p, StgClosure *val)
771 {
772 StgSelector *prev;
773
774 prev = NULL;
775 while (p)
776 {
777 #ifdef THREADED_RTS
778 ASSERT(p->header.info == &stg_WHITEHOLE_info);
779 #else
780 ASSERT(p->header.info == &stg_BLACKHOLE_info);
781 #endif
782 // val must be in to-space. Not always: when we recursively
783 // invoke eval_thunk_selector(), the recursive calls will not
784 // evacuate the value (because we want to select on the value,
785 // not evacuate it), so in this case val is in from-space.
786 // ASSERT(!HEAP_ALLOCED_GC(val) || Bdescr((P_)val)->gen_no > N || (Bdescr((P_)val)->flags & BF_EVACUATED));
787
788 prev = (StgSelector*)((StgClosure *)p)->payload[0];
789
790 // Update the THUNK_SELECTOR with an indirection to the
791 // value. The value is still in from-space at this stage.
792 //
793 // (old note: Why not do upd_evacuee(q,p)? Because we have an
794 // invariant that an EVACUATED closure always points to an
795 // object in the same or an older generation (required by
796 // the short-cut test in the EVACUATED case, below).
797 if ((StgClosure *)p == val) {
798 // must be a loop; just leave a BLACKHOLE in place. This
799 // can happen when we have a chain of selectors that
800 // eventually loops back on itself. We can't leave an
801 // indirection pointing to itself, and we want the program
802 // to deadlock if it ever enters this closure, so
803 // BLACKHOLE is correct.
804 SET_INFO(p, &stg_BLACKHOLE_info);
805 } else {
806 ((StgInd *)p)->indirectee = val;
807 write_barrier();
808 SET_INFO(p, &stg_IND_info);
809 }
810
811 // For the purposes of LDV profiling, we have created an
812 // indirection.
813 LDV_RECORD_CREATE(p);
814
815 p = prev;
816 }
817 }
818
819 static void
820 eval_thunk_selector (StgClosure **q, StgSelector * p, rtsBool evac)
821 // NB. for legacy reasons, p & q are swapped around :(
822 {
823 nat field;
824 StgInfoTable *info;
825 StgWord info_ptr;
826 StgClosure *selectee;
827 StgSelector *prev_thunk_selector;
828 bdescr *bd;
829 StgClosure *val;
830
831 prev_thunk_selector = NULL;
832 // this is a chain of THUNK_SELECTORs that we are going to update
833 // to point to the value of the current THUNK_SELECTOR. Each
834 // closure on the chain is a BLACKHOLE, and points to the next in the
835 // chain with payload[0].
836
837 selector_chain:
838
839 bd = Bdescr((StgPtr)p);
840 if (HEAP_ALLOCED_GC(p)) {
841 // If the THUNK_SELECTOR is in to-space or in a generation that we
842 // are not collecting, then bale out early. We won't be able to
843 // save any space in any case, and updating with an indirection is
844 // trickier in a non-collected gen: we would have to update the
845 // mutable list.
846 if (bd->flags & BF_EVACUATED) {
847 unchain_thunk_selectors(prev_thunk_selector, (StgClosure *)p);
848 *q = (StgClosure *)p;
849 // shortcut, behave as for: if (evac) evacuate(q);
850 if (evac && bd->step < gct->evac_step) {
851 gct->failed_to_evac = rtsTrue;
852 TICK_GC_FAILED_PROMOTION();
853 }
854 return;
855 }
856 // we don't update THUNK_SELECTORS in the compacted
857 // generation, because compaction does not remove the INDs
858 // that result, this causes confusion later
859 // (scavenge_mark_stack doesn't deal with IND). BEWARE! This
860 // bit is very tricky to get right. If you make changes
861 // around here, test by compiling stage 3 with +RTS -c -RTS.
862 if (bd->flags & BF_MARKED) {
863 // must call evacuate() to mark this closure if evac==rtsTrue
864 *q = (StgClosure *)p;
865 if (evac) evacuate(q);
866 unchain_thunk_selectors(prev_thunk_selector, (StgClosure *)p);
867 return;
868 }
869 }
870
871
872 // BLACKHOLE the selector thunk, since it is now under evaluation.
873 // This is important to stop us going into an infinite loop if
874 // this selector thunk eventually refers to itself.
875 #if defined(THREADED_RTS)
876 // In threaded mode, we'll use WHITEHOLE to lock the selector
877 // thunk while we evaluate it.
878 {
879 do {
880 info_ptr = xchg((StgPtr)&p->header.info, (W_)&stg_WHITEHOLE_info);
881 } while (info_ptr == (W_)&stg_WHITEHOLE_info);
882
883 // make sure someone else didn't get here first...
884 if (IS_FORWARDING_PTR(p) ||
885 INFO_PTR_TO_STRUCT(info_ptr)->type != THUNK_SELECTOR) {
886 // v. tricky now. The THUNK_SELECTOR has been evacuated
887 // by another thread, and is now either a forwarding ptr or IND.
888 // We need to extract ourselves from the current situation
889 // as cleanly as possible.
890 // - unlock the closure
891 // - update *q, we may have done *some* evaluation
892 // - if evac, we need to call evacuate(), because we
893 // need the write-barrier stuff.
894 // - undo the chain we've built to point to p.
895 SET_INFO(p, (const StgInfoTable *)info_ptr);
896 *q = (StgClosure *)p;
897 if (evac) evacuate(q);
898 unchain_thunk_selectors(prev_thunk_selector, (StgClosure *)p);
899 return;
900 }
901 }
902 #else
903 // Save the real info pointer (NOTE: not the same as get_itbl()).
904 info_ptr = (StgWord)p->header.info;
905 SET_INFO(p,&stg_BLACKHOLE_info);
906 #endif
907
908 field = INFO_PTR_TO_STRUCT(info_ptr)->layout.selector_offset;
909
910 // The selectee might be a constructor closure,
911 // so we untag the pointer.
912 selectee = UNTAG_CLOSURE(p->selectee);
913
914 selector_loop:
915 // selectee now points to the closure that we're trying to select
916 // a field from. It may or may not be in to-space: we try not to
917 // end up in to-space, but it's impractical to avoid it in
918 // general. The compacting GC scatters to-space pointers in
919 // from-space during marking, for example. We rely on the property
920 // that evacuate() doesn't mind if it gets passed a to-space pointer.
921
922 info = (StgInfoTable*)selectee->header.info;
923
924 if (IS_FORWARDING_PTR(info)) {
925 // We don't follow pointers into to-space; the constructor
926 // has already been evacuated, so we won't save any space
927 // leaks by evaluating this selector thunk anyhow.
928 goto bale_out;
929 }
930
931 info = INFO_PTR_TO_STRUCT(info);
932 switch (info->type) {
933 case WHITEHOLE:
934 goto bale_out; // about to be evacuated by another thread (or a loop).
935
936 case CONSTR:
937 case CONSTR_1_0:
938 case CONSTR_0_1:
939 case CONSTR_2_0:
940 case CONSTR_1_1:
941 case CONSTR_0_2:
942 case CONSTR_STATIC:
943 case CONSTR_NOCAF_STATIC:
944 {
945 // check that the size is in range
946 ASSERT(field < (StgWord32)(info->layout.payload.ptrs +
947 info->layout.payload.nptrs));
948
949 // Select the right field from the constructor
950 val = selectee->payload[field];
951
952 #ifdef PROFILING
953 // For the purposes of LDV profiling, we have destroyed
954 // the original selector thunk, p.
955 SET_INFO(p, (StgInfoTable *)info_ptr);
956 LDV_RECORD_DEAD_FILL_SLOP_DYNAMIC((StgClosure *)p);
957 #if defined(THREADED_RTS)
958 SET_INFO(p, &stg_WHITEHOLE_info);
959 #else
960 SET_INFO(p, &stg_BLACKHOLE_info);
961 #endif
962 #endif
963
964 // the closure in val is now the "value" of the
965 // THUNK_SELECTOR in p. However, val may itself be a
966 // THUNK_SELECTOR, in which case we want to continue
967 // evaluating until we find the real value, and then
968 // update the whole chain to point to the value.
969 val_loop:
970 info_ptr = (StgWord)UNTAG_CLOSURE(val)->header.info;
971 if (!IS_FORWARDING_PTR(info_ptr))
972 {
973 info = INFO_PTR_TO_STRUCT(info_ptr);
974 switch (info->type) {
975 case IND:
976 case IND_PERM:
977 case IND_OLDGEN:
978 case IND_OLDGEN_PERM:
979 case IND_STATIC:
980 val = ((StgInd *)val)->indirectee;
981 goto val_loop;
982 case THUNK_SELECTOR:
983 ((StgClosure*)p)->payload[0] = (StgClosure *)prev_thunk_selector;
984 prev_thunk_selector = p;
985 p = (StgSelector*)val;
986 goto selector_chain;
987 default:
988 break;
989 }
990 }
991 ((StgClosure*)p)->payload[0] = (StgClosure *)prev_thunk_selector;
992 prev_thunk_selector = p;
993
994 *q = val;
995
996 // update the other selectors in the chain *before*
997 // evacuating the value. This is necessary in the case
998 // where the value turns out to be one of the selectors
999 // in the chain (i.e. we have a loop), and evacuating it
1000 // would corrupt the chain.
1001 unchain_thunk_selectors(prev_thunk_selector, val);
1002
1003 // evacuate() cannot recurse through
1004 // eval_thunk_selector(), because we know val is not
1005 // a THUNK_SELECTOR.
1006 if (evac) evacuate(q);
1007 return;
1008 }
1009
1010 case IND:
1011 case IND_PERM:
1012 case IND_OLDGEN:
1013 case IND_OLDGEN_PERM:
1014 case IND_STATIC:
1015 // Again, we might need to untag a constructor.
1016 selectee = UNTAG_CLOSURE( ((StgInd *)selectee)->indirectee );
1017 goto selector_loop;
1018
1019 case THUNK_SELECTOR:
1020 {
1021 StgClosure *val;
1022
1023 // recursively evaluate this selector. We don't want to
1024 // recurse indefinitely, so we impose a depth bound.
1025 if (gct->thunk_selector_depth >= MAX_THUNK_SELECTOR_DEPTH) {
1026 goto bale_out;
1027 }
1028
1029 gct->thunk_selector_depth++;
1030 // rtsFalse says "don't evacuate the result". It will,
1031 // however, update any THUNK_SELECTORs that are evaluated
1032 // along the way.
1033 eval_thunk_selector(&val, (StgSelector*)selectee, rtsFalse);
1034 gct->thunk_selector_depth--;
1035
1036 // did we actually manage to evaluate it?
1037 if (val == selectee) goto bale_out;
1038
1039 // Of course this pointer might be tagged...
1040 selectee = UNTAG_CLOSURE(val);
1041 goto selector_loop;
1042 }
1043
1044 case AP:
1045 case AP_STACK:
1046 case THUNK:
1047 case THUNK_1_0:
1048 case THUNK_0_1:
1049 case THUNK_2_0:
1050 case THUNK_1_1:
1051 case THUNK_0_2:
1052 case THUNK_STATIC:
1053 case CAF_BLACKHOLE:
1054 case BLACKHOLE:
1055 // not evaluated yet
1056 goto bale_out;
1057
1058 default:
1059 barf("eval_thunk_selector: strange selectee %d",
1060 (int)(info->type));
1061 }
1062
1063 bale_out:
1064 // We didn't manage to evaluate this thunk; restore the old info
1065 // pointer. But don't forget: we still need to evacuate the thunk itself.
1066 SET_INFO(p, (const StgInfoTable *)info_ptr);
1067 // THREADED_RTS: we just unlocked the thunk, so another thread
1068 // might get in and update it. copy() will lock it again and
1069 // check whether it was updated in the meantime.
1070 *q = (StgClosure *)p;
1071 if (evac) {
1072 copy(q,(const StgInfoTable *)info_ptr,(StgClosure *)p,THUNK_SELECTOR_sizeW(),bd->step->to);
1073 }
1074 unchain_thunk_selectors(prev_thunk_selector, *q);
1075 return;
1076 }