source: trunk/gc6.8/mark.c@ 132

Last change on this file since 132 was 132, checked in by cinc, 19 years ago

Boehm-Demers-Weiser garbage collector. Single-threaded for OS/2.
File size: 54.6 KB
Line 
1
2/*
3 * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
4 * Copyright (c) 1991-1995 by Xerox Corporation. All rights reserved.
5 * Copyright (c) 2000 by Hewlett-Packard Company. All rights reserved.
6 *
7 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
8 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
9 *
10 * Permission is hereby granted to use or copy this program
11 * for any purpose, provided the above notices are retained on all copies.
12 * Permission to modify the code and to distribute modified code is granted,
13 * provided the above notices are retained, and a notice that the code was
14 * modified is included with the above copyright notice.
15 *
16 */
17
18
19# include <stdio.h>
20# include "private/gc_pmark.h"
21
22#if defined(MSWIN32) && defined(__GNUC__)
23# include <excpt.h>
24#endif
25
26/* We put this here to minimize the risk of inlining. */
27/*VARARGS*/
28#ifdef __WATCOMC__
29 void GC_noop(void *p, ...) {}
30#else
31 void GC_noop() {}
32#endif
33
34/* Single argument version, robust against whole program analysis. */
35void GC_noop1(x)
36word x;
37{
38 static VOLATILE word sink;
39
40 sink = x;
41}
42
43/* mark_proc GC_mark_procs[MAX_MARK_PROCS] = {0} -- declared in gc_priv.h */
44
45word GC_n_mark_procs = GC_RESERVED_MARK_PROCS;
46
47/* Initialize GC_obj_kinds properly and standard free lists properly. */
48/* This must be done statically since they may be accessed before */
49/* GC_init is called. */
50/* It's done here, since we need to deal with mark descriptors. */
51struct obj_kind GC_obj_kinds[MAXOBJKINDS] = {
52/* PTRFREE */ { &GC_aobjfreelist[0], 0 /* filled in dynamically */,
53 0 | GC_DS_LENGTH, FALSE, FALSE },
54/* NORMAL */ { &GC_objfreelist[0], 0,
55 0 | GC_DS_LENGTH, /* Adjusted in GC_init_inner for EXTRA_BYTES */
56 TRUE /* add length to descr */, TRUE },
57/* UNCOLLECTABLE */
58 { &GC_uobjfreelist[0], 0,
59 0 | GC_DS_LENGTH, TRUE /* add length to descr */, TRUE },
60# ifdef ATOMIC_UNCOLLECTABLE
61 /* AUNCOLLECTABLE */
62 { &GC_auobjfreelist[0], 0,
63 0 | GC_DS_LENGTH, FALSE /* add length to descr */, FALSE },
64# endif
65# ifdef STUBBORN_ALLOC
66/*STUBBORN*/ { &GC_sobjfreelist[0], 0,
67 0 | GC_DS_LENGTH, TRUE /* add length to descr */, TRUE },
68# endif
69};
70
71# ifdef ATOMIC_UNCOLLECTABLE
72# ifdef STUBBORN_ALLOC
73 int GC_n_kinds = 5;
74# else
75 int GC_n_kinds = 4;
76# endif
77# else
78# ifdef STUBBORN_ALLOC
79 int GC_n_kinds = 4;
80# else
81 int GC_n_kinds = 3;
82# endif
83# endif
84
85
86# ifndef INITIAL_MARK_STACK_SIZE
87# define INITIAL_MARK_STACK_SIZE (1*HBLKSIZE)
88 /* INITIAL_MARK_STACK_SIZE * sizeof(mse) should be a */
89 /* multiple of HBLKSIZE. */
90 /* The incremental collector actually likes a larger */
91 /* size, since it want to push all marked dirty objs */
92 /* before marking anything new. Currently we let it */
93 /* grow dynamically. */
94# endif
95
96/*
97 * Limits of stack for GC_mark routine.
98 * All ranges between GC_mark_stack(incl.) and GC_mark_stack_top(incl.) still
99 * need to be marked from.
100 */
101
102word GC_n_rescuing_pages; /* Number of dirty pages we marked from */
103 /* excludes ptrfree pages, etc. */
104
105mse * GC_mark_stack;
106
107mse * GC_mark_stack_limit;
108
109word GC_mark_stack_size = 0;
110
111#ifdef PARALLEL_MARK
112 mse * VOLATILE GC_mark_stack_top;
113#else
114 mse * GC_mark_stack_top;
115#endif
116
117static struct hblk * scan_ptr;
118
119mark_state_t GC_mark_state = MS_NONE;
120
121GC_bool GC_mark_stack_too_small = FALSE;
122
123GC_bool GC_objects_are_marked = FALSE; /* Are there collectable marked */
124 /* objects in the heap? */
125
126/* Is a collection in progress? Note that this can return true in the */
127/* nonincremental case, if a collection has been abandoned and the */
128/* mark state is now MS_INVALID. */
129GC_bool GC_collection_in_progress()
130{
131 return(GC_mark_state != MS_NONE);
132}
133
134/* clear all mark bits in the header */
135void GC_clear_hdr_marks(hhdr)
136register hdr * hhdr;
137{
138# ifdef USE_MARK_BYTES
139 BZERO(hhdr -> hb_marks, MARK_BITS_SZ);
140# else
141 BZERO(hhdr -> hb_marks, MARK_BITS_SZ*sizeof(word));
142# endif
143}
144
145/* Set all mark bits in the header. Used for uncollectable blocks. */
146void GC_set_hdr_marks(hhdr)
147register hdr * hhdr;
148{
149 register int i;
150
151 for (i = 0; i < MARK_BITS_SZ; ++i) {
152# ifdef USE_MARK_BYTES
153 hhdr -> hb_marks[i] = 1;
154# else
155 hhdr -> hb_marks[i] = ONES;
156# endif
157 }
158}
159
160/*
161 * Clear all mark bits associated with block h.
162 */
163/*ARGSUSED*/
164# if defined(__STDC__) || defined(__cplusplus)
165 static void clear_marks_for_block(struct hblk *h, word dummy)
166# else
167 static void clear_marks_for_block(h, dummy)
168 struct hblk *h;
169 word dummy;
170# endif
171{
172 register hdr * hhdr = HDR(h);
173
174 if (IS_UNCOLLECTABLE(hhdr -> hb_obj_kind)) return;
175 /* Mark bit for these is cleared only once the object is */
176 /* explicitly deallocated. This either frees the block, or */
177 /* the bit is cleared once the object is on the free list. */
178 GC_clear_hdr_marks(hhdr);
179}
180
181/* Slow but general routines for setting/clearing/asking about mark bits */
182void GC_set_mark_bit(p)
183ptr_t p;
184{
185 register struct hblk *h = HBLKPTR(p);
186 register hdr * hhdr = HDR(h);
187 register int word_no = (word *)p - (word *)h;
188
189 set_mark_bit_from_hdr(hhdr, word_no);
190}
191
192void GC_clear_mark_bit(p)
193ptr_t p;
194{
195 register struct hblk *h = HBLKPTR(p);
196 register hdr * hhdr = HDR(h);
197 register int word_no = (word *)p - (word *)h;
198
199 clear_mark_bit_from_hdr(hhdr, word_no);
200}
201
202GC_bool GC_is_marked(p)
203ptr_t p;
204{
205 register struct hblk *h = HBLKPTR(p);
206 register hdr * hhdr = HDR(h);
207 register int word_no = (word *)p - (word *)h;
208
209 return(mark_bit_from_hdr(hhdr, word_no));
210}
211
212
213/*
214 * Clear mark bits in all allocated heap blocks. This invalidates
215 * the marker invariant, and sets GC_mark_state to reflect this.
216 * (This implicitly starts marking to reestablish the invariant.)
217 */
218void GC_clear_marks()
219{
220 GC_apply_to_all_blocks(clear_marks_for_block, (word)0);
221 GC_objects_are_marked = FALSE;
222 GC_mark_state = MS_INVALID;
223 scan_ptr = 0;
224# ifdef GATHERSTATS
225 /* Counters reflect currently marked objects: reset here */
226 GC_composite_in_use = 0;
227 GC_atomic_in_use = 0;
228# endif
229
230}
231
232/* Initiate a garbage collection. Initiates a full collection if the */
233/* mark state is invalid. */
234/*ARGSUSED*/
235void GC_initiate_gc()
236{
237 if (GC_dirty_maintained) GC_read_dirty();
238# ifdef STUBBORN_ALLOC
239 GC_read_changed();
240# endif
241# ifdef CHECKSUMS
242 {
243 extern void GC_check_dirty();
244
245 if (GC_dirty_maintained) GC_check_dirty();
246 }
247# endif
248 GC_n_rescuing_pages = 0;
249 if (GC_mark_state == MS_NONE) {
250 GC_mark_state = MS_PUSH_RESCUERS;
251 } else if (GC_mark_state != MS_INVALID) {
252 ABORT("unexpected state");
253 } /* else this is really a full collection, and mark */
254 /* bits are invalid. */
255 scan_ptr = 0;
256}
257
258
259static void alloc_mark_stack();
260
261/* Perform a small amount of marking. */
262/* We try to touch roughly a page of memory. */
263/* Return TRUE if we just finished a mark phase. */
264/* Cold_gc_frame is an address inside a GC frame that */
265/* remains valid until all marking is complete. */
266/* A zero value indicates that it's OK to miss some */
267/* register values. */
268/* We hold the allocation lock. In the case of */
269/* incremental collection, the world may not be stopped.*/
270#ifdef MSWIN32
271 /* For win32, this is called after we establish a structured */
272 /* exception handler, in case Windows unmaps one of our root */
273 /* segments. See below. In either case, we acquire the */
274 /* allocator lock long before we get here. */
275 GC_bool GC_mark_some_inner(cold_gc_frame)
276 ptr_t cold_gc_frame;
277#else
278 GC_bool GC_mark_some(cold_gc_frame)
279 ptr_t cold_gc_frame;
280#endif
281{
282 switch(GC_mark_state) {
283 case MS_NONE:
284 return(FALSE);
285
286 case MS_PUSH_RESCUERS:
287 if (GC_mark_stack_top
288 >= GC_mark_stack_limit - INITIAL_MARK_STACK_SIZE/2) {
289 /* Go ahead and mark, even though that might cause us to */
290 /* see more marked dirty objects later on. Avoid this */
291 /* in the future. */
292 GC_mark_stack_too_small = TRUE;
293 MARK_FROM_MARK_STACK();
294 return(FALSE);
295 } else {
296 scan_ptr = GC_push_next_marked_dirty(scan_ptr);
297 if (scan_ptr == 0) {
298# ifdef CONDPRINT
299 if (GC_print_stats) {
300 GC_printf1("Marked from %lu dirty pages\n",
301 (unsigned long)GC_n_rescuing_pages);
302 }
303# endif
304 GC_push_roots(FALSE, cold_gc_frame);
305 GC_objects_are_marked = TRUE;
306 if (GC_mark_state != MS_INVALID) {
307 GC_mark_state = MS_ROOTS_PUSHED;
308 }
309 }
310 }
311 return(FALSE);
312
313 case MS_PUSH_UNCOLLECTABLE:
314 if (GC_mark_stack_top
315 >= GC_mark_stack + GC_mark_stack_size/4) {
316# ifdef PARALLEL_MARK
317 /* Avoid this, since we don't parallelize the marker */
318 /* here. */
319 if (GC_parallel) GC_mark_stack_too_small = TRUE;
320# endif
321 MARK_FROM_MARK_STACK();
322 return(FALSE);
323 } else {
324 scan_ptr = GC_push_next_marked_uncollectable(scan_ptr);
325 if (scan_ptr == 0) {
326 GC_push_roots(TRUE, cold_gc_frame);
327 GC_objects_are_marked = TRUE;
328 if (GC_mark_state != MS_INVALID) {
329 GC_mark_state = MS_ROOTS_PUSHED;
330 }
331 }
332 }
333 return(FALSE);
334
335 case MS_ROOTS_PUSHED:
336# ifdef PARALLEL_MARK
337 /* In the incremental GC case, this currently doesn't */
338 /* quite do the right thing, since it runs to */
339 /* completion. On the other hand, starting a */
340 /* parallel marker is expensive, so perhaps it is */
341 /* the right thing? */
342 /* Eventually, incremental marking should run */
343 /* asynchronously in multiple threads, without grabbing */
344 /* the allocation lock. */
345 if (GC_parallel) {
346 GC_do_parallel_mark();
347 GC_ASSERT(GC_mark_stack_top < GC_first_nonempty);
348 GC_mark_stack_top = GC_mark_stack - 1;
349 if (GC_mark_stack_too_small) {
350 alloc_mark_stack(2*GC_mark_stack_size);
351 }
352 if (GC_mark_state == MS_ROOTS_PUSHED) {
353 GC_mark_state = MS_NONE;
354 return(TRUE);
355 } else {
356 return(FALSE);
357 }
358 }
359# endif
360 if (GC_mark_stack_top >= GC_mark_stack) {
361 MARK_FROM_MARK_STACK();
362 return(FALSE);
363 } else {
364 GC_mark_state = MS_NONE;
365 if (GC_mark_stack_too_small) {
366 alloc_mark_stack(2*GC_mark_stack_size);
367 }
368 return(TRUE);
369 }
370
371 case MS_INVALID:
372 case MS_PARTIALLY_INVALID:
373 if (!GC_objects_are_marked) {
374 GC_mark_state = MS_PUSH_UNCOLLECTABLE;
375 return(FALSE);
376 }
377 if (GC_mark_stack_top >= GC_mark_stack) {
378 MARK_FROM_MARK_STACK();
379 return(FALSE);
380 }
381 if (scan_ptr == 0 && GC_mark_state == MS_INVALID) {
382 /* About to start a heap scan for marked objects. */
383 /* Mark stack is empty. OK to reallocate. */
384 if (GC_mark_stack_too_small) {
385 alloc_mark_stack(2*GC_mark_stack_size);
386 }
387 GC_mark_state = MS_PARTIALLY_INVALID;
388 }
389 scan_ptr = GC_push_next_marked(scan_ptr);
390 if (scan_ptr == 0 && GC_mark_state == MS_PARTIALLY_INVALID) {
391 GC_push_roots(TRUE, cold_gc_frame);
392 GC_objects_are_marked = TRUE;
393 if (GC_mark_state != MS_INVALID) {
394 GC_mark_state = MS_ROOTS_PUSHED;
395 }
396 }
397 return(FALSE);
398 default:
399 ABORT("GC_mark_some: bad state");
400 return(FALSE);
401 }
402}
403
404
405#ifdef MSWIN32
406
407# ifdef __GNUC__
408
409 typedef struct {
410 EXCEPTION_REGISTRATION ex_reg;
411 void *alt_path;
412 } ext_ex_regn;
413
414
415 static EXCEPTION_DISPOSITION mark_ex_handler(
416 struct _EXCEPTION_RECORD *ex_rec,
417 void *est_frame,
418 struct _CONTEXT *context,
419 void *disp_ctxt)
420 {
421 if (ex_rec->ExceptionCode == STATUS_ACCESS_VIOLATION) {
422 ext_ex_regn *xer = (ext_ex_regn *)est_frame;
423
424 /* Unwind from the inner function assuming the standard */
425 /* function prologue. */
426 /* Assumes code has not been compiled with */
427 /* -fomit-frame-pointer. */
428 context->Esp = context->Ebp;
429 context->Ebp = *((DWORD *)context->Esp);
430 context->Esp = context->Esp - 8;
431
432 /* Resume execution at the "real" handler within the */
433 /* wrapper function. */
434 context->Eip = (DWORD )(xer->alt_path);
435
436 return ExceptionContinueExecution;
437
438 } else {
439 return ExceptionContinueSearch;
440 }
441 }
442# endif /* __GNUC__ */
443
444
445 GC_bool GC_mark_some(cold_gc_frame)
446 ptr_t cold_gc_frame;
447 {
448 GC_bool ret_val;
449
450# ifndef __GNUC__
451 /* Windows 98 appears to asynchronously create and remove */
452 /* writable memory mappings, for reasons we haven't yet */
453 /* understood. Since we look for writable regions to */
454 /* determine the root set, we may try to mark from an */
455 /* address range that disappeared since we started the */
456 /* collection. Thus we have to recover from faults here. */
457 /* This code does not appear to be necessary for Windows */
458 /* 95/NT/2000. Note that this code should never generate */
459 /* an incremental GC write fault. */
460
461 __try {
462
463# else /* __GNUC__ */
464
465 /* Manually install an exception handler since GCC does */
466 /* not yet support Structured Exception Handling (SEH) on */
467 /* Win32. */
468
469 ext_ex_regn er;
470
471 er.alt_path = &&handle_ex;
472 er.ex_reg.handler = mark_ex_handler;
473 asm volatile ("movl %%fs:0, %0" : "=r" (er.ex_reg.prev));
474 asm volatile ("movl %0, %%fs:0" : : "r" (&er));
475
476# endif /* __GNUC__ */
477
478 ret_val = GC_mark_some_inner(cold_gc_frame);
479
480# ifndef __GNUC__
481
482 } __except (GetExceptionCode() == EXCEPTION_ACCESS_VIOLATION ?
483 EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH) {
484
485# else /* __GNUC__ */
486
487 /* Prevent GCC from considering the following code unreachable */
488 /* and thus eliminating it. */
489 if (er.alt_path != 0)
490 goto rm_handler;
491
492handle_ex:
493 /* Execution resumes from here on an access violation. */
494
495# endif /* __GNUC__ */
496
497# ifdef CONDPRINT
498 if (GC_print_stats) {
499 GC_printf0("Caught ACCESS_VIOLATION in marker. "
500 "Memory mapping disappeared.\n");
501 }
502# endif /* CONDPRINT */
503
504 /* We have bad roots on the stack. Discard mark stack. */
505 /* Rescan from marked objects. Redetermine roots. */
506 GC_invalidate_mark_state();
507 scan_ptr = 0;
508
509 ret_val = FALSE;
510
511# ifndef __GNUC__
512
513 }
514
515# else /* __GNUC__ */
516
517rm_handler:
518 /* Uninstall the exception handler */
519 asm volatile ("mov %0, %%fs:0" : : "r" (er.ex_reg.prev));
520
521# endif /* __GNUC__ */
522
523 return ret_val;
524 }
525#endif /* MSWIN32 */
526
527
528GC_bool GC_mark_stack_empty()
529{
530 return(GC_mark_stack_top < GC_mark_stack);
531}
532
533#ifdef PROF_MARKER
534 word GC_prof_array[10];
535# define PROF(n) GC_prof_array[n]++
536#else
537# define PROF(n)
538#endif
539
540/* Given a pointer to someplace other than a small object page or the */
541/* first page of a large object, either: */
542/* - return a pointer to somewhere in the first page of the large */
543/* object, if current points to a large object. */
544/* In this case *hhdr is replaced with a pointer to the header */
545/* for the large object. */
546/* - just return current if it does not point to a large object. */
547/*ARGSUSED*/
548ptr_t GC_find_start(current, hhdr, new_hdr_p)
549register ptr_t current;
550register hdr *hhdr, **new_hdr_p;
551{
552 if (GC_all_interior_pointers) {
553 if (hhdr != 0) {
554 register ptr_t orig = current;
555
556 current = (ptr_t)HBLKPTR(current);
557 do {
558 current = current - HBLKSIZE*(word)hhdr;
559 hhdr = HDR(current);
560 } while(IS_FORWARDING_ADDR_OR_NIL(hhdr));
561 /* current points to near the start of the large object */
562 if (hhdr -> hb_flags & IGNORE_OFF_PAGE) return(orig);
563 if ((word *)orig - (word *)current
564 >= (ptrdiff_t)(hhdr->hb_sz)) {
565 /* Pointer past the end of the block */
566 return(orig);
567 }
568 *new_hdr_p = hhdr;
569 return(current);
570 } else {
571 return(current);
572 }
573 } else {
574 return(current);
575 }
576}
577
578void GC_invalidate_mark_state()
579{
580 GC_mark_state = MS_INVALID;
581 GC_mark_stack_top = GC_mark_stack-1;
582}
583
584mse * GC_signal_mark_stack_overflow(msp)
585mse * msp;
586{
587 GC_mark_state = MS_INVALID;
588 GC_mark_stack_too_small = TRUE;
589# ifdef CONDPRINT
590 if (GC_print_stats) {
591 GC_printf1("Mark stack overflow; current size = %lu entries\n",
592 GC_mark_stack_size);
593 }
594# endif
595 return(msp - GC_MARK_STACK_DISCARDS);
596}
597
598/*
599 * Mark objects pointed to by the regions described by
600 * mark stack entries between GC_mark_stack and GC_mark_stack_top,
601 * inclusive. Assumes the upper limit of a mark stack entry
602 * is never 0. A mark stack entry never has size 0.
603 * We try to traverse on the order of a hblk of memory before we return.
604 * Caller is responsible for calling this until the mark stack is empty.
605 * Note that this is the most performance critical routine in the
606 * collector. Hence it contains all sorts of ugly hacks to speed
607 * things up. In particular, we avoid procedure calls on the common
608 * path, we take advantage of peculiarities of the mark descriptor
609 * encoding, we optionally maintain a cache for the block address to
610 * header mapping, we prefetch when an object is "grayed", etc.
611 */
612mse * GC_mark_from(mark_stack_top, mark_stack, mark_stack_limit)
613mse * mark_stack_top;
614mse * mark_stack;
615mse * mark_stack_limit;
616{
617 int credit = HBLKSIZE; /* Remaining credit for marking work */
618 register word * current_p; /* Pointer to current candidate ptr. */
619 register word current; /* Candidate pointer. */
620 register word * limit; /* (Incl) limit of current candidate */
621 /* range */
622 register word descr;
623 register ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
624 register ptr_t least_ha = GC_least_plausible_heap_addr;
625 DECLARE_HDR_CACHE;
626
627# define SPLIT_RANGE_WORDS 128 /* Must be power of 2. */
628
629 GC_objects_are_marked = TRUE;
630 INIT_HDR_CACHE;
631# ifdef OS2 /* Use untweaked version to circumvent compiler problem */
632 while (mark_stack_top >= mark_stack && credit >= 0) {
633# else
634 while ((((ptr_t)mark_stack_top - (ptr_t)mark_stack) | credit)
635 >= 0) {
636# endif
637 current_p = mark_stack_top -> mse_start;
638 descr = mark_stack_top -> mse_descr;
639 retry:
640 /* current_p and descr describe the current object. */
641 /* *mark_stack_top is vacant. */
642 /* The following is 0 only for small objects described by a simple */
643 /* length descriptor. For many applications this is the common */
644 /* case, so we try to detect it quickly. */
645 if (descr & ((~(WORDS_TO_BYTES(SPLIT_RANGE_WORDS) - 1)) | GC_DS_TAGS)) {
646 word tag = descr & GC_DS_TAGS;
647
648 switch(tag) {
649 case GC_DS_LENGTH:
650 /* Large length. */
651 /* Process part of the range to avoid pushing too much on the */
652 /* stack. */
653 GC_ASSERT(descr < (word)GC_greatest_plausible_heap_addr
654 - (word)GC_least_plausible_heap_addr);
655# ifdef PARALLEL_MARK
656# define SHARE_BYTES 2048
657 if (descr > SHARE_BYTES && GC_parallel
658 && mark_stack_top < mark_stack_limit - 1) {
659 int new_size = (descr/2) & ~(sizeof(word)-1);
660 mark_stack_top -> mse_start = current_p;
661 mark_stack_top -> mse_descr = new_size + sizeof(word);
662 /* makes sure we handle */
663 /* misaligned pointers. */
664 mark_stack_top++;
665 current_p = (word *) ((char *)current_p + new_size);
666 descr -= new_size;
667 goto retry;
668 }
669# endif /* PARALLEL_MARK */
670 mark_stack_top -> mse_start =
671 limit = current_p + SPLIT_RANGE_WORDS-1;
672 mark_stack_top -> mse_descr =
673 descr - WORDS_TO_BYTES(SPLIT_RANGE_WORDS-1);
674 /* Make sure that pointers overlapping the two ranges are */
675 /* considered. */
676 limit = (word *)((char *)limit + sizeof(word) - ALIGNMENT);
677 break;
678 case GC_DS_BITMAP:
679 mark_stack_top--;
680 descr &= ~GC_DS_TAGS;
681 credit -= WORDS_TO_BYTES(WORDSZ/2); /* guess */
682 while (descr != 0) {
683 if ((signed_word)descr < 0) {
684 current = *current_p;
685 FIXUP_POINTER(current);
686 if ((ptr_t)current >= least_ha && (ptr_t)current < greatest_ha) {
687 PREFETCH((ptr_t)current);
688 HC_PUSH_CONTENTS((ptr_t)current, mark_stack_top,
689 mark_stack_limit, current_p, exit1);
690 }
691 }
692 descr <<= 1;
693 ++ current_p;
694 }
695 continue;
696 case GC_DS_PROC:
697 mark_stack_top--;
698 credit -= GC_PROC_BYTES;
699 mark_stack_top =
700 (*PROC(descr))
701 (current_p, mark_stack_top,
702 mark_stack_limit, ENV(descr));
703 continue;
704 case GC_DS_PER_OBJECT:
705 if ((signed_word)descr >= 0) {
706 /* Descriptor is in the object. */
707 descr = *(word *)((ptr_t)current_p + descr - GC_DS_PER_OBJECT);
708 } else {
709 /* Descriptor is in type descriptor pointed to by first */
710 /* word in object. */
711 ptr_t type_descr = *(ptr_t *)current_p;
712 /* type_descr is either a valid pointer to the descriptor */
713 /* structure, or this object was on a free list. If it */
714 /* it was anything but the last object on the free list, */
715 /* we will misinterpret the next object on the free list as */
716 /* the type descriptor, and get a 0 GC descriptor, which */
717 /* is ideal. Unfortunately, we need to check for the last */
718 /* object case explicitly. */
719 if (0 == type_descr) {
720 /* Rarely executed. */
721 mark_stack_top--;
722 continue;
723 }
724 descr = *(word *)(type_descr
725 - (descr - (GC_DS_PER_OBJECT
726 - GC_INDIR_PER_OBJ_BIAS)));
727 }
728 if (0 == descr) {
729 /* Can happen either because we generated a 0 descriptor */
730 /* or we saw a pointer to a free object. */
731 mark_stack_top--;
732 continue;
733 }
734 goto retry;
735 }
736 } else /* Small object with length descriptor */ {
737 mark_stack_top--;
738 limit = (word *)(((ptr_t)current_p) + (word)descr);
739 }
740 /* The simple case in which we're scanning a range. */
741 GC_ASSERT(!((word)current_p & (ALIGNMENT-1)));
742 credit -= (ptr_t)limit - (ptr_t)current_p;
743 limit -= 1;
744 {
745# define PREF_DIST 4
746
747# ifndef SMALL_CONFIG
748 word deferred;
749
750 /* Try to prefetch the next pointer to be examined asap. */
751 /* Empirically, this also seems to help slightly without */
752 /* prefetches, at least on linux/X86. Presumably this loop */
753 /* ends up with less register pressure, and gcc thus ends up */
754 /* generating slightly better code. Overall gcc code quality */
755 /* for this loop is still not great. */
756 for(;;) {
757 PREFETCH((ptr_t)limit - PREF_DIST*CACHE_LINE_SIZE);
758 GC_ASSERT(limit >= current_p);
759 deferred = *limit;
760 FIXUP_POINTER(deferred);
761 limit = (word *)((char *)limit - ALIGNMENT);
762 if ((ptr_t)deferred >= least_ha && (ptr_t)deferred < greatest_ha) {
763 PREFETCH((ptr_t)deferred);
764 break;
765 }
766 if (current_p > limit) goto next_object;
767 /* Unroll once, so we don't do too many of the prefetches */
768 /* based on limit. */
769 deferred = *limit;
770 FIXUP_POINTER(deferred);
771 limit = (word *)((char *)limit - ALIGNMENT);
772 if ((ptr_t)deferred >= least_ha && (ptr_t)deferred < greatest_ha) {
773 PREFETCH((ptr_t)deferred);
774 break;
775 }
776 if (current_p > limit) goto next_object;
777 }
778# endif
779
780 while (current_p <= limit) {
781 /* Empirically, unrolling this loop doesn't help a lot. */
782 /* Since HC_PUSH_CONTENTS expands to a lot of code, */
783 /* we don't. */
784 current = *current_p;
785 FIXUP_POINTER(current);
786 PREFETCH((ptr_t)current_p + PREF_DIST*CACHE_LINE_SIZE);
787 if ((ptr_t)current >= least_ha && (ptr_t)current < greatest_ha) {
788 /* Prefetch the contents of the object we just pushed. It's */
789 /* likely we will need them soon. */
790 PREFETCH((ptr_t)current);
791 HC_PUSH_CONTENTS((ptr_t)current, mark_stack_top,
792 mark_stack_limit, current_p, exit2);
793 }
794 current_p = (word *)((char *)current_p + ALIGNMENT);
795 }
796
797# ifndef SMALL_CONFIG
798 /* We still need to mark the entry we previously prefetched. */
799 /* We alrady know that it passes the preliminary pointer */
800 /* validity test. */
801 HC_PUSH_CONTENTS((ptr_t)deferred, mark_stack_top,
802 mark_stack_limit, current_p, exit4);
803 next_object:;
804# endif
805 }
806 }
807 return mark_stack_top;
808}
809
810#ifdef PARALLEL_MARK
811
812/* We assume we have an ANSI C Compiler. */
813GC_bool GC_help_wanted = FALSE;
814unsigned GC_helper_count = 0;
815unsigned GC_active_count = 0;
816mse * VOLATILE GC_first_nonempty;
817word GC_mark_no = 0;
818
819#define LOCAL_MARK_STACK_SIZE HBLKSIZE
820 /* Under normal circumstances, this is big enough to guarantee */
821 /* We don't overflow half of it in a single call to */
822 /* GC_mark_from. */
823
824
825/* Steal mark stack entries starting at mse low into mark stack local */
826/* until we either steal mse high, or we have max entries. */
827/* Return a pointer to the top of the local mark stack. */
828/* *next is replaced by a pointer to the next unscanned mark stack */
829/* entry. */
830mse * GC_steal_mark_stack(mse * low, mse * high, mse * local,
831 unsigned max, mse **next)
832{
833 mse *p;
834 mse *top = local - 1;
835 unsigned i = 0;
836
837 /* Make sure that prior writes to the mark stack are visible. */
838 /* On some architectures, the fact that the reads are */
839 /* volatile should suffice. */
840# if !defined(IA64) && !defined(HP_PA) && !defined(I386)
841 GC_memory_barrier();
842# endif
843 GC_ASSERT(high >= low-1 && high - low + 1 <= GC_mark_stack_size);
844 for (p = low; p <= high && i <= max; ++p) {
845 word descr = *(volatile word *) &(p -> mse_descr);
846 /* In the IA64 memory model, the following volatile store is */
847 /* ordered after this read of descr. Thus a thread must read */
848 /* the original nonzero value. HP_PA appears to be similar, */
849 /* and if I'm reading the P4 spec correctly, X86 is probably */
850 /* also OK. In some other cases we need a barrier. */
851# if !defined(IA64) && !defined(HP_PA) && !defined(I386)
852 GC_memory_barrier();
853# endif
854 if (descr != 0) {
855 *(volatile word *) &(p -> mse_descr) = 0;
856 /* More than one thread may get this entry, but that's only */
857 /* a minor performance problem. */
858 ++top;
859 top -> mse_descr = descr;
860 top -> mse_start = p -> mse_start;
861 GC_ASSERT( (top -> mse_descr & GC_DS_TAGS) != GC_DS_LENGTH ||
862 top -> mse_descr < (ptr_t)GC_greatest_plausible_heap_addr
863 - (ptr_t)GC_least_plausible_heap_addr);
864 /* If this is a big object, count it as */
865 /* size/256 + 1 objects. */
866 ++i;
867 if ((descr & GC_DS_TAGS) == GC_DS_LENGTH) i += (descr >> 8);
868 }
869 }
870 *next = p;
871 return top;
872}
873
874/* Copy back a local mark stack. */
875/* low and high are inclusive bounds. */
876void GC_return_mark_stack(mse * low, mse * high)
877{
878 mse * my_top;
879 mse * my_start;
880 size_t stack_size;
881
882 if (high < low) return;
883 stack_size = high - low + 1;
884 GC_acquire_mark_lock();
885 my_top = GC_mark_stack_top;
886 my_start = my_top + 1;
887 if (my_start - GC_mark_stack + stack_size > GC_mark_stack_size) {
888# ifdef CONDPRINT
889 if (GC_print_stats) {
890 GC_printf0("No room to copy back mark stack.");
891 }
892# endif
893 GC_mark_state = MS_INVALID;
894 GC_mark_stack_too_small = TRUE;
895 /* We drop the local mark stack. We'll fix things later. */
896 } else {
897 BCOPY(low, my_start, stack_size * sizeof(mse));
898 GC_ASSERT(GC_mark_stack_top = my_top);
899# if !defined(IA64) && !defined(HP_PA)
900 GC_memory_barrier();
901# endif
902 /* On IA64, the volatile write acts as a release barrier. */
903 GC_mark_stack_top = my_top + stack_size;
904 }
905 GC_release_mark_lock();
906 GC_notify_all_marker();
907}
908
909/* Mark from the local mark stack. */
910/* On return, the local mark stack is empty. */
911/* But this may be achieved by copying the */
912/* local mark stack back into the global one. */
913void GC_do_local_mark(mse *local_mark_stack, mse *local_top)
914{
915 unsigned n;
916# define N_LOCAL_ITERS 1
917
918# ifdef GC_ASSERTIONS
919 /* Make sure we don't hold mark lock. */
920 GC_acquire_mark_lock();
921 GC_release_mark_lock();
922# endif
923 for (;;) {
924 for (n = 0; n < N_LOCAL_ITERS; ++n) {
925 local_top = GC_mark_from(local_top, local_mark_stack,
926 local_mark_stack + LOCAL_MARK_STACK_SIZE);
927 if (local_top < local_mark_stack) return;
928 if (local_top - local_mark_stack >= LOCAL_MARK_STACK_SIZE/2) {
929 GC_return_mark_stack(local_mark_stack, local_top);
930 return;
931 }
932 }
933 if (GC_mark_stack_top < GC_first_nonempty &&
934 GC_active_count < GC_helper_count
935 && local_top > local_mark_stack + 1) {
936 /* Try to share the load, since the main stack is empty, */
937 /* and helper threads are waiting for a refill. */
938 /* The entries near the bottom of the stack are likely */
939 /* to require more work. Thus we return those, eventhough */
940 /* it's harder. */
941 mse * p;
942 mse * new_bottom = local_mark_stack
943 + (local_top - local_mark_stack)/2;
944 GC_ASSERT(new_bottom > local_mark_stack
945 && new_bottom < local_top);
946 GC_return_mark_stack(local_mark_stack, new_bottom - 1);
947 memmove(local_mark_stack, new_bottom,
948 (local_top - new_bottom + 1) * sizeof(mse));
949 local_top -= (new_bottom - local_mark_stack);
950 }
951 }
952}
953
954#define ENTRIES_TO_GET 5
955
956long GC_markers = 2; /* Normally changed by thread-library- */
957 /* -specific code. */
958
959/* Mark using the local mark stack until the global mark stack is empty */
960/* and there are no active workers. Update GC_first_nonempty to reflect */
961/* progress. */
962/* Caller does not hold mark lock. */
963/* Caller has already incremented GC_helper_count. We decrement it, */
964/* and maintain GC_active_count. */
965void GC_mark_local(mse *local_mark_stack, int id)
966{
967 mse * my_first_nonempty;
968
969 GC_acquire_mark_lock();
970 GC_active_count++;
971 my_first_nonempty = GC_first_nonempty;
972 GC_ASSERT(GC_first_nonempty >= GC_mark_stack &&
973 GC_first_nonempty <= GC_mark_stack_top + 1);
974# ifdef PRINTSTATS
975 GC_printf1("Starting mark helper %lu\n", (unsigned long)id);
976# endif
977 GC_release_mark_lock();
978 for (;;) {
979 size_t n_on_stack;
980 size_t n_to_get;
981 mse *next;
982 mse * my_top;
983 mse * local_top;
984 mse * global_first_nonempty = GC_first_nonempty;
985
986 GC_ASSERT(my_first_nonempty >= GC_mark_stack &&
987 my_first_nonempty <= GC_mark_stack_top + 1);
988 GC_ASSERT(global_first_nonempty >= GC_mark_stack &&
989 global_first_nonempty <= GC_mark_stack_top + 1);
990 if (my_first_nonempty < global_first_nonempty) {
991 my_first_nonempty = global_first_nonempty;
992 } else if (global_first_nonempty < my_first_nonempty) {
993 GC_compare_and_exchange((word *)(&GC_first_nonempty),
994 (word) global_first_nonempty,
995 (word) my_first_nonempty);
996 /* If this fails, we just go ahead, without updating */
997 /* GC_first_nonempty. */
998 }
999 /* Perhaps we should also update GC_first_nonempty, if it */
1000 /* is less. But that would require using atomic updates. */
1001 my_top = GC_mark_stack_top;
1002 n_on_stack = my_top - my_first_nonempty + 1;
1003 if (0 == n_on_stack) {
1004 GC_acquire_mark_lock();
1005 my_top = GC_mark_stack_top;
1006 n_on_stack = my_top - my_first_nonempty + 1;
1007 if (0 == n_on_stack) {
1008 GC_active_count--;
1009 GC_ASSERT(GC_active_count <= GC_helper_count);
1010 /* Other markers may redeposit objects */
1011 /* on the stack. */
1012 if (0 == GC_active_count) GC_notify_all_marker();
1013 while (GC_active_count > 0
1014 && GC_first_nonempty > GC_mark_stack_top) {
1015 /* We will be notified if either GC_active_count */
1016 /* reaches zero, or if more objects are pushed on */
1017 /* the global mark stack. */
1018 GC_wait_marker();
1019 }
1020 if (GC_active_count == 0 &&
1021 GC_first_nonempty > GC_mark_stack_top) {
1022 GC_bool need_to_notify = FALSE;
1023 /* The above conditions can't be falsified while we */
1024 /* hold the mark lock, since neither */
1025 /* GC_active_count nor GC_mark_stack_top can */
1026 /* change. GC_first_nonempty can only be */
1027 /* incremented asynchronously. Thus we know that */
1028 /* both conditions actually held simultaneously. */
1029 GC_helper_count--;
1030 if (0 == GC_helper_count) need_to_notify = TRUE;
1031# ifdef PRINTSTATS
1032 GC_printf1(
1033 "Finished mark helper %lu\n", (unsigned long)id);
1034# endif
1035 GC_release_mark_lock();
1036 if (need_to_notify) GC_notify_all_marker();
1037 return;
1038 }
1039 /* else there's something on the stack again, or */
1040 /* another helper may push something. */
1041 GC_active_count++;
1042 GC_ASSERT(GC_active_count > 0);
1043 GC_release_mark_lock();
1044 continue;
1045 } else {
1046 GC_release_mark_lock();
1047 }
1048 }
1049 n_to_get = ENTRIES_TO_GET;
1050 if (n_on_stack < 2 * ENTRIES_TO_GET) n_to_get = 1;
1051 local_top = GC_steal_mark_stack(my_first_nonempty, my_top,
1052 local_mark_stack, n_to_get,
1053 &my_first_nonempty);
1054 GC_ASSERT(my_first_nonempty >= GC_mark_stack &&
1055 my_first_nonempty <= GC_mark_stack_top + 1);
1056 GC_do_local_mark(local_mark_stack, local_top);
1057 }
1058}
1059
1060/* Perform Parallel mark. */
1061/* We hold the GC lock, not the mark lock. */
1062/* Currently runs until the mark stack is */
1063/* empty. */
1064void GC_do_parallel_mark()
1065{
1066 mse local_mark_stack[LOCAL_MARK_STACK_SIZE];
1067 mse * local_top;
1068 mse * my_top;
1069
1070 GC_acquire_mark_lock();
1071 GC_ASSERT(I_HOLD_LOCK());
1072 /* This could be a GC_ASSERT, but it seems safer to keep it on */
1073 /* all the time, especially since it's cheap. */
1074 if (GC_help_wanted || GC_active_count != 0 || GC_helper_count != 0)
1075 ABORT("Tried to start parallel mark in bad state");
1076# ifdef PRINTSTATS
1077 GC_printf1("Starting marking for mark phase number %lu\n",
1078 (unsigned long)GC_mark_no);
1079# endif
1080 GC_first_nonempty = GC_mark_stack;
1081 GC_active_count = 0;
1082 GC_helper_count = 1;
1083 GC_help_wanted = TRUE;
1084 GC_release_mark_lock();
1085 GC_notify_all_marker();
1086 /* Wake up potential helpers. */
1087 GC_mark_local(local_mark_stack, 0);
1088 GC_acquire_mark_lock();
1089 GC_help_wanted = FALSE;
1090 /* Done; clean up. */
1091 while (GC_helper_count > 0) GC_wait_marker();
1092 /* GC_helper_count cannot be incremented while GC_help_wanted == FALSE */
1093# ifdef PRINTSTATS
1094 GC_printf1(
1095 "Finished marking for mark phase number %lu\n",
1096 (unsigned long)GC_mark_no);
1097# endif
1098 GC_mark_no++;
1099 GC_release_mark_lock();
1100 GC_notify_all_marker();
1101}
1102
1103
1104/* Try to help out the marker, if it's running. */
1105/* We do not hold the GC lock, but the requestor does. */
1106void GC_help_marker(word my_mark_no)
1107{
1108 mse local_mark_stack[LOCAL_MARK_STACK_SIZE];
1109 unsigned my_id;
1110 mse * my_first_nonempty;
1111
1112 if (!GC_parallel) return;
1113 GC_acquire_mark_lock();
1114 while (GC_mark_no < my_mark_no
1115 || !GC_help_wanted && GC_mark_no == my_mark_no) {
1116 GC_wait_marker();
1117 }
1118 my_id = GC_helper_count;
1119 if (GC_mark_no != my_mark_no || my_id >= GC_markers) {
1120 /* Second test is useful only if original threads can also */
1121 /* act as helpers. Under Linux they can't. */
1122 GC_release_mark_lock();
1123 return;
1124 }
1125 GC_helper_count = my_id + 1;
1126 GC_release_mark_lock();
1127 GC_mark_local(local_mark_stack, my_id);
1128 /* GC_mark_local decrements GC_helper_count. */
1129}
1130
1131#endif /* PARALLEL_MARK */
1132
1133/* Allocate or reallocate space for mark stack of size s words */
1134/* May silently fail. */
1135static void alloc_mark_stack(n)
1136word n;
1137{
1138 mse * new_stack = (mse *)GC_scratch_alloc(n * sizeof(struct GC_ms_entry));
1139
1140 GC_mark_stack_too_small = FALSE;
1141 if (GC_mark_stack_size != 0) {
1142 if (new_stack != 0) {
1143 word displ = (word)GC_mark_stack & (GC_page_size - 1);
1144 signed_word size = GC_mark_stack_size * sizeof(struct GC_ms_entry);
1145
1146 /* Recycle old space */
1147 if (0 != displ) displ = GC_page_size - displ;
1148 size = (size - displ) & ~(GC_page_size - 1);
1149 if (size > 0) {
1150 GC_add_to_heap((struct hblk *)
1151 ((word)GC_mark_stack + displ), (word)size);
1152 }
1153 GC_mark_stack = new_stack;
1154 GC_mark_stack_size = n;
1155 GC_mark_stack_limit = new_stack + n;
1156# ifdef CONDPRINT
1157 if (GC_print_stats) {
1158 GC_printf1("Grew mark stack to %lu frames\n",
1159 (unsigned long) GC_mark_stack_size);
1160 }
1161# endif
1162 } else {
1163# ifdef CONDPRINT
1164 if (GC_print_stats) {
1165 GC_printf1("Failed to grow mark stack to %lu frames\n",
1166 (unsigned long) n);
1167 }
1168# endif
1169 }
1170 } else {
1171 if (new_stack == 0) {
1172 GC_err_printf0("No space for mark stack\n");
1173 EXIT();
1174 }
1175 GC_mark_stack = new_stack;
1176 GC_mark_stack_size = n;
1177 GC_mark_stack_limit = new_stack + n;
1178 }
1179 GC_mark_stack_top = GC_mark_stack-1;
1180}
1181
1182void GC_mark_init()
1183{
1184 alloc_mark_stack(INITIAL_MARK_STACK_SIZE);
1185}
1186
1187/*
1188 * Push all locations between b and t onto the mark stack.
1189 * b is the first location to be checked. t is one past the last
1190 * location to be checked.
1191 * Should only be used if there is no possibility of mark stack
1192 * overflow.
1193 */
1194void GC_push_all(bottom, top)
1195ptr_t bottom;
1196ptr_t top;
1197{
1198 register word length;
1199
1200 bottom = (ptr_t)(((word) bottom + ALIGNMENT-1) & ~(ALIGNMENT-1));
1201 top = (ptr_t)(((word) top) & ~(ALIGNMENT-1));
1202 if (top == 0 || bottom == top) return;
1203 GC_mark_stack_top++;
1204 if (GC_mark_stack_top >= GC_mark_stack_limit) {
1205 ABORT("unexpected mark stack overflow");
1206 }
1207 length = top - bottom;
1208# if GC_DS_TAGS > ALIGNMENT - 1
1209 length += GC_DS_TAGS;
1210 length &= ~GC_DS_TAGS;
1211# endif
1212 GC_mark_stack_top -> mse_start = (word *)bottom;
1213 GC_mark_stack_top -> mse_descr = length;
1214}
1215
1216/*
1217 * Analogous to the above, but push only those pages h with dirty_fn(h) != 0.
1218 * We use push_fn to actually push the block.
1219 * Used both to selectively push dirty pages, or to push a block
1220 * in piecemeal fashion, to allow for more marking concurrency.
1221 * Will not overflow mark stack if push_fn pushes a small fixed number
1222 * of entries. (This is invoked only if push_fn pushes a single entry,
1223 * or if it marks each object before pushing it, thus ensuring progress
1224 * in the event of a stack overflow.)
1225 */
1226void GC_push_selected(bottom, top, dirty_fn, push_fn)
1227ptr_t bottom;
1228ptr_t top;
1229int (*dirty_fn) GC_PROTO((struct hblk * h));
1230void (*push_fn) GC_PROTO((ptr_t bottom, ptr_t top));
1231{
1232 register struct hblk * h;
1233
1234 bottom = (ptr_t)(((long) bottom + ALIGNMENT-1) & ~(ALIGNMENT-1));
1235 top = (ptr_t)(((long) top) & ~(ALIGNMENT-1));
1236
1237 if (top == 0 || bottom == top) return;
1238 h = HBLKPTR(bottom + HBLKSIZE);
1239 if (top <= (ptr_t) h) {
1240 if ((*dirty_fn)(h-1)) {
1241 (*push_fn)(bottom, top);
1242 }
1243 return;
1244 }
1245 if ((*dirty_fn)(h-1)) {
1246 (*push_fn)(bottom, (ptr_t)h);
1247 }
1248 while ((ptr_t)(h+1) <= top) {
1249 if ((*dirty_fn)(h)) {
1250 if ((word)(GC_mark_stack_top - GC_mark_stack)
1251 > 3 * GC_mark_stack_size / 4) {
1252 /* Danger of mark stack overflow */
1253 (*push_fn)((ptr_t)h, top);
1254 return;
1255 } else {
1256 (*push_fn)((ptr_t)h, (ptr_t)(h+1));
1257 }
1258 }
1259 h++;
1260 }
1261 if ((ptr_t)h != top) {
1262 if ((*dirty_fn)(h)) {
1263 (*push_fn)((ptr_t)h, top);
1264 }
1265 }
1266 if (GC_mark_stack_top >= GC_mark_stack_limit) {
1267 ABORT("unexpected mark stack overflow");
1268 }
1269}
1270
1271# ifndef SMALL_CONFIG
1272
1273#ifdef PARALLEL_MARK
1274 /* Break up root sections into page size chunks to better spread */
1275 /* out work. */
1276 GC_bool GC_true_func(struct hblk *h) { return TRUE; }
1277# define GC_PUSH_ALL(b,t) GC_push_selected(b,t,GC_true_func,GC_push_all);
1278#else
1279# define GC_PUSH_ALL(b,t) GC_push_all(b,t);
1280#endif
1281
1282
1283void GC_push_conditional(bottom, top, all)
1284ptr_t bottom;
1285ptr_t top;
1286int all;
1287{
1288 if (all) {
1289 if (GC_dirty_maintained) {
1290# ifdef PROC_VDB
1291 /* Pages that were never dirtied cannot contain pointers */
1292 GC_push_selected(bottom, top, GC_page_was_ever_dirty, GC_push_all);
1293# else
1294 GC_push_all(bottom, top);
1295# endif
1296 } else {
1297 GC_push_all(bottom, top);
1298 }
1299 } else {
1300 GC_push_selected(bottom, top, GC_page_was_dirty, GC_push_all);
1301 }
1302}
1303#endif
1304
1305# if defined(MSWIN32) || defined(MSWINCE)
1306 void __cdecl GC_push_one(p)
1307# else
1308 void GC_push_one(p)
1309# endif
1310word p;
1311{
1312 GC_PUSH_ONE_STACK(p, MARKED_FROM_REGISTER);
1313}
1314
1315struct GC_ms_entry *GC_mark_and_push(obj, mark_stack_ptr, mark_stack_limit, src)
1316GC_PTR obj;
1317struct GC_ms_entry * mark_stack_ptr;
1318struct GC_ms_entry * mark_stack_limit;
1319GC_PTR *src;
1320{
1321 PREFETCH(obj);
1322 PUSH_CONTENTS(obj, mark_stack_ptr /* modified */, mark_stack_limit, src,
1323 was_marked /* internally generated exit label */);
1324 return mark_stack_ptr;
1325}
1326
1327# ifdef __STDC__
1328# define BASE(p) (word)GC_base((void *)(p))
1329# else
1330# define BASE(p) (word)GC_base((char *)(p))
1331# endif
1332
1333/* Mark and push (i.e. gray) a single object p onto the main */
1334/* mark stack. Consider p to be valid if it is an interior */
1335/* pointer. */
1336/* The object p has passed a preliminary pointer validity */
1337/* test, but we do not definitely know whether it is valid. */
1338/* Mark bits are NOT atomically updated. Thus this must be the */
1339/* only thread setting them. */
1340# if defined(PRINT_BLACK_LIST) || defined(KEEP_BACK_PTRS)
1341 void GC_mark_and_push_stack(p, source)
1342 ptr_t source;
1343# else
1344 void GC_mark_and_push_stack(p)
1345# define source 0
1346# endif
1347register word p;
1348{
1349 register word r;
1350 register hdr * hhdr;
1351 register int displ;
1352
1353 GET_HDR(p, hhdr);
1354 if (IS_FORWARDING_ADDR_OR_NIL(hhdr)) {
1355 if (hhdr != 0) {
1356 r = BASE(p);
1357 hhdr = HDR(r);
1358 displ = BYTES_TO_WORDS(HBLKDISPL(r));
1359 }
1360 } else {
1361 register map_entry_type map_entry;
1362
1363 displ = HBLKDISPL(p);
1364 map_entry = MAP_ENTRY((hhdr -> hb_map), displ);
1365 if (map_entry >= MAX_OFFSET) {
1366 if (map_entry == OFFSET_TOO_BIG || !GC_all_interior_pointers) {
1367 r = BASE(p);
1368 displ = BYTES_TO_WORDS(HBLKDISPL(r));
1369 if (r == 0) hhdr = 0;
1370 } else {
1371 /* Offset invalid, but map reflects interior pointers */
1372 hhdr = 0;
1373 }
1374 } else {
1375 displ = BYTES_TO_WORDS(displ);
1376 displ -= map_entry;
1377 r = (word)((word *)(HBLKPTR(p)) + displ);
1378 }
1379 }
1380 /* If hhdr != 0 then r == GC_base(p), only we did it faster. */
1381 /* displ is the word index within the block. */
1382 if (hhdr == 0) {
1383# ifdef PRINT_BLACK_LIST
1384 GC_add_to_black_list_stack(p, source);
1385# else
1386 GC_add_to_black_list_stack(p);
1387# endif
1388# undef source /* In case we had to define it. */
1389 } else {
1390 if (!mark_bit_from_hdr(hhdr, displ)) {
1391 set_mark_bit_from_hdr(hhdr, displ);
1392 GC_STORE_BACK_PTR(source, (ptr_t)r);
1393 PUSH_OBJ((word *)r, hhdr, GC_mark_stack_top,
1394 GC_mark_stack_limit);
1395 }
1396 }
1397}
1398
1399# ifdef TRACE_BUF
1400
1401# define TRACE_ENTRIES 1000
1402
1403struct trace_entry {
1404 char * kind;
1405 word gc_no;
1406 word words_allocd;
1407 word arg1;
1408 word arg2;
1409} GC_trace_buf[TRACE_ENTRIES];
1410
1411int GC_trace_buf_ptr = 0;
1412
1413void GC_add_trace_entry(char *kind, word arg1, word arg2)
1414{
1415 GC_trace_buf[GC_trace_buf_ptr].kind = kind;
1416 GC_trace_buf[GC_trace_buf_ptr].gc_no = GC_gc_no;
1417 GC_trace_buf[GC_trace_buf_ptr].words_allocd = GC_words_allocd;
1418 GC_trace_buf[GC_trace_buf_ptr].arg1 = arg1 ^ 0x80000000;
1419 GC_trace_buf[GC_trace_buf_ptr].arg2 = arg2 ^ 0x80000000;
1420 GC_trace_buf_ptr++;
1421 if (GC_trace_buf_ptr >= TRACE_ENTRIES) GC_trace_buf_ptr = 0;
1422}
1423
1424void GC_print_trace(word gc_no, GC_bool lock)
1425{
1426 int i;
1427 struct trace_entry *p;
1428
1429 if (lock) LOCK();
1430 for (i = GC_trace_buf_ptr-1; i != GC_trace_buf_ptr; i--) {
1431 if (i < 0) i = TRACE_ENTRIES-1;
1432 p = GC_trace_buf + i;
1433 if (p -> gc_no < gc_no || p -> kind == 0) return;
1434 printf("Trace:%s (gc:%d,words:%d) 0x%X, 0x%X\n",
1435 p -> kind, p -> gc_no, p -> words_allocd,
1436 (p -> arg1) ^ 0x80000000, (p -> arg2) ^ 0x80000000);
1437 }
1438 printf("Trace incomplete\n");
1439 if (lock) UNLOCK();
1440}
1441
1442# endif /* TRACE_BUF */
1443
1444/*
1445 * A version of GC_push_all that treats all interior pointers as valid
1446 * and scans the entire region immediately, in case the contents
1447 * change.
1448 */
1449void GC_push_all_eager(bottom, top)
1450ptr_t bottom;
1451ptr_t top;
1452{
1453 word * b = (word *)(((word) bottom + ALIGNMENT-1) & ~(ALIGNMENT-1));
1454 word * t = (word *)(((word) top) & ~(ALIGNMENT-1));
1455 register word *p;
1456 register word q;
1457 register word *lim;
1458 register ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
1459 register ptr_t least_ha = GC_least_plausible_heap_addr;
1460# define GC_greatest_plausible_heap_addr greatest_ha
1461# define GC_least_plausible_heap_addr least_ha
1462
1463 if (top == 0) return;
1464 /* check all pointers in range and push if they appear */
1465 /* to be valid. */
1466 lim = t - 1 /* longword */;
1467 for (p = b; p <= lim; p = (word *)(((char *)p) + ALIGNMENT)) {
1468 q = *p;
1469 GC_PUSH_ONE_STACK(q, p);
1470 }
1471# undef GC_greatest_plausible_heap_addr
1472# undef GC_least_plausible_heap_addr
1473}
1474
1475#ifndef THREADS
1476/*
1477 * A version of GC_push_all that treats all interior pointers as valid
1478 * and scans part of the area immediately, to make sure that saved
1479 * register values are not lost.
1480 * Cold_gc_frame delimits the stack section that must be scanned
1481 * eagerly. A zero value indicates that no eager scanning is needed.
1482 */
1483void GC_push_all_stack_partially_eager(bottom, top, cold_gc_frame)
1484ptr_t bottom;
1485ptr_t top;
1486ptr_t cold_gc_frame;
1487{
1488 if (!NEED_FIXUP_POINTER && GC_all_interior_pointers) {
1489 /* Push the hot end of the stack eagerly, so that register values */
1490 /* saved inside GC frames are marked before they disappear. */
1491 /* The rest of the marking can be deferred until later. */
1492 if (0 == cold_gc_frame) {
1493 GC_push_all_stack(bottom, top);
1494 return;
1495 }
1496 GC_ASSERT(bottom <= cold_gc_frame && cold_gc_frame <= top);
1497# ifdef STACK_GROWS_DOWN
1498 GC_push_all(cold_gc_frame - sizeof(ptr_t), top);
1499 GC_push_all_eager(bottom, cold_gc_frame);
1500# else /* STACK_GROWS_UP */
1501 GC_push_all(bottom, cold_gc_frame + sizeof(ptr_t));
1502 GC_push_all_eager(cold_gc_frame, top);
1503# endif /* STACK_GROWS_UP */
1504 } else {
1505 GC_push_all_eager(bottom, top);
1506 }
1507# ifdef TRACE_BUF
1508 GC_add_trace_entry("GC_push_all_stack", bottom, top);
1509# endif
1510}
1511#endif /* !THREADS */
1512
1513void GC_push_all_stack(bottom, top)
1514ptr_t bottom;
1515ptr_t top;
1516{
1517 if (!NEED_FIXUP_POINTER && GC_all_interior_pointers) {
1518 GC_push_all(bottom, top);
1519 } else {
1520 GC_push_all_eager(bottom, top);
1521 }
1522}
1523
1524#if !defined(SMALL_CONFIG) && !defined(USE_MARK_BYTES)
1525/* Push all objects reachable from marked objects in the given block */
1526/* of size 1 objects. */
1527void GC_push_marked1(h, hhdr)
1528struct hblk *h;
1529register hdr * hhdr;
1530{
1531 word * mark_word_addr = &(hhdr->hb_marks[0]);
1532 register word *p;
1533 word *plim;
1534 register int i;
1535 register word q;
1536 register word mark_word;
1537 register ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
1538 register ptr_t least_ha = GC_least_plausible_heap_addr;
1539 register mse * mark_stack_top = GC_mark_stack_top;
1540 register mse * mark_stack_limit = GC_mark_stack_limit;
1541# define GC_mark_stack_top mark_stack_top
1542# define GC_mark_stack_limit mark_stack_limit
1543# define GC_greatest_plausible_heap_addr greatest_ha
1544# define GC_least_plausible_heap_addr least_ha
1545
1546 p = (word *)(h->hb_body);
1547 plim = (word *)(((word)h) + HBLKSIZE);
1548
1549 /* go through all words in block */
1550 while( p < plim ) {
1551 mark_word = *mark_word_addr++;
1552 i = 0;
1553 while(mark_word != 0) {
1554 if (mark_word & 1) {
1555 q = p[i];
1556 GC_PUSH_ONE_HEAP(q, p + i);
1557 }
1558 i++;
1559 mark_word >>= 1;
1560 }
1561 p += WORDSZ;
1562 }
1563# undef GC_greatest_plausible_heap_addr
1564# undef GC_least_plausible_heap_addr
1565# undef GC_mark_stack_top
1566# undef GC_mark_stack_limit
1567 GC_mark_stack_top = mark_stack_top;
1568}
1569
1570
1571#ifndef UNALIGNED
1572
1573/* Push all objects reachable from marked objects in the given block */
1574/* of size 2 objects. */
1575void GC_push_marked2(h, hhdr)
1576struct hblk *h;
1577register hdr * hhdr;
1578{
1579 word * mark_word_addr = &(hhdr->hb_marks[0]);
1580 register word *p;
1581 word *plim;
1582 register int i;
1583 register word q;
1584 register word mark_word;
1585 register ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
1586 register ptr_t least_ha = GC_least_plausible_heap_addr;
1587 register mse * mark_stack_top = GC_mark_stack_top;
1588 register mse * mark_stack_limit = GC_mark_stack_limit;
1589# define GC_mark_stack_top mark_stack_top
1590# define GC_mark_stack_limit mark_stack_limit
1591# define GC_greatest_plausible_heap_addr greatest_ha
1592# define GC_least_plausible_heap_addr least_ha
1593
1594 p = (word *)(h->hb_body);
1595 plim = (word *)(((word)h) + HBLKSIZE);
1596
1597 /* go through all words in block */
1598 while( p < plim ) {
1599 mark_word = *mark_word_addr++;
1600 i = 0;
1601 while(mark_word != 0) {
1602 if (mark_word & 1) {
1603 q = p[i];
1604 GC_PUSH_ONE_HEAP(q, p + i);
1605 q = p[i+1];
1606 GC_PUSH_ONE_HEAP(q, p + i);
1607 }
1608 i += 2;
1609 mark_word >>= 2;
1610 }
1611 p += WORDSZ;
1612 }
1613# undef GC_greatest_plausible_heap_addr
1614# undef GC_least_plausible_heap_addr
1615# undef GC_mark_stack_top
1616# undef GC_mark_stack_limit
1617 GC_mark_stack_top = mark_stack_top;
1618}
1619
1620/* Push all objects reachable from marked objects in the given block */
1621/* of size 4 objects. */
1622/* There is a risk of mark stack overflow here. But we handle that. */
1623/* And only unmarked objects get pushed, so it's not very likely. */
1624void GC_push_marked4(h, hhdr)
1625struct hblk *h;
1626register hdr * hhdr;
1627{
1628 word * mark_word_addr = &(hhdr->hb_marks[0]);
1629 register word *p;
1630 word *plim;
1631 register int i;
1632 register word q;
1633 register word mark_word;
1634 register ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
1635 register ptr_t least_ha = GC_least_plausible_heap_addr;
1636 register mse * mark_stack_top = GC_mark_stack_top;
1637 register mse * mark_stack_limit = GC_mark_stack_limit;
1638# define GC_mark_stack_top mark_stack_top
1639# define GC_mark_stack_limit mark_stack_limit
1640# define GC_greatest_plausible_heap_addr greatest_ha
1641# define GC_least_plausible_heap_addr least_ha
1642
1643 p = (word *)(h->hb_body);
1644 plim = (word *)(((word)h) + HBLKSIZE);
1645
1646 /* go through all words in block */
1647 while( p < plim ) {
1648 mark_word = *mark_word_addr++;
1649 i = 0;
1650 while(mark_word != 0) {
1651 if (mark_word & 1) {
1652 q = p[i];
1653 GC_PUSH_ONE_HEAP(q, p + i);
1654 q = p[i+1];
1655 GC_PUSH_ONE_HEAP(q, p + i + 1);
1656 q = p[i+2];
1657 GC_PUSH_ONE_HEAP(q, p + i + 2);
1658 q = p[i+3];
1659 GC_PUSH_ONE_HEAP(q, p + i + 3);
1660 }
1661 i += 4;
1662 mark_word >>= 4;
1663 }
1664 p += WORDSZ;
1665 }
1666# undef GC_greatest_plausible_heap_addr
1667# undef GC_least_plausible_heap_addr
1668# undef GC_mark_stack_top
1669# undef GC_mark_stack_limit
1670 GC_mark_stack_top = mark_stack_top;
1671}
1672
1673#endif /* UNALIGNED */
1674
1675#endif /* SMALL_CONFIG */
1676
1677/* Push all objects reachable from marked objects in the given block */
1678void GC_push_marked(h, hhdr)
1679struct hblk *h;
1680register hdr * hhdr;
1681{
1682 register int sz = hhdr -> hb_sz;
1683 register int descr = hhdr -> hb_descr;
1684 register word * p;
1685 register int word_no;
1686 register word * lim;
1687 register mse * GC_mark_stack_top_reg;
1688 register mse * mark_stack_limit = GC_mark_stack_limit;
1689
1690 /* Some quick shortcuts: */
1691 if ((0 | GC_DS_LENGTH) == descr) return;
1692 if (GC_block_empty(hhdr)/* nothing marked */) return;
1693 GC_n_rescuing_pages++;
1694 GC_objects_are_marked = TRUE;
1695 if (sz > MAXOBJSZ) {
1696 lim = (word *)h;
1697 } else {
1698 lim = (word *)(h + 1) - sz;
1699 }
1700
1701 switch(sz) {
1702# if !defined(SMALL_CONFIG) && !defined(USE_MARK_BYTES)
1703 case 1:
1704 GC_push_marked1(h, hhdr);
1705 break;
1706# endif
1707# if !defined(SMALL_CONFIG) && !defined(UNALIGNED) && \
1708 !defined(USE_MARK_BYTES)
1709 case 2:
1710 GC_push_marked2(h, hhdr);
1711 break;
1712 case 4:
1713 GC_push_marked4(h, hhdr);
1714 break;
1715# endif
1716 default:
1717 GC_mark_stack_top_reg = GC_mark_stack_top;
1718 for (p = (word *)h, word_no = 0; p <= lim; p += sz, word_no += sz) {
1719 if (mark_bit_from_hdr(hhdr, word_no)) {
1720 /* Mark from fields inside the object */
1721 PUSH_OBJ((word *)p, hhdr, GC_mark_stack_top_reg, mark_stack_limit);
1722# ifdef GATHERSTATS
1723 /* Subtract this object from total, since it was */
1724 /* added in twice. */
1725 GC_composite_in_use -= sz;
1726# endif
1727 }
1728 }
1729 GC_mark_stack_top = GC_mark_stack_top_reg;
1730 }
1731}
1732
1733#ifndef SMALL_CONFIG
1734/* Test whether any page in the given block is dirty */
1735GC_bool GC_block_was_dirty(h, hhdr)
1736struct hblk *h;
1737register hdr * hhdr;
1738{
1739 register int sz = hhdr -> hb_sz;
1740
1741 if (sz <= MAXOBJSZ) {
1742 return(GC_page_was_dirty(h));
1743 } else {
1744 register ptr_t p = (ptr_t)h;
1745 sz = WORDS_TO_BYTES(sz);
1746 while (p < (ptr_t)h + sz) {
1747 if (GC_page_was_dirty((struct hblk *)p)) return(TRUE);
1748 p += HBLKSIZE;
1749 }
1750 return(FALSE);
1751 }
1752}
1753#endif /* SMALL_CONFIG */
1754
1755/* Similar to GC_push_next_marked, but return address of next block */
1756struct hblk * GC_push_next_marked(h)
1757struct hblk *h;
1758{
1759 register hdr * hhdr;
1760
1761 h = GC_next_used_block(h);
1762 if (h == 0) return(0);
1763 hhdr = HDR(h);
1764 GC_push_marked(h, hhdr);
1765 return(h + OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz));
1766}
1767
1768#ifndef SMALL_CONFIG
1769/* Identical to above, but mark only from dirty pages */
1770struct hblk * GC_push_next_marked_dirty(h)
1771struct hblk *h;
1772{
1773 register hdr * hhdr;
1774
1775 if (!GC_dirty_maintained) { ABORT("dirty bits not set up"); }
1776 for (;;) {
1777 h = GC_next_used_block(h);
1778 if (h == 0) return(0);
1779 hhdr = HDR(h);
1780# ifdef STUBBORN_ALLOC
1781 if (hhdr -> hb_obj_kind == STUBBORN) {
1782 if (GC_page_was_changed(h) && GC_block_was_dirty(h, hhdr)) {
1783 break;
1784 }
1785 } else {
1786 if (GC_block_was_dirty(h, hhdr)) break;
1787 }
1788# else
1789 if (GC_block_was_dirty(h, hhdr)) break;
1790# endif
1791 h += OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz);
1792 }
1793 GC_push_marked(h, hhdr);
1794 return(h + OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz));
1795}
1796#endif
1797
1798/* Similar to above, but for uncollectable pages. Needed since we */
1799/* do not clear marks for such pages, even for full collections. */
1800struct hblk * GC_push_next_marked_uncollectable(h)
1801struct hblk *h;
1802{
1803 register hdr * hhdr = HDR(h);
1804
1805 for (;;) {
1806 h = GC_next_used_block(h);
1807 if (h == 0) return(0);
1808 hhdr = HDR(h);
1809 if (hhdr -> hb_obj_kind == UNCOLLECTABLE) break;
1810 h += OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz);
1811 }
1812 GC_push_marked(h, hhdr);
1813 return(h + OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz));
1814}
1815
1816
Note: See TracBrowser for help on using the repository browser.