0
|
1 /*
|
|
2 * Copyright 2005-2007 Sun Microsystems, Inc. All Rights Reserved.
|
|
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
|
|
4 *
|
|
5 * This code is free software; you can redistribute it and/or modify it
|
|
6 * under the terms of the GNU General Public License version 2 only, as
|
|
7 * published by the Free Software Foundation.
|
|
8 *
|
|
9 * This code is distributed in the hope that it will be useful, but WITHOUT
|
|
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
12 * version 2 for more details (a copy is included in the LICENSE file that
|
|
13 * accompanied this code).
|
|
14 *
|
|
15 * You should have received a copy of the GNU General Public License version
|
|
16 * 2 along with this work; if not, write to the Free Software Foundation,
|
|
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
|
|
18 *
|
|
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
|
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or
|
|
21 * have any questions.
|
|
22 *
|
|
23 */
|
|
24
|
|
25 #include "incls/_precompiled.incl"
|
|
26 #include "incls/_psParallelCompact.cpp.incl"
|
|
27
|
|
28 #include <math.h>
|
|
29
|
|
30 // All sizes are in HeapWords.
|
|
31 const size_t ParallelCompactData::Log2ChunkSize = 9; // 512 words
|
|
32 const size_t ParallelCompactData::ChunkSize = (size_t)1 << Log2ChunkSize;
|
|
33 const size_t ParallelCompactData::ChunkSizeBytes = ChunkSize << LogHeapWordSize;
|
|
34 const size_t ParallelCompactData::ChunkSizeOffsetMask = ChunkSize - 1;
|
|
35 const size_t ParallelCompactData::ChunkAddrOffsetMask = ChunkSizeBytes - 1;
|
|
36 const size_t ParallelCompactData::ChunkAddrMask = ~ChunkAddrOffsetMask;
|
|
37
|
|
38 // 32-bit: 128 words covers 4 bitmap words
|
|
39 // 64-bit: 128 words covers 2 bitmap words
|
|
40 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words
|
|
41 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize;
|
|
42 const size_t ParallelCompactData::BlockOffsetMask = BlockSize - 1;
|
|
43 const size_t ParallelCompactData::BlockMask = ~BlockOffsetMask;
|
|
44
|
|
45 const size_t ParallelCompactData::BlocksPerChunk = ChunkSize / BlockSize;
|
|
46
|
|
47 const ParallelCompactData::ChunkData::chunk_sz_t
|
|
48 ParallelCompactData::ChunkData::dc_shift = 27;
|
|
49
|
|
50 const ParallelCompactData::ChunkData::chunk_sz_t
|
|
51 ParallelCompactData::ChunkData::dc_mask = ~0U << dc_shift;
|
|
52
|
|
53 const ParallelCompactData::ChunkData::chunk_sz_t
|
|
54 ParallelCompactData::ChunkData::dc_one = 0x1U << dc_shift;
|
|
55
|
|
56 const ParallelCompactData::ChunkData::chunk_sz_t
|
|
57 ParallelCompactData::ChunkData::los_mask = ~dc_mask;
|
|
58
|
|
59 const ParallelCompactData::ChunkData::chunk_sz_t
|
|
60 ParallelCompactData::ChunkData::dc_claimed = 0x8U << dc_shift;
|
|
61
|
|
62 const ParallelCompactData::ChunkData::chunk_sz_t
|
|
63 ParallelCompactData::ChunkData::dc_completed = 0xcU << dc_shift;
|
|
64
|
|
65 #ifdef ASSERT
|
|
66 short ParallelCompactData::BlockData::_cur_phase = 0;
|
|
67 #endif
|
|
68
|
|
69 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
|
|
70 bool PSParallelCompact::_print_phases = false;
|
|
71
|
|
72 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
|
|
73 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL;
|
|
74
|
|
75 double PSParallelCompact::_dwl_mean;
|
|
76 double PSParallelCompact::_dwl_std_dev;
|
|
77 double PSParallelCompact::_dwl_first_term;
|
|
78 double PSParallelCompact::_dwl_adjustment;
|
|
79 #ifdef ASSERT
|
|
80 bool PSParallelCompact::_dwl_initialized = false;
|
|
81 #endif // #ifdef ASSERT
|
|
82
|
|
83 #ifdef VALIDATE_MARK_SWEEP
|
|
84 GrowableArray<oop*>* PSParallelCompact::_root_refs_stack = NULL;
|
|
85 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL;
|
|
86 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL;
|
|
87 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL;
|
|
88 size_t PSParallelCompact::_live_oops_index = 0;
|
|
89 size_t PSParallelCompact::_live_oops_index_at_perm = 0;
|
|
90 GrowableArray<oop*>* PSParallelCompact::_other_refs_stack = NULL;
|
|
91 GrowableArray<oop*>* PSParallelCompact::_adjusted_pointers = NULL;
|
|
92 bool PSParallelCompact::_pointer_tracking = false;
|
|
93 bool PSParallelCompact::_root_tracking = true;
|
|
94
|
|
95 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL;
|
|
96 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL;
|
|
97 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL;
|
|
98 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL;
|
|
99 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL;
|
|
100 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL;
|
|
101 #endif
|
|
102
|
|
103 // XXX beg - verification code; only works while we also mark in object headers
|
|
104 static void
|
|
105 verify_mark_bitmap(ParMarkBitMap& _mark_bitmap)
|
|
106 {
|
|
107 ParallelScavengeHeap* heap = PSParallelCompact::gc_heap();
|
|
108
|
|
109 PSPermGen* perm_gen = heap->perm_gen();
|
|
110 PSOldGen* old_gen = heap->old_gen();
|
|
111 PSYoungGen* young_gen = heap->young_gen();
|
|
112
|
|
113 MutableSpace* perm_space = perm_gen->object_space();
|
|
114 MutableSpace* old_space = old_gen->object_space();
|
|
115 MutableSpace* eden_space = young_gen->eden_space();
|
|
116 MutableSpace* from_space = young_gen->from_space();
|
|
117 MutableSpace* to_space = young_gen->to_space();
|
|
118
|
|
119 // 'from_space' here is the survivor space at the lower address.
|
|
120 if (to_space->bottom() < from_space->bottom()) {
|
|
121 from_space = to_space;
|
|
122 to_space = young_gen->from_space();
|
|
123 }
|
|
124
|
|
125 HeapWord* boundaries[12];
|
|
126 unsigned int bidx = 0;
|
|
127 const unsigned int bidx_max = sizeof(boundaries) / sizeof(boundaries[0]);
|
|
128
|
|
129 boundaries[0] = perm_space->bottom();
|
|
130 boundaries[1] = perm_space->top();
|
|
131 boundaries[2] = old_space->bottom();
|
|
132 boundaries[3] = old_space->top();
|
|
133 boundaries[4] = eden_space->bottom();
|
|
134 boundaries[5] = eden_space->top();
|
|
135 boundaries[6] = from_space->bottom();
|
|
136 boundaries[7] = from_space->top();
|
|
137 boundaries[8] = to_space->bottom();
|
|
138 boundaries[9] = to_space->top();
|
|
139 boundaries[10] = to_space->end();
|
|
140 boundaries[11] = to_space->end();
|
|
141
|
|
142 BitMap::idx_t beg_bit = 0;
|
|
143 BitMap::idx_t end_bit;
|
|
144 BitMap::idx_t tmp_bit;
|
|
145 const BitMap::idx_t last_bit = _mark_bitmap.size();
|
|
146 do {
|
|
147 HeapWord* addr = _mark_bitmap.bit_to_addr(beg_bit);
|
|
148 if (_mark_bitmap.is_marked(beg_bit)) {
|
|
149 oop obj = (oop)addr;
|
|
150 assert(obj->is_gc_marked(), "obj header is not marked");
|
|
151 end_bit = _mark_bitmap.find_obj_end(beg_bit, last_bit);
|
|
152 const size_t size = _mark_bitmap.obj_size(beg_bit, end_bit);
|
|
153 assert(size == (size_t)obj->size(), "end bit wrong?");
|
|
154 beg_bit = _mark_bitmap.find_obj_beg(beg_bit + 1, last_bit);
|
|
155 assert(beg_bit > end_bit, "bit set in middle of an obj");
|
|
156 } else {
|
|
157 if (addr >= boundaries[bidx] && addr < boundaries[bidx + 1]) {
|
|
158 // a dead object in the current space.
|
|
159 oop obj = (oop)addr;
|
|
160 end_bit = _mark_bitmap.addr_to_bit(addr + obj->size());
|
|
161 assert(!obj->is_gc_marked(), "obj marked in header, not in bitmap");
|
|
162 tmp_bit = beg_bit + 1;
|
|
163 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, end_bit);
|
|
164 assert(beg_bit == end_bit, "beg bit set in unmarked obj");
|
|
165 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, end_bit);
|
|
166 assert(beg_bit == end_bit, "end bit set in unmarked obj");
|
|
167 } else if (addr < boundaries[bidx + 2]) {
|
|
168 // addr is between top in the current space and bottom in the next.
|
|
169 end_bit = beg_bit + pointer_delta(boundaries[bidx + 2], addr);
|
|
170 tmp_bit = beg_bit;
|
|
171 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, end_bit);
|
|
172 assert(beg_bit == end_bit, "beg bit set above top");
|
|
173 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, end_bit);
|
|
174 assert(beg_bit == end_bit, "end bit set above top");
|
|
175 bidx += 2;
|
|
176 } else if (bidx < bidx_max - 2) {
|
|
177 bidx += 2; // ???
|
|
178 } else {
|
|
179 tmp_bit = beg_bit;
|
|
180 beg_bit = _mark_bitmap.find_obj_beg(tmp_bit, last_bit);
|
|
181 assert(beg_bit == last_bit, "beg bit set outside heap");
|
|
182 beg_bit = _mark_bitmap.find_obj_end(tmp_bit, last_bit);
|
|
183 assert(beg_bit == last_bit, "end bit set outside heap");
|
|
184 }
|
|
185 }
|
|
186 } while (beg_bit < last_bit);
|
|
187 }
|
|
188 // XXX end - verification code; only works while we also mark in object headers
|
|
189
|
|
190 #ifndef PRODUCT
|
|
191 const char* PSParallelCompact::space_names[] = {
|
|
192 "perm", "old ", "eden", "from", "to "
|
|
193 };
|
|
194
|
|
195 void PSParallelCompact::print_chunk_ranges()
|
|
196 {
|
|
197 tty->print_cr("space bottom top end new_top");
|
|
198 tty->print_cr("------ ---------- ---------- ---------- ----------");
|
|
199
|
|
200 for (unsigned int id = 0; id < last_space_id; ++id) {
|
|
201 const MutableSpace* space = _space_info[id].space();
|
|
202 tty->print_cr("%u %s "
|
|
203 SIZE_FORMAT_W("10") " " SIZE_FORMAT_W("10") " "
|
|
204 SIZE_FORMAT_W("10") " " SIZE_FORMAT_W("10") " ",
|
|
205 id, space_names[id],
|
|
206 summary_data().addr_to_chunk_idx(space->bottom()),
|
|
207 summary_data().addr_to_chunk_idx(space->top()),
|
|
208 summary_data().addr_to_chunk_idx(space->end()),
|
|
209 summary_data().addr_to_chunk_idx(_space_info[id].new_top()));
|
|
210 }
|
|
211 }
|
|
212
|
|
213 void
|
|
214 print_generic_summary_chunk(size_t i, const ParallelCompactData::ChunkData* c)
|
|
215 {
|
|
216 #define CHUNK_IDX_FORMAT SIZE_FORMAT_W("7")
|
|
217 #define CHUNK_DATA_FORMAT SIZE_FORMAT_W("5")
|
|
218
|
|
219 ParallelCompactData& sd = PSParallelCompact::summary_data();
|
|
220 size_t dci = c->destination() ? sd.addr_to_chunk_idx(c->destination()) : 0;
|
|
221 tty->print_cr(CHUNK_IDX_FORMAT " " PTR_FORMAT " "
|
|
222 CHUNK_IDX_FORMAT " " PTR_FORMAT " "
|
|
223 CHUNK_DATA_FORMAT " " CHUNK_DATA_FORMAT " "
|
|
224 CHUNK_DATA_FORMAT " " CHUNK_IDX_FORMAT " %d",
|
|
225 i, c->data_location(), dci, c->destination(),
|
|
226 c->partial_obj_size(), c->live_obj_size(),
|
|
227 c->data_size(), c->source_chunk(), c->destination_count());
|
|
228
|
|
229 #undef CHUNK_IDX_FORMAT
|
|
230 #undef CHUNK_DATA_FORMAT
|
|
231 }
|
|
232
|
|
233 void
|
|
234 print_generic_summary_data(ParallelCompactData& summary_data,
|
|
235 HeapWord* const beg_addr,
|
|
236 HeapWord* const end_addr)
|
|
237 {
|
|
238 size_t total_words = 0;
|
|
239 size_t i = summary_data.addr_to_chunk_idx(beg_addr);
|
|
240 const size_t last = summary_data.addr_to_chunk_idx(end_addr);
|
|
241 HeapWord* pdest = 0;
|
|
242
|
|
243 while (i <= last) {
|
|
244 ParallelCompactData::ChunkData* c = summary_data.chunk(i);
|
|
245 if (c->data_size() != 0 || c->destination() != pdest) {
|
|
246 print_generic_summary_chunk(i, c);
|
|
247 total_words += c->data_size();
|
|
248 pdest = c->destination();
|
|
249 }
|
|
250 ++i;
|
|
251 }
|
|
252
|
|
253 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
|
|
254 }
|
|
255
|
|
256 void
|
|
257 print_generic_summary_data(ParallelCompactData& summary_data,
|
|
258 SpaceInfo* space_info)
|
|
259 {
|
|
260 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
|
|
261 const MutableSpace* space = space_info[id].space();
|
|
262 print_generic_summary_data(summary_data, space->bottom(),
|
|
263 MAX2(space->top(), space_info[id].new_top()));
|
|
264 }
|
|
265 }
|
|
266
|
|
267 void
|
|
268 print_initial_summary_chunk(size_t i,
|
|
269 const ParallelCompactData::ChunkData* c,
|
|
270 bool newline = true)
|
|
271 {
|
|
272 tty->print(SIZE_FORMAT_W("5") " " PTR_FORMAT " "
|
|
273 SIZE_FORMAT_W("5") " " SIZE_FORMAT_W("5") " "
|
|
274 SIZE_FORMAT_W("5") " " SIZE_FORMAT_W("5") " %d",
|
|
275 i, c->destination(),
|
|
276 c->partial_obj_size(), c->live_obj_size(),
|
|
277 c->data_size(), c->source_chunk(), c->destination_count());
|
|
278 if (newline) tty->cr();
|
|
279 }
|
|
280
|
|
281 void
|
|
282 print_initial_summary_data(ParallelCompactData& summary_data,
|
|
283 const MutableSpace* space) {
|
|
284 if (space->top() == space->bottom()) {
|
|
285 return;
|
|
286 }
|
|
287
|
|
288 const size_t chunk_size = ParallelCompactData::ChunkSize;
|
|
289 HeapWord* const top_aligned_up = summary_data.chunk_align_up(space->top());
|
|
290 const size_t end_chunk = summary_data.addr_to_chunk_idx(top_aligned_up);
|
|
291 const ParallelCompactData::ChunkData* c = summary_data.chunk(end_chunk - 1);
|
|
292 HeapWord* end_addr = c->destination() + c->data_size();
|
|
293 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
|
|
294
|
|
295 // Print (and count) the full chunks at the beginning of the space.
|
|
296 size_t full_chunk_count = 0;
|
|
297 size_t i = summary_data.addr_to_chunk_idx(space->bottom());
|
|
298 while (i < end_chunk && summary_data.chunk(i)->data_size() == chunk_size) {
|
|
299 print_initial_summary_chunk(i, summary_data.chunk(i));
|
|
300 ++full_chunk_count;
|
|
301 ++i;
|
|
302 }
|
|
303
|
|
304 size_t live_to_right = live_in_space - full_chunk_count * chunk_size;
|
|
305
|
|
306 double max_reclaimed_ratio = 0.0;
|
|
307 size_t max_reclaimed_ratio_chunk = 0;
|
|
308 size_t max_dead_to_right = 0;
|
|
309 size_t max_live_to_right = 0;
|
|
310
|
|
311 // Print the 'reclaimed ratio' for chunks while there is something live in the
|
|
312 // chunk or to the right of it. The remaining chunks are empty (and
|
|
313 // uninteresting), and computing the ratio will result in division by 0.
|
|
314 while (i < end_chunk && live_to_right > 0) {
|
|
315 c = summary_data.chunk(i);
|
|
316 HeapWord* const chunk_addr = summary_data.chunk_to_addr(i);
|
|
317 const size_t used_to_right = pointer_delta(space->top(), chunk_addr);
|
|
318 const size_t dead_to_right = used_to_right - live_to_right;
|
|
319 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
|
|
320
|
|
321 if (reclaimed_ratio > max_reclaimed_ratio) {
|
|
322 max_reclaimed_ratio = reclaimed_ratio;
|
|
323 max_reclaimed_ratio_chunk = i;
|
|
324 max_dead_to_right = dead_to_right;
|
|
325 max_live_to_right = live_to_right;
|
|
326 }
|
|
327
|
|
328 print_initial_summary_chunk(i, c, false);
|
|
329 tty->print_cr(" %12.10f " SIZE_FORMAT_W("10") " " SIZE_FORMAT_W("10"),
|
|
330 reclaimed_ratio, dead_to_right, live_to_right);
|
|
331
|
|
332 live_to_right -= c->data_size();
|
|
333 ++i;
|
|
334 }
|
|
335
|
|
336 // Any remaining chunks are empty. Print one more if there is one.
|
|
337 if (i < end_chunk) {
|
|
338 print_initial_summary_chunk(i, summary_data.chunk(i));
|
|
339 }
|
|
340
|
|
341 tty->print_cr("max: " SIZE_FORMAT_W("4") " d2r=" SIZE_FORMAT_W("10") " "
|
|
342 "l2r=" SIZE_FORMAT_W("10") " max_ratio=%14.12f",
|
|
343 max_reclaimed_ratio_chunk, max_dead_to_right,
|
|
344 max_live_to_right, max_reclaimed_ratio);
|
|
345 }
|
|
346
|
|
347 void
|
|
348 print_initial_summary_data(ParallelCompactData& summary_data,
|
|
349 SpaceInfo* space_info) {
|
|
350 unsigned int id = PSParallelCompact::perm_space_id;
|
|
351 const MutableSpace* space;
|
|
352 do {
|
|
353 space = space_info[id].space();
|
|
354 print_initial_summary_data(summary_data, space);
|
|
355 } while (++id < PSParallelCompact::eden_space_id);
|
|
356
|
|
357 do {
|
|
358 space = space_info[id].space();
|
|
359 print_generic_summary_data(summary_data, space->bottom(), space->top());
|
|
360 } while (++id < PSParallelCompact::last_space_id);
|
|
361 }
|
|
362 #endif // #ifndef PRODUCT
|
|
363
|
|
364 #ifdef ASSERT
|
|
365 size_t add_obj_count;
|
|
366 size_t add_obj_size;
|
|
367 size_t mark_bitmap_count;
|
|
368 size_t mark_bitmap_size;
|
|
369 #endif // #ifdef ASSERT
|
|
370
|
|
371 ParallelCompactData::ParallelCompactData()
|
|
372 {
|
|
373 _region_start = 0;
|
|
374
|
|
375 _chunk_vspace = 0;
|
|
376 _chunk_data = 0;
|
|
377 _chunk_count = 0;
|
|
378
|
|
379 _block_vspace = 0;
|
|
380 _block_data = 0;
|
|
381 _block_count = 0;
|
|
382 }
|
|
383
|
|
384 bool ParallelCompactData::initialize(MemRegion covered_region)
|
|
385 {
|
|
386 _region_start = covered_region.start();
|
|
387 const size_t region_size = covered_region.word_size();
|
|
388 DEBUG_ONLY(_region_end = _region_start + region_size;)
|
|
389
|
|
390 assert(chunk_align_down(_region_start) == _region_start,
|
|
391 "region start not aligned");
|
|
392 assert((region_size & ChunkSizeOffsetMask) == 0,
|
|
393 "region size not a multiple of ChunkSize");
|
|
394
|
|
395 bool result = initialize_chunk_data(region_size);
|
|
396
|
|
397 // Initialize the block data if it will be used for updating pointers, or if
|
|
398 // this is a debug build.
|
|
399 if (!UseParallelOldGCChunkPointerCalc || trueInDebug) {
|
|
400 result = result && initialize_block_data(region_size);
|
|
401 }
|
|
402
|
|
403 return result;
|
|
404 }
|
|
405
|
|
406 PSVirtualSpace*
|
|
407 ParallelCompactData::create_vspace(size_t count, size_t element_size)
|
|
408 {
|
|
409 const size_t raw_bytes = count * element_size;
|
|
410 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
|
|
411 const size_t granularity = os::vm_allocation_granularity();
|
|
412 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
|
|
413
|
|
414 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
|
|
415 MAX2(page_sz, granularity);
|
|
416 ReservedSpace rs(bytes, rs_align, false);
|
|
417 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
|
|
418 rs.size());
|
|
419 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
|
|
420 if (vspace != 0) {
|
|
421 if (vspace->expand_by(bytes)) {
|
|
422 return vspace;
|
|
423 }
|
|
424 delete vspace;
|
|
425 }
|
|
426
|
|
427 return 0;
|
|
428 }
|
|
429
|
|
430 bool ParallelCompactData::initialize_chunk_data(size_t region_size)
|
|
431 {
|
|
432 const size_t count = (region_size + ChunkSizeOffsetMask) >> Log2ChunkSize;
|
|
433 _chunk_vspace = create_vspace(count, sizeof(ChunkData));
|
|
434 if (_chunk_vspace != 0) {
|
|
435 _chunk_data = (ChunkData*)_chunk_vspace->reserved_low_addr();
|
|
436 _chunk_count = count;
|
|
437 return true;
|
|
438 }
|
|
439 return false;
|
|
440 }
|
|
441
|
|
442 bool ParallelCompactData::initialize_block_data(size_t region_size)
|
|
443 {
|
|
444 const size_t count = (region_size + BlockOffsetMask) >> Log2BlockSize;
|
|
445 _block_vspace = create_vspace(count, sizeof(BlockData));
|
|
446 if (_block_vspace != 0) {
|
|
447 _block_data = (BlockData*)_block_vspace->reserved_low_addr();
|
|
448 _block_count = count;
|
|
449 return true;
|
|
450 }
|
|
451 return false;
|
|
452 }
|
|
453
|
|
454 void ParallelCompactData::clear()
|
|
455 {
|
|
456 if (_block_data) {
|
|
457 memset(_block_data, 0, _block_vspace->committed_size());
|
|
458 }
|
|
459 memset(_chunk_data, 0, _chunk_vspace->committed_size());
|
|
460 }
|
|
461
|
|
462 void ParallelCompactData::clear_range(size_t beg_chunk, size_t end_chunk) {
|
|
463 assert(beg_chunk <= _chunk_count, "beg_chunk out of range");
|
|
464 assert(end_chunk <= _chunk_count, "end_chunk out of range");
|
|
465 assert(ChunkSize % BlockSize == 0, "ChunkSize not a multiple of BlockSize");
|
|
466
|
|
467 const size_t chunk_cnt = end_chunk - beg_chunk;
|
|
468
|
|
469 if (_block_data) {
|
|
470 const size_t blocks_per_chunk = ChunkSize / BlockSize;
|
|
471 const size_t beg_block = beg_chunk * blocks_per_chunk;
|
|
472 const size_t block_cnt = chunk_cnt * blocks_per_chunk;
|
|
473 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
|
|
474 }
|
|
475 memset(_chunk_data + beg_chunk, 0, chunk_cnt * sizeof(ChunkData));
|
|
476 }
|
|
477
|
|
478 HeapWord* ParallelCompactData::partial_obj_end(size_t chunk_idx) const
|
|
479 {
|
|
480 const ChunkData* cur_cp = chunk(chunk_idx);
|
|
481 const ChunkData* const end_cp = chunk(chunk_count() - 1);
|
|
482
|
|
483 HeapWord* result = chunk_to_addr(chunk_idx);
|
|
484 if (cur_cp < end_cp) {
|
|
485 do {
|
|
486 result += cur_cp->partial_obj_size();
|
|
487 } while (cur_cp->partial_obj_size() == ChunkSize && ++cur_cp < end_cp);
|
|
488 }
|
|
489 return result;
|
|
490 }
|
|
491
|
|
492 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
|
|
493 {
|
|
494 const size_t obj_ofs = pointer_delta(addr, _region_start);
|
|
495 const size_t beg_chunk = obj_ofs >> Log2ChunkSize;
|
|
496 const size_t end_chunk = (obj_ofs + len - 1) >> Log2ChunkSize;
|
|
497
|
|
498 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
|
|
499 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
|
|
500
|
|
501 if (beg_chunk == end_chunk) {
|
|
502 // All in one chunk.
|
|
503 _chunk_data[beg_chunk].add_live_obj(len);
|
|
504 return;
|
|
505 }
|
|
506
|
|
507 // First chunk.
|
|
508 const size_t beg_ofs = chunk_offset(addr);
|
|
509 _chunk_data[beg_chunk].add_live_obj(ChunkSize - beg_ofs);
|
|
510
|
|
511 klassOop klass = ((oop)addr)->klass();
|
|
512 // Middle chunks--completely spanned by this object.
|
|
513 for (size_t chunk = beg_chunk + 1; chunk < end_chunk; ++chunk) {
|
|
514 _chunk_data[chunk].set_partial_obj_size(ChunkSize);
|
|
515 _chunk_data[chunk].set_partial_obj_addr(addr);
|
|
516 }
|
|
517
|
|
518 // Last chunk.
|
|
519 const size_t end_ofs = chunk_offset(addr + len - 1);
|
|
520 _chunk_data[end_chunk].set_partial_obj_size(end_ofs + 1);
|
|
521 _chunk_data[end_chunk].set_partial_obj_addr(addr);
|
|
522 }
|
|
523
|
|
524 void
|
|
525 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
|
|
526 {
|
|
527 assert(chunk_offset(beg) == 0, "not ChunkSize aligned");
|
|
528 assert(chunk_offset(end) == 0, "not ChunkSize aligned");
|
|
529
|
|
530 size_t cur_chunk = addr_to_chunk_idx(beg);
|
|
531 const size_t end_chunk = addr_to_chunk_idx(end);
|
|
532 HeapWord* addr = beg;
|
|
533 while (cur_chunk < end_chunk) {
|
|
534 _chunk_data[cur_chunk].set_destination(addr);
|
|
535 _chunk_data[cur_chunk].set_destination_count(0);
|
|
536 _chunk_data[cur_chunk].set_source_chunk(cur_chunk);
|
|
537 _chunk_data[cur_chunk].set_data_location(addr);
|
|
538
|
|
539 // Update live_obj_size so the chunk appears completely full.
|
|
540 size_t live_size = ChunkSize - _chunk_data[cur_chunk].partial_obj_size();
|
|
541 _chunk_data[cur_chunk].set_live_obj_size(live_size);
|
|
542
|
|
543 ++cur_chunk;
|
|
544 addr += ChunkSize;
|
|
545 }
|
|
546 }
|
|
547
|
|
548 bool ParallelCompactData::summarize(HeapWord* target_beg, HeapWord* target_end,
|
|
549 HeapWord* source_beg, HeapWord* source_end,
|
|
550 HeapWord** target_next,
|
|
551 HeapWord** source_next) {
|
|
552 // This is too strict.
|
|
553 // assert(chunk_offset(source_beg) == 0, "not ChunkSize aligned");
|
|
554
|
|
555 if (TraceParallelOldGCSummaryPhase) {
|
|
556 tty->print_cr("tb=" PTR_FORMAT " te=" PTR_FORMAT " "
|
|
557 "sb=" PTR_FORMAT " se=" PTR_FORMAT " "
|
|
558 "tn=" PTR_FORMAT " sn=" PTR_FORMAT,
|
|
559 target_beg, target_end,
|
|
560 source_beg, source_end,
|
|
561 target_next != 0 ? *target_next : (HeapWord*) 0,
|
|
562 source_next != 0 ? *source_next : (HeapWord*) 0);
|
|
563 }
|
|
564
|
|
565 size_t cur_chunk = addr_to_chunk_idx(source_beg);
|
|
566 const size_t end_chunk = addr_to_chunk_idx(chunk_align_up(source_end));
|
|
567
|
|
568 HeapWord *dest_addr = target_beg;
|
|
569 while (cur_chunk < end_chunk) {
|
|
570 size_t words = _chunk_data[cur_chunk].data_size();
|
|
571
|
|
572 #if 1
|
|
573 assert(pointer_delta(target_end, dest_addr) >= words,
|
|
574 "source region does not fit into target region");
|
|
575 #else
|
|
576 // XXX - need some work on the corner cases here. If the chunk does not
|
|
577 // fit, then must either make sure any partial_obj from the chunk fits, or
|
|
578 // 'undo' the initial part of the partial_obj that is in the previous chunk.
|
|
579 if (dest_addr + words >= target_end) {
|
|
580 // Let the caller know where to continue.
|
|
581 *target_next = dest_addr;
|
|
582 *source_next = chunk_to_addr(cur_chunk);
|
|
583 return false;
|
|
584 }
|
|
585 #endif // #if 1
|
|
586
|
|
587 _chunk_data[cur_chunk].set_destination(dest_addr);
|
|
588
|
|
589 // Set the destination_count for cur_chunk, and if necessary, update
|
|
590 // source_chunk for a destination chunk. The source_chunk field is updated
|
|
591 // if cur_chunk is the first (left-most) chunk to be copied to a destination
|
|
592 // chunk.
|
|
593 //
|
|
594 // The destination_count calculation is a bit subtle. A chunk that has data
|
|
595 // that compacts into itself does not count itself as a destination. This
|
|
596 // maintains the invariant that a zero count means the chunk is available
|
|
597 // and can be claimed and then filled.
|
|
598 if (words > 0) {
|
|
599 HeapWord* const last_addr = dest_addr + words - 1;
|
|
600 const size_t dest_chunk_1 = addr_to_chunk_idx(dest_addr);
|
|
601 const size_t dest_chunk_2 = addr_to_chunk_idx(last_addr);
|
|
602 #if 0
|
|
603 // Initially assume that the destination chunks will be the same and
|
|
604 // adjust the value below if necessary. Under this assumption, if
|
|
605 // cur_chunk == dest_chunk_2, then cur_chunk will be compacted completely
|
|
606 // into itself.
|
|
607 uint destination_count = cur_chunk == dest_chunk_2 ? 0 : 1;
|
|
608 if (dest_chunk_1 != dest_chunk_2) {
|
|
609 // Destination chunks differ; adjust destination_count.
|
|
610 destination_count += 1;
|
|
611 // Data from cur_chunk will be copied to the start of dest_chunk_2.
|
|
612 _chunk_data[dest_chunk_2].set_source_chunk(cur_chunk);
|
|
613 } else if (chunk_offset(dest_addr) == 0) {
|
|
614 // Data from cur_chunk will be copied to the start of the destination
|
|
615 // chunk.
|
|
616 _chunk_data[dest_chunk_1].set_source_chunk(cur_chunk);
|
|
617 }
|
|
618 #else
|
|
619 // Initially assume that the destination chunks will be different and
|
|
620 // adjust the value below if necessary. Under this assumption, if
|
|
621 // cur_chunk == dest_chunk2, then cur_chunk will be compacted partially
|
|
622 // into dest_chunk_1 and partially into itself.
|
|
623 uint destination_count = cur_chunk == dest_chunk_2 ? 1 : 2;
|
|
624 if (dest_chunk_1 != dest_chunk_2) {
|
|
625 // Data from cur_chunk will be copied to the start of dest_chunk_2.
|
|
626 _chunk_data[dest_chunk_2].set_source_chunk(cur_chunk);
|
|
627 } else {
|
|
628 // Destination chunks are the same; adjust destination_count.
|
|
629 destination_count -= 1;
|
|
630 if (chunk_offset(dest_addr) == 0) {
|
|
631 // Data from cur_chunk will be copied to the start of the destination
|
|
632 // chunk.
|
|
633 _chunk_data[dest_chunk_1].set_source_chunk(cur_chunk);
|
|
634 }
|
|
635 }
|
|
636 #endif // #if 0
|
|
637
|
|
638 _chunk_data[cur_chunk].set_destination_count(destination_count);
|
|
639 _chunk_data[cur_chunk].set_data_location(chunk_to_addr(cur_chunk));
|
|
640 dest_addr += words;
|
|
641 }
|
|
642
|
|
643 ++cur_chunk;
|
|
644 }
|
|
645
|
|
646 *target_next = dest_addr;
|
|
647 return true;
|
|
648 }
|
|
649
|
|
650 bool ParallelCompactData::partial_obj_ends_in_block(size_t block_index) {
|
|
651 HeapWord* block_addr = block_to_addr(block_index);
|
|
652 HeapWord* block_end_addr = block_addr + BlockSize;
|
|
653 size_t chunk_index = addr_to_chunk_idx(block_addr);
|
|
654 HeapWord* partial_obj_end_addr = partial_obj_end(chunk_index);
|
|
655
|
|
656 // An object that ends at the end of the block, ends
|
|
657 // in the block (the last word of the object is to
|
|
658 // the left of the end).
|
|
659 if ((block_addr < partial_obj_end_addr) &&
|
|
660 (partial_obj_end_addr <= block_end_addr)) {
|
|
661 return true;
|
|
662 }
|
|
663
|
|
664 return false;
|
|
665 }
|
|
666
|
|
667 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
|
|
668 HeapWord* result = NULL;
|
|
669 if (UseParallelOldGCChunkPointerCalc) {
|
|
670 result = chunk_calc_new_pointer(addr);
|
|
671 } else {
|
|
672 result = block_calc_new_pointer(addr);
|
|
673 }
|
|
674 return result;
|
|
675 }
|
|
676
|
|
677 // This method is overly complicated (expensive) to be called
|
|
678 // for every reference.
|
|
679 // Try to restructure this so that a NULL is returned if
|
|
680 // the object is dead. But don't wast the cycles to explicitly check
|
|
681 // that it is dead since only live objects should be passed in.
|
|
682
|
|
683 HeapWord* ParallelCompactData::chunk_calc_new_pointer(HeapWord* addr) {
|
|
684 assert(addr != NULL, "Should detect NULL oop earlier");
|
|
685 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
|
|
686 #ifdef ASSERT
|
|
687 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
|
|
688 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
|
|
689 }
|
|
690 #endif
|
|
691 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
|
|
692
|
|
693 // Chunk covering the object.
|
|
694 size_t chunk_index = addr_to_chunk_idx(addr);
|
|
695 const ChunkData* const chunk_ptr = chunk(chunk_index);
|
|
696 HeapWord* const chunk_addr = chunk_align_down(addr);
|
|
697
|
|
698 assert(addr < chunk_addr + ChunkSize, "Chunk does not cover object");
|
|
699 assert(addr_to_chunk_ptr(chunk_addr) == chunk_ptr, "sanity check");
|
|
700
|
|
701 HeapWord* result = chunk_ptr->destination();
|
|
702
|
|
703 // If all the data in the chunk is live, then the new location of the object
|
|
704 // can be calculated from the destination of the chunk plus the offset of the
|
|
705 // object in the chunk.
|
|
706 if (chunk_ptr->data_size() == ChunkSize) {
|
|
707 result += pointer_delta(addr, chunk_addr);
|
|
708 return result;
|
|
709 }
|
|
710
|
|
711 // The new location of the object is
|
|
712 // chunk destination +
|
|
713 // size of the partial object extending onto the chunk +
|
|
714 // sizes of the live objects in the Chunk that are to the left of addr
|
|
715 const size_t partial_obj_size = chunk_ptr->partial_obj_size();
|
|
716 HeapWord* const search_start = chunk_addr + partial_obj_size;
|
|
717
|
|
718 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
|
|
719 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
|
|
720
|
|
721 result += partial_obj_size + live_to_left;
|
|
722 assert(result <= addr, "object cannot move to the right");
|
|
723 return result;
|
|
724 }
|
|
725
|
|
726 HeapWord* ParallelCompactData::block_calc_new_pointer(HeapWord* addr) {
|
|
727 assert(addr != NULL, "Should detect NULL oop earlier");
|
|
728 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
|
|
729 #ifdef ASSERT
|
|
730 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
|
|
731 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
|
|
732 }
|
|
733 #endif
|
|
734 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
|
|
735
|
|
736 // Chunk covering the object.
|
|
737 size_t chunk_index = addr_to_chunk_idx(addr);
|
|
738 const ChunkData* const chunk_ptr = chunk(chunk_index);
|
|
739 HeapWord* const chunk_addr = chunk_align_down(addr);
|
|
740
|
|
741 assert(addr < chunk_addr + ChunkSize, "Chunk does not cover object");
|
|
742 assert(addr_to_chunk_ptr(chunk_addr) == chunk_ptr, "sanity check");
|
|
743
|
|
744 HeapWord* result = chunk_ptr->destination();
|
|
745
|
|
746 // If all the data in the chunk is live, then the new location of the object
|
|
747 // can be calculated from the destination of the chunk plus the offset of the
|
|
748 // object in the chunk.
|
|
749 if (chunk_ptr->data_size() == ChunkSize) {
|
|
750 result += pointer_delta(addr, chunk_addr);
|
|
751 return result;
|
|
752 }
|
|
753
|
|
754 // The new location of the object is
|
|
755 // chunk destination +
|
|
756 // block offset +
|
|
757 // sizes of the live objects in the Block that are to the left of addr
|
|
758 const size_t block_offset = addr_to_block_ptr(addr)->offset();
|
|
759 HeapWord* const search_start = chunk_addr + block_offset;
|
|
760
|
|
761 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
|
|
762 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
|
|
763
|
|
764 result += block_offset + live_to_left;
|
|
765 assert(result <= addr, "object cannot move to the right");
|
|
766 assert(result == chunk_calc_new_pointer(addr), "Should match");
|
|
767 return result;
|
|
768 }
|
|
769
|
|
770 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
|
|
771 klassOop updated_klass;
|
|
772 if (PSParallelCompact::should_update_klass(old_klass)) {
|
|
773 updated_klass = (klassOop) calc_new_pointer(old_klass);
|
|
774 } else {
|
|
775 updated_klass = old_klass;
|
|
776 }
|
|
777
|
|
778 return updated_klass;
|
|
779 }
|
|
780
|
|
781 #ifdef ASSERT
|
|
782 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
|
|
783 {
|
|
784 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
|
|
785 const size_t* const end = (const size_t*)vspace->committed_high_addr();
|
|
786 for (const size_t* p = beg; p < end; ++p) {
|
|
787 assert(*p == 0, "not zero");
|
|
788 }
|
|
789 }
|
|
790
|
|
791 void ParallelCompactData::verify_clear()
|
|
792 {
|
|
793 verify_clear(_chunk_vspace);
|
|
794 verify_clear(_block_vspace);
|
|
795 }
|
|
796 #endif // #ifdef ASSERT
|
|
797
|
|
798 #ifdef NOT_PRODUCT
|
|
799 ParallelCompactData::ChunkData* debug_chunk(size_t chunk_index) {
|
|
800 ParallelCompactData& sd = PSParallelCompact::summary_data();
|
|
801 return sd.chunk(chunk_index);
|
|
802 }
|
|
803 #endif
|
|
804
|
|
805 elapsedTimer PSParallelCompact::_accumulated_time;
|
|
806 unsigned int PSParallelCompact::_total_invocations = 0;
|
|
807 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
|
|
808 jlong PSParallelCompact::_time_of_last_gc = 0;
|
|
809 CollectorCounters* PSParallelCompact::_counters = NULL;
|
|
810 ParMarkBitMap PSParallelCompact::_mark_bitmap;
|
|
811 ParallelCompactData PSParallelCompact::_summary_data;
|
|
812
|
|
813 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
|
|
814 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
|
|
815 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
|
|
816
|
|
817 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) {
|
|
818 #ifdef VALIDATE_MARK_SWEEP
|
|
819 if (ValidateMarkSweep) {
|
|
820 if (!Universe::heap()->is_in_reserved(p)) {
|
|
821 _root_refs_stack->push(p);
|
|
822 } else {
|
|
823 _other_refs_stack->push(p);
|
|
824 }
|
|
825 }
|
|
826 #endif
|
|
827 mark_and_push(_compaction_manager, p);
|
|
828 }
|
|
829
|
|
830 void PSParallelCompact::mark_and_follow(ParCompactionManager* cm,
|
|
831 oop* p) {
|
|
832 assert(Universe::heap()->is_in_reserved(p),
|
|
833 "we should only be traversing objects here");
|
|
834 oop m = *p;
|
|
835 if (m != NULL && mark_bitmap()->is_unmarked(m)) {
|
|
836 if (mark_obj(m)) {
|
|
837 m->follow_contents(cm); // Follow contents of the marked object
|
|
838 }
|
|
839 }
|
|
840 }
|
|
841
|
|
842 // Anything associated with this variable is temporary.
|
|
843
|
|
844 void PSParallelCompact::mark_and_push_internal(ParCompactionManager* cm,
|
|
845 oop* p) {
|
|
846 // Push marked object, contents will be followed later
|
|
847 oop m = *p;
|
|
848 if (mark_obj(m)) {
|
|
849 // This thread marked the object and
|
|
850 // owns the subsequent processing of it.
|
|
851 cm->save_for_scanning(m);
|
|
852 }
|
|
853 }
|
|
854
|
|
855 void PSParallelCompact::post_initialize() {
|
|
856 ParallelScavengeHeap* heap = gc_heap();
|
|
857 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
|
|
858
|
|
859 MemRegion mr = heap->reserved_region();
|
|
860 _ref_processor = ReferenceProcessor::create_ref_processor(
|
|
861 mr, // span
|
|
862 true, // atomic_discovery
|
|
863 true, // mt_discovery
|
|
864 &_is_alive_closure,
|
|
865 ParallelGCThreads,
|
|
866 ParallelRefProcEnabled);
|
|
867 _counters = new CollectorCounters("PSParallelCompact", 1);
|
|
868
|
|
869 // Initialize static fields in ParCompactionManager.
|
|
870 ParCompactionManager::initialize(mark_bitmap());
|
|
871 }
|
|
872
|
|
873 bool PSParallelCompact::initialize() {
|
|
874 ParallelScavengeHeap* heap = gc_heap();
|
|
875 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
|
|
876 MemRegion mr = heap->reserved_region();
|
|
877
|
|
878 // Was the old gen get allocated successfully?
|
|
879 if (!heap->old_gen()->is_allocated()) {
|
|
880 return false;
|
|
881 }
|
|
882
|
|
883 initialize_space_info();
|
|
884 initialize_dead_wood_limiter();
|
|
885
|
|
886 if (!_mark_bitmap.initialize(mr)) {
|
|
887 vm_shutdown_during_initialization("Unable to allocate bit map for "
|
|
888 "parallel garbage collection for the requested heap size.");
|
|
889 return false;
|
|
890 }
|
|
891
|
|
892 if (!_summary_data.initialize(mr)) {
|
|
893 vm_shutdown_during_initialization("Unable to allocate tables for "
|
|
894 "parallel garbage collection for the requested heap size.");
|
|
895 return false;
|
|
896 }
|
|
897
|
|
898 return true;
|
|
899 }
|
|
900
|
|
901 void PSParallelCompact::initialize_space_info()
|
|
902 {
|
|
903 memset(&_space_info, 0, sizeof(_space_info));
|
|
904
|
|
905 ParallelScavengeHeap* heap = gc_heap();
|
|
906 PSYoungGen* young_gen = heap->young_gen();
|
|
907 MutableSpace* perm_space = heap->perm_gen()->object_space();
|
|
908
|
|
909 _space_info[perm_space_id].set_space(perm_space);
|
|
910 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
|
|
911 _space_info[eden_space_id].set_space(young_gen->eden_space());
|
|
912 _space_info[from_space_id].set_space(young_gen->from_space());
|
|
913 _space_info[to_space_id].set_space(young_gen->to_space());
|
|
914
|
|
915 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
|
|
916 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
|
|
917
|
|
918 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
|
|
919 if (TraceParallelOldGCDensePrefix) {
|
|
920 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
|
|
921 _space_info[perm_space_id].min_dense_prefix());
|
|
922 }
|
|
923 }
|
|
924
|
|
925 void PSParallelCompact::initialize_dead_wood_limiter()
|
|
926 {
|
|
927 const size_t max = 100;
|
|
928 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
|
|
929 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
|
|
930 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
|
|
931 DEBUG_ONLY(_dwl_initialized = true;)
|
|
932 _dwl_adjustment = normal_distribution(1.0);
|
|
933 }
|
|
934
|
|
935 // Simple class for storing info about the heap at the start of GC, to be used
|
|
936 // after GC for comparison/printing.
|
|
937 class PreGCValues {
|
|
938 public:
|
|
939 PreGCValues() { }
|
|
940 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
|
|
941
|
|
942 void fill(ParallelScavengeHeap* heap) {
|
|
943 _heap_used = heap->used();
|
|
944 _young_gen_used = heap->young_gen()->used_in_bytes();
|
|
945 _old_gen_used = heap->old_gen()->used_in_bytes();
|
|
946 _perm_gen_used = heap->perm_gen()->used_in_bytes();
|
|
947 };
|
|
948
|
|
949 size_t heap_used() const { return _heap_used; }
|
|
950 size_t young_gen_used() const { return _young_gen_used; }
|
|
951 size_t old_gen_used() const { return _old_gen_used; }
|
|
952 size_t perm_gen_used() const { return _perm_gen_used; }
|
|
953
|
|
954 private:
|
|
955 size_t _heap_used;
|
|
956 size_t _young_gen_used;
|
|
957 size_t _old_gen_used;
|
|
958 size_t _perm_gen_used;
|
|
959 };
|
|
960
|
|
961 void
|
|
962 PSParallelCompact::clear_data_covering_space(SpaceId id)
|
|
963 {
|
|
964 // At this point, top is the value before GC, new_top() is the value that will
|
|
965 // be set at the end of GC. The marking bitmap is cleared to top; nothing
|
|
966 // should be marked above top. The summary data is cleared to the larger of
|
|
967 // top & new_top.
|
|
968 MutableSpace* const space = _space_info[id].space();
|
|
969 HeapWord* const bot = space->bottom();
|
|
970 HeapWord* const top = space->top();
|
|
971 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
|
|
972
|
|
973 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
|
|
974 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
|
|
975 _mark_bitmap.clear_range(beg_bit, end_bit);
|
|
976
|
|
977 const size_t beg_chunk = _summary_data.addr_to_chunk_idx(bot);
|
|
978 const size_t end_chunk =
|
|
979 _summary_data.addr_to_chunk_idx(_summary_data.chunk_align_up(max_top));
|
|
980 _summary_data.clear_range(beg_chunk, end_chunk);
|
|
981 }
|
|
982
|
|
983 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
|
|
984 {
|
|
985 // Update the from & to space pointers in space_info, since they are swapped
|
|
986 // at each young gen gc. Do the update unconditionally (even though a
|
|
987 // promotion failure does not swap spaces) because an unknown number of minor
|
|
988 // collections will have swapped the spaces an unknown number of times.
|
|
989 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
|
|
990 ParallelScavengeHeap* heap = gc_heap();
|
|
991 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
|
|
992 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
|
|
993
|
|
994 pre_gc_values->fill(heap);
|
|
995
|
|
996 ParCompactionManager::reset();
|
|
997 NOT_PRODUCT(_mark_bitmap.reset_counters());
|
|
998 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
|
|
999 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
|
|
1000
|
|
1001 // Increment the invocation count
|
|
1002 heap->increment_total_collections();
|
|
1003
|
|
1004 // We need to track unique mark sweep invocations as well.
|
|
1005 _total_invocations++;
|
|
1006
|
|
1007 if (PrintHeapAtGC) {
|
|
1008 Universe::print_heap_before_gc();
|
|
1009 }
|
|
1010
|
|
1011 // Fill in TLABs
|
|
1012 heap->accumulate_statistics_all_tlabs();
|
|
1013 heap->ensure_parsability(true); // retire TLABs
|
|
1014
|
|
1015 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
|
|
1016 HandleMark hm; // Discard invalid handles created during verification
|
|
1017 gclog_or_tty->print(" VerifyBeforeGC:");
|
|
1018 Universe::verify(true);
|
|
1019 }
|
|
1020
|
|
1021 // Verify object start arrays
|
|
1022 if (VerifyObjectStartArray &&
|
|
1023 VerifyBeforeGC) {
|
|
1024 heap->old_gen()->verify_object_start_array();
|
|
1025 heap->perm_gen()->verify_object_start_array();
|
|
1026 }
|
|
1027
|
|
1028 DEBUG_ONLY(mark_bitmap()->verify_clear();)
|
|
1029 DEBUG_ONLY(summary_data().verify_clear();)
|
|
1030 }
|
|
1031
|
|
1032 void PSParallelCompact::post_compact()
|
|
1033 {
|
|
1034 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
|
|
1035
|
|
1036 // Clear the marking bitmap and summary data and update top() in each space.
|
|
1037 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
|
|
1038 clear_data_covering_space(SpaceId(id));
|
|
1039 _space_info[id].space()->set_top(_space_info[id].new_top());
|
|
1040 }
|
|
1041
|
|
1042 MutableSpace* const eden_space = _space_info[eden_space_id].space();
|
|
1043 MutableSpace* const from_space = _space_info[from_space_id].space();
|
|
1044 MutableSpace* const to_space = _space_info[to_space_id].space();
|
|
1045
|
|
1046 ParallelScavengeHeap* heap = gc_heap();
|
|
1047 bool eden_empty = eden_space->is_empty();
|
|
1048 if (!eden_empty) {
|
|
1049 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
|
|
1050 heap->young_gen(), heap->old_gen());
|
|
1051 }
|
|
1052
|
|
1053 // Update heap occupancy information which is used as input to the soft ref
|
|
1054 // clearing policy at the next gc.
|
|
1055 Universe::update_heap_info_at_gc();
|
|
1056
|
|
1057 bool young_gen_empty = eden_empty && from_space->is_empty() &&
|
|
1058 to_space->is_empty();
|
|
1059
|
|
1060 BarrierSet* bs = heap->barrier_set();
|
|
1061 if (bs->is_a(BarrierSet::ModRef)) {
|
|
1062 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
|
|
1063 MemRegion old_mr = heap->old_gen()->reserved();
|
|
1064 MemRegion perm_mr = heap->perm_gen()->reserved();
|
|
1065 assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
|
|
1066
|
|
1067 if (young_gen_empty) {
|
|
1068 modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
|
|
1069 } else {
|
|
1070 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
|
|
1071 }
|
|
1072 }
|
|
1073
|
|
1074 Threads::gc_epilogue();
|
|
1075 CodeCache::gc_epilogue();
|
|
1076
|
|
1077 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
|
|
1078
|
|
1079 ref_processor()->enqueue_discovered_references(NULL);
|
|
1080
|
|
1081 // Update time of last GC
|
|
1082 reset_millis_since_last_gc();
|
|
1083 }
|
|
1084
|
|
1085 HeapWord*
|
|
1086 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
|
|
1087 bool maximum_compaction)
|
|
1088 {
|
|
1089 const size_t chunk_size = ParallelCompactData::ChunkSize;
|
|
1090 const ParallelCompactData& sd = summary_data();
|
|
1091
|
|
1092 const MutableSpace* const space = _space_info[id].space();
|
|
1093 HeapWord* const top_aligned_up = sd.chunk_align_up(space->top());
|
|
1094 const ChunkData* const beg_cp = sd.addr_to_chunk_ptr(space->bottom());
|
|
1095 const ChunkData* const end_cp = sd.addr_to_chunk_ptr(top_aligned_up);
|
|
1096
|
|
1097 // Skip full chunks at the beginning of the space--they are necessarily part
|
|
1098 // of the dense prefix.
|
|
1099 size_t full_count = 0;
|
|
1100 const ChunkData* cp;
|
|
1101 for (cp = beg_cp; cp < end_cp && cp->data_size() == chunk_size; ++cp) {
|
|
1102 ++full_count;
|
|
1103 }
|
|
1104
|
|
1105 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
|
|
1106 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
|
|
1107 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
|
|
1108 if (maximum_compaction || cp == end_cp || interval_ended) {
|
|
1109 _maximum_compaction_gc_num = total_invocations();
|
|
1110 return sd.chunk_to_addr(cp);
|
|
1111 }
|
|
1112
|
|
1113 HeapWord* const new_top = _space_info[id].new_top();
|
|
1114 const size_t space_live = pointer_delta(new_top, space->bottom());
|
|
1115 const size_t space_used = space->used_in_words();
|
|
1116 const size_t space_capacity = space->capacity_in_words();
|
|
1117
|
|
1118 const double cur_density = double(space_live) / space_capacity;
|
|
1119 const double deadwood_density =
|
|
1120 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
|
|
1121 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
|
|
1122
|
|
1123 if (TraceParallelOldGCDensePrefix) {
|
|
1124 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
|
|
1125 cur_density, deadwood_density, deadwood_goal);
|
|
1126 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
|
|
1127 "space_cap=" SIZE_FORMAT,
|
|
1128 space_live, space_used,
|
|
1129 space_capacity);
|
|
1130 }
|
|
1131
|
|
1132 // XXX - Use binary search?
|
|
1133 HeapWord* dense_prefix = sd.chunk_to_addr(cp);
|
|
1134 const ChunkData* full_cp = cp;
|
|
1135 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(space->top() - 1);
|
|
1136 while (cp < end_cp) {
|
|
1137 HeapWord* chunk_destination = cp->destination();
|
|
1138 const size_t cur_deadwood = pointer_delta(dense_prefix, chunk_destination);
|
|
1139 if (TraceParallelOldGCDensePrefix && Verbose) {
|
|
1140 tty->print_cr("c#=" SIZE_FORMAT_W("04") " dst=" PTR_FORMAT " "
|
|
1141 "dp=" SIZE_FORMAT_W("08") " " "cdw=" SIZE_FORMAT_W("08"),
|
|
1142 sd.chunk(cp), chunk_destination,
|
|
1143 dense_prefix, cur_deadwood);
|
|
1144 }
|
|
1145
|
|
1146 if (cur_deadwood >= deadwood_goal) {
|
|
1147 // Found the chunk that has the correct amount of deadwood to the left.
|
|
1148 // This typically occurs after crossing a fairly sparse set of chunks, so
|
|
1149 // iterate backwards over those sparse chunks, looking for the chunk that
|
|
1150 // has the lowest density of live objects 'to the right.'
|
|
1151 size_t space_to_left = sd.chunk(cp) * chunk_size;
|
|
1152 size_t live_to_left = space_to_left - cur_deadwood;
|
|
1153 size_t space_to_right = space_capacity - space_to_left;
|
|
1154 size_t live_to_right = space_live - live_to_left;
|
|
1155 double density_to_right = double(live_to_right) / space_to_right;
|
|
1156 while (cp > full_cp) {
|
|
1157 --cp;
|
|
1158 const size_t prev_chunk_live_to_right = live_to_right - cp->data_size();
|
|
1159 const size_t prev_chunk_space_to_right = space_to_right + chunk_size;
|
|
1160 double prev_chunk_density_to_right =
|
|
1161 double(prev_chunk_live_to_right) / prev_chunk_space_to_right;
|
|
1162 if (density_to_right <= prev_chunk_density_to_right) {
|
|
1163 return dense_prefix;
|
|
1164 }
|
|
1165 if (TraceParallelOldGCDensePrefix && Verbose) {
|
|
1166 tty->print_cr("backing up from c=" SIZE_FORMAT_W("4") " d2r=%10.8f "
|
|
1167 "pc_d2r=%10.8f", sd.chunk(cp), density_to_right,
|
|
1168 prev_chunk_density_to_right);
|
|
1169 }
|
|
1170 dense_prefix -= chunk_size;
|
|
1171 live_to_right = prev_chunk_live_to_right;
|
|
1172 space_to_right = prev_chunk_space_to_right;
|
|
1173 density_to_right = prev_chunk_density_to_right;
|
|
1174 }
|
|
1175 return dense_prefix;
|
|
1176 }
|
|
1177
|
|
1178 dense_prefix += chunk_size;
|
|
1179 ++cp;
|
|
1180 }
|
|
1181
|
|
1182 return dense_prefix;
|
|
1183 }
|
|
1184
|
|
1185 #ifndef PRODUCT
|
|
1186 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
|
|
1187 const SpaceId id,
|
|
1188 const bool maximum_compaction,
|
|
1189 HeapWord* const addr)
|
|
1190 {
|
|
1191 const size_t chunk_idx = summary_data().addr_to_chunk_idx(addr);
|
|
1192 ChunkData* const cp = summary_data().chunk(chunk_idx);
|
|
1193 const MutableSpace* const space = _space_info[id].space();
|
|
1194 HeapWord* const new_top = _space_info[id].new_top();
|
|
1195
|
|
1196 const size_t space_live = pointer_delta(new_top, space->bottom());
|
|
1197 const size_t dead_to_left = pointer_delta(addr, cp->destination());
|
|
1198 const size_t space_cap = space->capacity_in_words();
|
|
1199 const double dead_to_left_pct = double(dead_to_left) / space_cap;
|
|
1200 const size_t live_to_right = new_top - cp->destination();
|
|
1201 const size_t dead_to_right = space->top() - addr - live_to_right;
|
|
1202
|
|
1203 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W("05") " "
|
|
1204 "spl=" SIZE_FORMAT " "
|
|
1205 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
|
|
1206 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
|
|
1207 " ratio=%10.8f",
|
|
1208 algorithm, addr, chunk_idx,
|
|
1209 space_live,
|
|
1210 dead_to_left, dead_to_left_pct,
|
|
1211 dead_to_right, live_to_right,
|
|
1212 double(dead_to_right) / live_to_right);
|
|
1213 }
|
|
1214 #endif // #ifndef PRODUCT
|
|
1215
|
|
1216 // Return a fraction indicating how much of the generation can be treated as
|
|
1217 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
|
|
1218 // based on the density of live objects in the generation to determine a limit,
|
|
1219 // which is then adjusted so the return value is min_percent when the density is
|
|
1220 // 1.
|
|
1221 //
|
|
1222 // The following table shows some return values for a different values of the
|
|
1223 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
|
|
1224 // min_percent is 1.
|
|
1225 //
|
|
1226 // fraction allowed as dead wood
|
|
1227 // -----------------------------------------------------------------
|
|
1228 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
|
|
1229 // ------- ---------- ---------- ---------- ---------- ---------- ----------
|
|
1230 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
|
|
1231 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
|
|
1232 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
|
|
1233 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
|
|
1234 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
|
|
1235 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
|
|
1236 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
|
|
1237 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
|
|
1238 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
|
|
1239 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
|
|
1240 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
|
|
1241 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
|
|
1242 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
|
|
1243 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
|
|
1244 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
|
|
1245 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
|
|
1246 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
|
|
1247 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
|
|
1248 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
|
|
1249 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
|
|
1250 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
|
|
1251
|
|
1252 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
|
|
1253 {
|
|
1254 assert(_dwl_initialized, "uninitialized");
|
|
1255
|
|
1256 // The raw limit is the value of the normal distribution at x = density.
|
|
1257 const double raw_limit = normal_distribution(density);
|
|
1258
|
|
1259 // Adjust the raw limit so it becomes the minimum when the density is 1.
|
|
1260 //
|
|
1261 // First subtract the adjustment value (which is simply the precomputed value
|
|
1262 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
|
|
1263 // Then add the minimum value, so the minimum is returned when the density is
|
|
1264 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
|
|
1265 const double min = double(min_percent) / 100.0;
|
|
1266 const double limit = raw_limit - _dwl_adjustment + min;
|
|
1267 return MAX2(limit, 0.0);
|
|
1268 }
|
|
1269
|
|
1270 ParallelCompactData::ChunkData*
|
|
1271 PSParallelCompact::first_dead_space_chunk(const ChunkData* beg,
|
|
1272 const ChunkData* end)
|
|
1273 {
|
|
1274 const size_t chunk_size = ParallelCompactData::ChunkSize;
|
|
1275 ParallelCompactData& sd = summary_data();
|
|
1276 size_t left = sd.chunk(beg);
|
|
1277 size_t right = end > beg ? sd.chunk(end) - 1 : left;
|
|
1278
|
|
1279 // Binary search.
|
|
1280 while (left < right) {
|
|
1281 // Equivalent to (left + right) / 2, but does not overflow.
|
|
1282 const size_t middle = left + (right - left) / 2;
|
|
1283 ChunkData* const middle_ptr = sd.chunk(middle);
|
|
1284 HeapWord* const dest = middle_ptr->destination();
|
|
1285 HeapWord* const addr = sd.chunk_to_addr(middle);
|
|
1286 assert(dest != NULL, "sanity");
|
|
1287 assert(dest <= addr, "must move left");
|
|
1288
|
|
1289 if (middle > left && dest < addr) {
|
|
1290 right = middle - 1;
|
|
1291 } else if (middle < right && middle_ptr->data_size() == chunk_size) {
|
|
1292 left = middle + 1;
|
|
1293 } else {
|
|
1294 return middle_ptr;
|
|
1295 }
|
|
1296 }
|
|
1297 return sd.chunk(left);
|
|
1298 }
|
|
1299
|
|
1300 ParallelCompactData::ChunkData*
|
|
1301 PSParallelCompact::dead_wood_limit_chunk(const ChunkData* beg,
|
|
1302 const ChunkData* end,
|
|
1303 size_t dead_words)
|
|
1304 {
|
|
1305 ParallelCompactData& sd = summary_data();
|
|
1306 size_t left = sd.chunk(beg);
|
|
1307 size_t right = end > beg ? sd.chunk(end) - 1 : left;
|
|
1308
|
|
1309 // Binary search.
|
|
1310 while (left < right) {
|
|
1311 // Equivalent to (left + right) / 2, but does not overflow.
|
|
1312 const size_t middle = left + (right - left) / 2;
|
|
1313 ChunkData* const middle_ptr = sd.chunk(middle);
|
|
1314 HeapWord* const dest = middle_ptr->destination();
|
|
1315 HeapWord* const addr = sd.chunk_to_addr(middle);
|
|
1316 assert(dest != NULL, "sanity");
|
|
1317 assert(dest <= addr, "must move left");
|
|
1318
|
|
1319 const size_t dead_to_left = pointer_delta(addr, dest);
|
|
1320 if (middle > left && dead_to_left > dead_words) {
|
|
1321 right = middle - 1;
|
|
1322 } else if (middle < right && dead_to_left < dead_words) {
|
|
1323 left = middle + 1;
|
|
1324 } else {
|
|
1325 return middle_ptr;
|
|
1326 }
|
|
1327 }
|
|
1328 return sd.chunk(left);
|
|
1329 }
|
|
1330
|
|
1331 // The result is valid during the summary phase, after the initial summarization
|
|
1332 // of each space into itself, and before final summarization.
|
|
1333 inline double
|
|
1334 PSParallelCompact::reclaimed_ratio(const ChunkData* const cp,
|
|
1335 HeapWord* const bottom,
|
|
1336 HeapWord* const top,
|
|
1337 HeapWord* const new_top)
|
|
1338 {
|
|
1339 ParallelCompactData& sd = summary_data();
|
|
1340
|
|
1341 assert(cp != NULL, "sanity");
|
|
1342 assert(bottom != NULL, "sanity");
|
|
1343 assert(top != NULL, "sanity");
|
|
1344 assert(new_top != NULL, "sanity");
|
|
1345 assert(top >= new_top, "summary data problem?");
|
|
1346 assert(new_top > bottom, "space is empty; should not be here");
|
|
1347 assert(new_top >= cp->destination(), "sanity");
|
|
1348 assert(top >= sd.chunk_to_addr(cp), "sanity");
|
|
1349
|
|
1350 HeapWord* const destination = cp->destination();
|
|
1351 const size_t dense_prefix_live = pointer_delta(destination, bottom);
|
|
1352 const size_t compacted_region_live = pointer_delta(new_top, destination);
|
|
1353 const size_t compacted_region_used = pointer_delta(top, sd.chunk_to_addr(cp));
|
|
1354 const size_t reclaimable = compacted_region_used - compacted_region_live;
|
|
1355
|
|
1356 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
|
|
1357 return double(reclaimable) / divisor;
|
|
1358 }
|
|
1359
|
|
1360 // Return the address of the end of the dense prefix, a.k.a. the start of the
|
|
1361 // compacted region. The address is always on a chunk boundary.
|
|
1362 //
|
|
1363 // Completely full chunks at the left are skipped, since no compaction can occur
|
|
1364 // in those chunks. Then the maximum amount of dead wood to allow is computed,
|
|
1365 // based on the density (amount live / capacity) of the generation; the chunk
|
|
1366 // with approximately that amount of dead space to the left is identified as the
|
|
1367 // limit chunk. Chunks between the last completely full chunk and the limit
|
|
1368 // chunk are scanned and the one that has the best (maximum) reclaimed_ratio()
|
|
1369 // is selected.
|
|
1370 HeapWord*
|
|
1371 PSParallelCompact::compute_dense_prefix(const SpaceId id,
|
|
1372 bool maximum_compaction)
|
|
1373 {
|
|
1374 const size_t chunk_size = ParallelCompactData::ChunkSize;
|
|
1375 const ParallelCompactData& sd = summary_data();
|
|
1376
|
|
1377 const MutableSpace* const space = _space_info[id].space();
|
|
1378 HeapWord* const top = space->top();
|
|
1379 HeapWord* const top_aligned_up = sd.chunk_align_up(top);
|
|
1380 HeapWord* const new_top = _space_info[id].new_top();
|
|
1381 HeapWord* const new_top_aligned_up = sd.chunk_align_up(new_top);
|
|
1382 HeapWord* const bottom = space->bottom();
|
|
1383 const ChunkData* const beg_cp = sd.addr_to_chunk_ptr(bottom);
|
|
1384 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(top_aligned_up);
|
|
1385 const ChunkData* const new_top_cp = sd.addr_to_chunk_ptr(new_top_aligned_up);
|
|
1386
|
|
1387 // Skip full chunks at the beginning of the space--they are necessarily part
|
|
1388 // of the dense prefix.
|
|
1389 const ChunkData* const full_cp = first_dead_space_chunk(beg_cp, new_top_cp);
|
|
1390 assert(full_cp->destination() == sd.chunk_to_addr(full_cp) ||
|
|
1391 space->is_empty(), "no dead space allowed to the left");
|
|
1392 assert(full_cp->data_size() < chunk_size || full_cp == new_top_cp - 1,
|
|
1393 "chunk must have dead space");
|
|
1394
|
|
1395 // The gc number is saved whenever a maximum compaction is done, and used to
|
|
1396 // determine when the maximum compaction interval has expired. This avoids
|
|
1397 // successive max compactions for different reasons.
|
|
1398 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
|
|
1399 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
|
|
1400 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
|
|
1401 total_invocations() == HeapFirstMaximumCompactionCount;
|
|
1402 if (maximum_compaction || full_cp == top_cp || interval_ended) {
|
|
1403 _maximum_compaction_gc_num = total_invocations();
|
|
1404 return sd.chunk_to_addr(full_cp);
|
|
1405 }
|
|
1406
|
|
1407 const size_t space_live = pointer_delta(new_top, bottom);
|
|
1408 const size_t space_used = space->used_in_words();
|
|
1409 const size_t space_capacity = space->capacity_in_words();
|
|
1410
|
|
1411 const double density = double(space_live) / double(space_capacity);
|
|
1412 const size_t min_percent_free =
|
|
1413 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
|
|
1414 const double limiter = dead_wood_limiter(density, min_percent_free);
|
|
1415 const size_t dead_wood_max = space_used - space_live;
|
|
1416 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
|
|
1417 dead_wood_max);
|
|
1418
|
|
1419 if (TraceParallelOldGCDensePrefix) {
|
|
1420 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
|
|
1421 "space_cap=" SIZE_FORMAT,
|
|
1422 space_live, space_used,
|
|
1423 space_capacity);
|
|
1424 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
|
|
1425 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
|
|
1426 density, min_percent_free, limiter,
|
|
1427 dead_wood_max, dead_wood_limit);
|
|
1428 }
|
|
1429
|
|
1430 // Locate the chunk with the desired amount of dead space to the left.
|
|
1431 const ChunkData* const limit_cp =
|
|
1432 dead_wood_limit_chunk(full_cp, top_cp, dead_wood_limit);
|
|
1433
|
|
1434 // Scan from the first chunk with dead space to the limit chunk and find the
|
|
1435 // one with the best (largest) reclaimed ratio.
|
|
1436 double best_ratio = 0.0;
|
|
1437 const ChunkData* best_cp = full_cp;
|
|
1438 for (const ChunkData* cp = full_cp; cp < limit_cp; ++cp) {
|
|
1439 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
|
|
1440 if (tmp_ratio > best_ratio) {
|
|
1441 best_cp = cp;
|
|
1442 best_ratio = tmp_ratio;
|
|
1443 }
|
|
1444 }
|
|
1445
|
|
1446 #if 0
|
|
1447 // Something to consider: if the chunk with the best ratio is 'close to' the
|
|
1448 // first chunk w/free space, choose the first chunk with free space
|
|
1449 // ("first-free"). The first-free chunk is usually near the start of the
|
|
1450 // heap, which means we are copying most of the heap already, so copy a bit
|
|
1451 // more to get complete compaction.
|
|
1452 if (pointer_delta(best_cp, full_cp, sizeof(ChunkData)) < 4) {
|
|
1453 _maximum_compaction_gc_num = total_invocations();
|
|
1454 best_cp = full_cp;
|
|
1455 }
|
|
1456 #endif // #if 0
|
|
1457
|
|
1458 return sd.chunk_to_addr(best_cp);
|
|
1459 }
|
|
1460
|
|
1461 void PSParallelCompact::summarize_spaces_quick()
|
|
1462 {
|
|
1463 for (unsigned int i = 0; i < last_space_id; ++i) {
|
|
1464 const MutableSpace* space = _space_info[i].space();
|
|
1465 bool result = _summary_data.summarize(space->bottom(), space->end(),
|
|
1466 space->bottom(), space->top(),
|
|
1467 _space_info[i].new_top_addr());
|
|
1468 assert(result, "should never fail");
|
|
1469 _space_info[i].set_dense_prefix(space->bottom());
|
|
1470 }
|
|
1471 }
|
|
1472
|
|
1473 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
|
|
1474 {
|
|
1475 HeapWord* const dense_prefix_end = dense_prefix(id);
|
|
1476 const ChunkData* chunk = _summary_data.addr_to_chunk_ptr(dense_prefix_end);
|
|
1477 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
|
|
1478 if (dead_space_crosses_boundary(chunk, dense_prefix_bit)) {
|
|
1479 // Only enough dead space is filled so that any remaining dead space to the
|
|
1480 // left is larger than the minimum filler object. (The remainder is filled
|
|
1481 // during the copy/update phase.)
|
|
1482 //
|
|
1483 // The size of the dead space to the right of the boundary is not a
|
|
1484 // concern, since compaction will be able to use whatever space is
|
|
1485 // available.
|
|
1486 //
|
|
1487 // Here '||' is the boundary, 'x' represents a don't care bit and a box
|
|
1488 // surrounds the space to be filled with an object.
|
|
1489 //
|
|
1490 // In the 32-bit VM, each bit represents two 32-bit words:
|
|
1491 // +---+
|
|
1492 // a) beg_bits: ... x x x | 0 | || 0 x x ...
|
|
1493 // end_bits: ... x x x | 0 | || 0 x x ...
|
|
1494 // +---+
|
|
1495 //
|
|
1496 // In the 64-bit VM, each bit represents one 64-bit word:
|
|
1497 // +------------+
|
|
1498 // b) beg_bits: ... x x x | 0 || 0 | x x ...
|
|
1499 // end_bits: ... x x 1 | 0 || 0 | x x ...
|
|
1500 // +------------+
|
|
1501 // +-------+
|
|
1502 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
|
|
1503 // end_bits: ... x 1 | 0 0 | || 0 x x ...
|
|
1504 // +-------+
|
|
1505 // +-----------+
|
|
1506 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
|
|
1507 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
|
|
1508 // +-----------+
|
|
1509 // +-------+
|
|
1510 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
|
|
1511 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
|
|
1512 // +-------+
|
|
1513
|
|
1514 // Initially assume case a, c or e will apply.
|
|
1515 size_t obj_len = (size_t)oopDesc::header_size();
|
|
1516 HeapWord* obj_beg = dense_prefix_end - obj_len;
|
|
1517
|
|
1518 #ifdef _LP64
|
|
1519 if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
|
|
1520 // Case b above.
|
|
1521 obj_beg = dense_prefix_end - 1;
|
|
1522 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
|
|
1523 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
|
|
1524 // Case d above.
|
|
1525 obj_beg = dense_prefix_end - 3;
|
|
1526 obj_len = 3;
|
|
1527 }
|
|
1528 #endif // #ifdef _LP64
|
|
1529
|
|
1530 MemRegion region(obj_beg, obj_len);
|
|
1531 SharedHeap::fill_region_with_object(region);
|
|
1532 _mark_bitmap.mark_obj(obj_beg, obj_len);
|
|
1533 _summary_data.add_obj(obj_beg, obj_len);
|
|
1534 assert(start_array(id) != NULL, "sanity");
|
|
1535 start_array(id)->allocate_block(obj_beg);
|
|
1536 }
|
|
1537 }
|
|
1538
|
|
1539 void
|
|
1540 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
|
|
1541 {
|
|
1542 assert(id < last_space_id, "id out of range");
|
|
1543
|
|
1544 const MutableSpace* space = _space_info[id].space();
|
|
1545 HeapWord** new_top_addr = _space_info[id].new_top_addr();
|
|
1546
|
|
1547 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
|
|
1548 _space_info[id].set_dense_prefix(dense_prefix_end);
|
|
1549
|
|
1550 #ifndef PRODUCT
|
|
1551 if (TraceParallelOldGCDensePrefix) {
|
|
1552 print_dense_prefix_stats("ratio", id, maximum_compaction, dense_prefix_end);
|
|
1553 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
|
|
1554 print_dense_prefix_stats("density", id, maximum_compaction, addr);
|
|
1555 }
|
|
1556 #endif // #ifndef PRODUCT
|
|
1557
|
|
1558 // If dead space crosses the dense prefix boundary, it is (at least partially)
|
|
1559 // filled with a dummy object, marked live and added to the summary data.
|
|
1560 // This simplifies the copy/update phase and must be done before the final
|
|
1561 // locations of objects are determined, to prevent leaving a fragment of dead
|
|
1562 // space that is too small to fill with an object.
|
|
1563 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
|
|
1564 fill_dense_prefix_end(id);
|
|
1565 }
|
|
1566
|
|
1567 // Compute the destination of each Chunk, and thus each object.
|
|
1568 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
|
|
1569 _summary_data.summarize(dense_prefix_end, space->end(),
|
|
1570 dense_prefix_end, space->top(),
|
|
1571 new_top_addr);
|
|
1572
|
|
1573 if (TraceParallelOldGCSummaryPhase) {
|
|
1574 const size_t chunk_size = ParallelCompactData::ChunkSize;
|
|
1575 const size_t dp_chunk = _summary_data.addr_to_chunk_idx(dense_prefix_end);
|
|
1576 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
|
|
1577 const HeapWord* nt_aligned_up = _summary_data.chunk_align_up(*new_top_addr);
|
|
1578 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
|
|
1579 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
|
|
1580 "dp_chunk=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
|
|
1581 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
|
|
1582 id, space->capacity_in_words(), dense_prefix_end,
|
|
1583 dp_chunk, dp_words / chunk_size,
|
|
1584 cr_words / chunk_size, *new_top_addr);
|
|
1585 }
|
|
1586 }
|
|
1587
|
|
1588 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
|
|
1589 bool maximum_compaction)
|
|
1590 {
|
|
1591 EventMark m("2 summarize");
|
|
1592 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
|
|
1593 // trace("2");
|
|
1594
|
|
1595 #ifdef ASSERT
|
|
1596 if (VerifyParallelOldWithMarkSweep &&
|
|
1597 (PSParallelCompact::total_invocations() %
|
|
1598 VerifyParallelOldWithMarkSweepInterval) == 0) {
|
|
1599 verify_mark_bitmap(_mark_bitmap);
|
|
1600 }
|
|
1601 if (TraceParallelOldGCMarkingPhase) {
|
|
1602 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
|
|
1603 "add_obj_bytes=" SIZE_FORMAT,
|
|
1604 add_obj_count, add_obj_size * HeapWordSize);
|
|
1605 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
|
|
1606 "mark_bitmap_bytes=" SIZE_FORMAT,
|
|
1607 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
|
|
1608 }
|
|
1609 #endif // #ifdef ASSERT
|
|
1610
|
|
1611 // Quick summarization of each space into itself, to see how much is live.
|
|
1612 summarize_spaces_quick();
|
|
1613
|
|
1614 if (TraceParallelOldGCSummaryPhase) {
|
|
1615 tty->print_cr("summary_phase: after summarizing each space to self");
|
|
1616 Universe::print();
|
|
1617 NOT_PRODUCT(print_chunk_ranges());
|
|
1618 if (Verbose) {
|
|
1619 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
|
|
1620 }
|
|
1621 }
|
|
1622
|
|
1623 // The amount of live data that will end up in old space (assuming it fits).
|
|
1624 size_t old_space_total_live = 0;
|
|
1625 unsigned int id;
|
|
1626 for (id = old_space_id; id < last_space_id; ++id) {
|
|
1627 old_space_total_live += pointer_delta(_space_info[id].new_top(),
|
|
1628 _space_info[id].space()->bottom());
|
|
1629 }
|
|
1630
|
|
1631 const MutableSpace* old_space = _space_info[old_space_id].space();
|
|
1632 if (old_space_total_live > old_space->capacity_in_words()) {
|
|
1633 // XXX - should also try to expand
|
|
1634 maximum_compaction = true;
|
|
1635 } else if (!UseParallelOldGCDensePrefix) {
|
|
1636 maximum_compaction = true;
|
|
1637 }
|
|
1638
|
|
1639 // Permanent and Old generations.
|
|
1640 summarize_space(perm_space_id, maximum_compaction);
|
|
1641 summarize_space(old_space_id, maximum_compaction);
|
|
1642
|
|
1643 // Summarize the remaining spaces (those in the young gen) into old space. If
|
|
1644 // the live data from a space doesn't fit, the existing summarization is left
|
|
1645 // intact, so the data is compacted down within the space itself.
|
|
1646 HeapWord** new_top_addr = _space_info[old_space_id].new_top_addr();
|
|
1647 HeapWord* const target_space_end = old_space->end();
|
|
1648 for (id = eden_space_id; id < last_space_id; ++id) {
|
|
1649 const MutableSpace* space = _space_info[id].space();
|
|
1650 const size_t live = pointer_delta(_space_info[id].new_top(),
|
|
1651 space->bottom());
|
|
1652 const size_t available = pointer_delta(target_space_end, *new_top_addr);
|
|
1653 if (live <= available) {
|
|
1654 // All the live data will fit.
|
|
1655 if (TraceParallelOldGCSummaryPhase) {
|
|
1656 tty->print_cr("summarizing %d into old_space @ " PTR_FORMAT,
|
|
1657 id, *new_top_addr);
|
|
1658 }
|
|
1659 _summary_data.summarize(*new_top_addr, target_space_end,
|
|
1660 space->bottom(), space->top(),
|
|
1661 new_top_addr);
|
|
1662
|
|
1663 // Reset the new_top value for the space.
|
|
1664 _space_info[id].set_new_top(space->bottom());
|
|
1665
|
|
1666 // Clear the source_chunk field for each chunk in the space.
|
|
1667 ChunkData* beg_chunk = _summary_data.addr_to_chunk_ptr(space->bottom());
|
|
1668 ChunkData* end_chunk = _summary_data.addr_to_chunk_ptr(space->top() - 1);
|
|
1669 while (beg_chunk <= end_chunk) {
|
|
1670 beg_chunk->set_source_chunk(0);
|
|
1671 ++beg_chunk;
|
|
1672 }
|
|
1673 }
|
|
1674 }
|
|
1675
|
|
1676 // Fill in the block data after any changes to the chunks have
|
|
1677 // been made.
|
|
1678 #ifdef ASSERT
|
|
1679 summarize_blocks(cm, perm_space_id);
|
|
1680 summarize_blocks(cm, old_space_id);
|
|
1681 #else
|
|
1682 if (!UseParallelOldGCChunkPointerCalc) {
|
|
1683 summarize_blocks(cm, perm_space_id);
|
|
1684 summarize_blocks(cm, old_space_id);
|
|
1685 }
|
|
1686 #endif
|
|
1687
|
|
1688 if (TraceParallelOldGCSummaryPhase) {
|
|
1689 tty->print_cr("summary_phase: after final summarization");
|
|
1690 Universe::print();
|
|
1691 NOT_PRODUCT(print_chunk_ranges());
|
|
1692 if (Verbose) {
|
|
1693 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
|
|
1694 }
|
|
1695 }
|
|
1696 }
|
|
1697
|
|
1698 // Fill in the BlockData.
|
|
1699 // Iterate over the spaces and within each space iterate over
|
|
1700 // the chunks and fill in the BlockData for each chunk.
|
|
1701
|
|
1702 void PSParallelCompact::summarize_blocks(ParCompactionManager* cm,
|
|
1703 SpaceId first_compaction_space_id) {
|
|
1704 #if 0
|
|
1705 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(1);)
|
|
1706 for (SpaceId cur_space_id = first_compaction_space_id;
|
|
1707 cur_space_id != last_space_id;
|
|
1708 cur_space_id = next_compaction_space_id(cur_space_id)) {
|
|
1709 // Iterate over the chunks in the space
|
|
1710 size_t start_chunk_index =
|
|
1711 _summary_data.addr_to_chunk_idx(space(cur_space_id)->bottom());
|
|
1712 BitBlockUpdateClosure bbu(mark_bitmap(),
|
|
1713 cm,
|
|
1714 start_chunk_index);
|
|
1715 // Iterate over blocks.
|
|
1716 for (size_t chunk_index = start_chunk_index;
|
|
1717 chunk_index < _summary_data.chunk_count() &&
|
|
1718 _summary_data.chunk_to_addr(chunk_index) < space(cur_space_id)->top();
|
|
1719 chunk_index++) {
|
|
1720
|
|
1721 // Reset the closure for the new chunk. Note that the closure
|
|
1722 // maintains some data that does not get reset for each chunk
|
|
1723 // so a new instance of the closure is no appropriate.
|
|
1724 bbu.reset_chunk(chunk_index);
|
|
1725
|
|
1726 // Start the iteration with the first live object. This
|
|
1727 // may return the end of the chunk. That is acceptable since
|
|
1728 // it will properly limit the iterations.
|
|
1729 ParMarkBitMap::idx_t left_offset = mark_bitmap()->addr_to_bit(
|
|
1730 _summary_data.first_live_or_end_in_chunk(chunk_index));
|
|
1731
|
|
1732 // End the iteration at the end of the chunk.
|
|
1733 HeapWord* chunk_addr = _summary_data.chunk_to_addr(chunk_index);
|
|
1734 HeapWord* chunk_end = chunk_addr + ParallelCompactData::ChunkSize;
|
|
1735 ParMarkBitMap::idx_t right_offset =
|
|
1736 mark_bitmap()->addr_to_bit(chunk_end);
|
|
1737
|
|
1738 // Blocks that have not objects starting in them can be
|
|
1739 // skipped because their data will never be used.
|
|
1740 if (left_offset < right_offset) {
|
|
1741
|
|
1742 // Iterate through the objects in the chunk.
|
|
1743 ParMarkBitMap::idx_t last_offset =
|
|
1744 mark_bitmap()->pair_iterate(&bbu, left_offset, right_offset);
|
|
1745
|
|
1746 // If last_offset is less than right_offset, then the iterations
|
|
1747 // terminated while it was looking for an end bit. "last_offset"
|
|
1748 // is then the offset for the last start bit. In this situation
|
|
1749 // the "offset" field for the next block to the right (_cur_block + 1)
|
|
1750 // will not have been update although there may be live data
|
|
1751 // to the left of the chunk.
|
|
1752
|
|
1753 size_t cur_block_plus_1 = bbu.cur_block() + 1;
|
|
1754 HeapWord* cur_block_plus_1_addr =
|
|
1755 _summary_data.block_to_addr(bbu.cur_block()) +
|
|
1756 ParallelCompactData::BlockSize;
|
|
1757 HeapWord* last_offset_addr = mark_bitmap()->bit_to_addr(last_offset);
|
|
1758 #if 1 // This code works. The else doesn't but should. Why does it?
|
|
1759 // The current block (cur_block()) has already been updated.
|
|
1760 // The last block that may need to be updated is either the
|
|
1761 // next block (current block + 1) or the block where the
|
|
1762 // last object starts (which can be greater than the
|
|
1763 // next block if there were no objects found in intervening
|
|
1764 // blocks).
|
|
1765 size_t last_block =
|
|
1766 MAX2(bbu.cur_block() + 1,
|
|
1767 _summary_data.addr_to_block_idx(last_offset_addr));
|
|
1768 #else
|
|
1769 // The current block has already been updated. The only block
|
|
1770 // that remains to be updated is the block where the last
|
|
1771 // object in the chunk starts.
|
|
1772 size_t last_block = _summary_data.addr_to_block_idx(last_offset_addr);
|
|
1773 #endif
|
|
1774 assert_bit_is_start(last_offset);
|
|
1775 assert((last_block == _summary_data.block_count()) ||
|
|
1776 (_summary_data.block(last_block)->raw_offset() == 0),
|
|
1777 "Should not have been set");
|
|
1778 // Is the last block still in the current chunk? If still
|
|
1779 // in this chunk, update the last block (the counting that
|
|
1780 // included the current block is meant for the offset of the last
|
|
1781 // block). If not in this chunk, do nothing. Should not
|
|
1782 // update a block in the next chunk.
|
|
1783 if (ParallelCompactData::chunk_contains_block(bbu.chunk_index(),
|
|
1784 last_block)) {
|
|
1785 if (last_offset < right_offset) {
|
|
1786 // The last object started in this chunk but ends beyond
|
|
1787 // this chunk. Update the block for this last object.
|
|
1788 assert(mark_bitmap()->is_marked(last_offset), "Should be marked");
|
|
1789 // No end bit was found. The closure takes care of
|
|
1790 // the cases where
|
|
1791 // an objects crosses over into the next block
|
|
1792 // an objects starts and ends in the next block
|
|
1793 // It does not handle the case where an object is
|
|
1794 // the first object in a later block and extends
|
|
1795 // past the end of the chunk (i.e., the closure
|
|
1796 // only handles complete objects that are in the range
|
|
1797 // it is given). That object is handed back here
|
|
1798 // for any special consideration necessary.
|
|
1799 //
|
|
1800 // Is the first bit in the last block a start or end bit?
|
|
1801 //
|
|
1802 // If the partial object ends in the last block L,
|
|
1803 // then the 1st bit in L may be an end bit.
|
|
1804 //
|
|
1805 // Else does the last object start in a block after the current
|
|
1806 // block? A block AA will already have been updated if an
|
|
1807 // object ends in the next block AA+1. An object found to end in
|
|
1808 // the AA+1 is the trigger that updates AA. Objects are being
|
|
1809 // counted in the current block for updaing a following
|
|
1810 // block. An object may start in later block
|
|
1811 // block but may extend beyond the last block in the chunk.
|
|
1812 // Updates are only done when the end of an object has been
|
|
1813 // found. If the last object (covered by block L) starts
|
|
1814 // beyond the current block, then no object ends in L (otherwise
|
|
1815 // L would be the current block). So the first bit in L is
|
|
1816 // a start bit.
|
|
1817 //
|
|
1818 // Else the last objects start in the current block and ends
|
|
1819 // beyond the chunk. The current block has already been
|
|
1820 // updated and there is no later block (with an object
|
|
1821 // starting in it) that needs to be updated.
|
|
1822 //
|
|
1823 if (_summary_data.partial_obj_ends_in_block(last_block)) {
|
|
1824 _summary_data.block(last_block)->set_end_bit_offset(
|
|
1825 bbu.live_data_left());
|
|
1826 } else if (last_offset_addr >= cur_block_plus_1_addr) {
|
|
1827 // The start of the object is on a later block
|
|
1828 // (to the right of the current block and there are no
|
|
1829 // complete live objects to the left of this last object
|
|
1830 // within the chunk.
|
|
1831 // The first bit in the block is for the start of the
|
|
1832 // last object.
|
|
1833 _summary_data.block(last_block)->set_start_bit_offset(
|
|
1834 bbu.live_data_left());
|
|
1835 } else {
|
|
1836 // The start of the last object was found in
|
|
1837 // the current chunk (which has already
|
|
1838 // been updated).
|
|
1839 assert(bbu.cur_block() ==
|
|
1840 _summary_data.addr_to_block_idx(last_offset_addr),
|
|
1841 "Should be a block already processed");
|
|
1842 }
|
|
1843 #ifdef ASSERT
|
|
1844 // Is there enough block information to find this object?
|
|
1845 // The destination of the chunk has not been set so the
|
|
1846 // values returned by calc_new_pointer() and
|
|
1847 // block_calc_new_pointer() will only be
|
|
1848 // offsets. But they should agree.
|
|
1849 HeapWord* moved_obj_with_chunks =
|
|
1850 _summary_data.chunk_calc_new_pointer(last_offset_addr);
|
|
1851 HeapWord* moved_obj_with_blocks =
|
|
1852 _summary_data.calc_new_pointer(last_offset_addr);
|
|
1853 assert(moved_obj_with_chunks == moved_obj_with_blocks,
|
|
1854 "Block calculation is wrong");
|
|
1855 #endif
|
|
1856 } else if (last_block < _summary_data.block_count()) {
|
|
1857 // Iterations ended looking for a start bit (but
|
|
1858 // did not run off the end of the block table).
|
|
1859 _summary_data.block(last_block)->set_start_bit_offset(
|
|
1860 bbu.live_data_left());
|
|
1861 }
|
|
1862 }
|
|
1863 #ifdef ASSERT
|
|
1864 // Is there enough block information to find this object?
|
|
1865 HeapWord* left_offset_addr = mark_bitmap()->bit_to_addr(left_offset);
|
|
1866 HeapWord* moved_obj_with_chunks =
|
|
1867 _summary_data.calc_new_pointer(left_offset_addr);
|
|
1868 HeapWord* moved_obj_with_blocks =
|
|
1869 _summary_data.calc_new_pointer(left_offset_addr);
|
|
1870 assert(moved_obj_with_chunks == moved_obj_with_blocks,
|
|
1871 "Block calculation is wrong");
|
|
1872 #endif
|
|
1873
|
|
1874 // Is there another block after the end of this chunk?
|
|
1875 #ifdef ASSERT
|
|
1876 if (last_block < _summary_data.block_count()) {
|
|
1877 // No object may have been found in a block. If that
|
|
1878 // block is at the end of the chunk, the iteration will
|
|
1879 // terminate without incrementing the current block so
|
|
1880 // that the current block is not the last block in the
|
|
1881 // chunk. That situation precludes asserting that the
|
|
1882 // current block is the last block in the chunk. Assert
|
|
1883 // the lesser condition that the current block does not
|
|
1884 // exceed the chunk.
|
|
1885 assert(_summary_data.block_to_addr(last_block) <=
|
|
1886 (_summary_data.chunk_to_addr(chunk_index) +
|
|
1887 ParallelCompactData::ChunkSize),
|
|
1888 "Chunk and block inconsistency");
|
|
1889 assert(last_offset <= right_offset, "Iteration over ran end");
|
|
1890 }
|
|
1891 #endif
|
|
1892 }
|
|
1893 #ifdef ASSERT
|
|
1894 if (PrintGCDetails && Verbose) {
|
|
1895 if (_summary_data.chunk(chunk_index)->partial_obj_size() == 1) {
|
|
1896 size_t first_block =
|
|
1897 chunk_index / ParallelCompactData::BlocksPerChunk;
|
|
1898 gclog_or_tty->print_cr("first_block " PTR_FORMAT
|
|
1899 " _offset " PTR_FORMAT
|
|
1900 "_first_is_start_bit %d",
|
|
1901 first_block,
|
|
1902 _summary_data.block(first_block)->raw_offset(),
|
|
1903 _summary_data.block(first_block)->first_is_start_bit());
|
|
1904 }
|
|
1905 }
|
|
1906 #endif
|
|
1907 }
|
|
1908 }
|
|
1909 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(16);)
|
|
1910 #endif // #if 0
|
|
1911 }
|
|
1912
|
|
1913 // This method should contain all heap-specific policy for invoking a full
|
|
1914 // collection. invoke_no_policy() will only attempt to compact the heap; it
|
|
1915 // will do nothing further. If we need to bail out for policy reasons, scavenge
|
|
1916 // before full gc, or any other specialized behavior, it needs to be added here.
|
|
1917 //
|
|
1918 // Note that this method should only be called from the vm_thread while at a
|
|
1919 // safepoint.
|
|
1920 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
|
|
1921 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
|
|
1922 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
|
|
1923 "should be in vm thread");
|
|
1924 ParallelScavengeHeap* heap = gc_heap();
|
|
1925 GCCause::Cause gc_cause = heap->gc_cause();
|
|
1926 assert(!heap->is_gc_active(), "not reentrant");
|
|
1927
|
|
1928 PSAdaptiveSizePolicy* policy = heap->size_policy();
|
|
1929
|
|
1930 // Before each allocation/collection attempt, find out from the
|
|
1931 // policy object if GCs are, on the whole, taking too long. If so,
|
|
1932 // bail out without attempting a collection. The exceptions are
|
|
1933 // for explicitly requested GC's.
|
|
1934 if (!policy->gc_time_limit_exceeded() ||
|
|
1935 GCCause::is_user_requested_gc(gc_cause) ||
|
|
1936 GCCause::is_serviceability_requested_gc(gc_cause)) {
|
|
1937 IsGCActiveMark mark;
|
|
1938
|
|
1939 if (ScavengeBeforeFullGC) {
|
|
1940 PSScavenge::invoke_no_policy();
|
|
1941 }
|
|
1942
|
|
1943 PSParallelCompact::invoke_no_policy(maximum_heap_compaction);
|
|
1944 }
|
|
1945 }
|
|
1946
|
|
1947 bool ParallelCompactData::chunk_contains(size_t chunk_index, HeapWord* addr) {
|
|
1948 size_t addr_chunk_index = addr_to_chunk_idx(addr);
|
|
1949 return chunk_index == addr_chunk_index;
|
|
1950 }
|
|
1951
|
|
1952 bool ParallelCompactData::chunk_contains_block(size_t chunk_index,
|
|
1953 size_t block_index) {
|
|
1954 size_t first_block_in_chunk = chunk_index * BlocksPerChunk;
|
|
1955 size_t last_block_in_chunk = (chunk_index + 1) * BlocksPerChunk - 1;
|
|
1956
|
|
1957 return (first_block_in_chunk <= block_index) &&
|
|
1958 (block_index <= last_block_in_chunk);
|
|
1959 }
|
|
1960
|
|
1961 // This method contains no policy. You should probably
|
|
1962 // be calling invoke() instead.
|
|
1963 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
|
|
1964 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
|
|
1965 assert(ref_processor() != NULL, "Sanity");
|
|
1966
|
|
1967 if (GC_locker::is_active()) {
|
|
1968 return;
|
|
1969 }
|
|
1970
|
|
1971 TimeStamp marking_start;
|
|
1972 TimeStamp compaction_start;
|
|
1973 TimeStamp collection_exit;
|
|
1974
|
|
1975 // "serial_CM" is needed until the parallel implementation
|
|
1976 // of the move and update is done.
|
|
1977 ParCompactionManager* serial_CM = new ParCompactionManager();
|
|
1978 // Don't initialize more than once.
|
|
1979 // serial_CM->initialize(&summary_data(), mark_bitmap());
|
|
1980
|
|
1981 ParallelScavengeHeap* heap = gc_heap();
|
|
1982 GCCause::Cause gc_cause = heap->gc_cause();
|
|
1983 PSYoungGen* young_gen = heap->young_gen();
|
|
1984 PSOldGen* old_gen = heap->old_gen();
|
|
1985 PSPermGen* perm_gen = heap->perm_gen();
|
|
1986 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
|
|
1987
|
|
1988 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
|
|
1989
|
|
1990 // Make sure data structures are sane, make the heap parsable, and do other
|
|
1991 // miscellaneous bookkeeping.
|
|
1992 PreGCValues pre_gc_values;
|
|
1993 pre_compact(&pre_gc_values);
|
|
1994
|
|
1995 // Place after pre_compact() where the number of invocations is incremented.
|
|
1996 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
|
|
1997
|
|
1998 {
|
|
1999 ResourceMark rm;
|
|
2000 HandleMark hm;
|
|
2001
|
|
2002 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc;
|
|
2003
|
|
2004 // This is useful for debugging but don't change the output the
|
|
2005 // the customer sees.
|
|
2006 const char* gc_cause_str = "Full GC";
|
|
2007 if (is_system_gc && PrintGCDetails) {
|
|
2008 gc_cause_str = "Full GC (System)";
|
|
2009 }
|
|
2010 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
|
|
2011 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
|
|
2012 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty);
|
|
2013 TraceCollectorStats tcs(counters());
|
|
2014 TraceMemoryManagerStats tms(true /* Full GC */);
|
|
2015
|
|
2016 if (TraceGen1Time) accumulated_time()->start();
|
|
2017
|
|
2018 // Let the size policy know we're starting
|
|
2019 size_policy->major_collection_begin();
|
|
2020
|
|
2021 // When collecting the permanent generation methodOops may be moving,
|
|
2022 // so we either have to flush all bcp data or convert it into bci.
|
|
2023 CodeCache::gc_prologue();
|
|
2024 Threads::gc_prologue();
|
|
2025
|
|
2026 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
|
|
2027 COMPILER2_PRESENT(DerivedPointerTable::clear());
|
|
2028
|
|
2029 ref_processor()->enable_discovery();
|
|
2030
|
|
2031 bool marked_for_unloading = false;
|
|
2032
|
|
2033 marking_start.update();
|
|
2034 marking_phase(serial_CM, maximum_heap_compaction);
|
|
2035
|
|
2036 #ifndef PRODUCT
|
|
2037 if (TraceParallelOldGCMarkingPhase) {
|
|
2038 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d "
|
|
2039 "cas_by_another %d",
|
|
2040 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
|
|
2041 mark_bitmap()->cas_by_another());
|
|
2042 }
|
|
2043 #endif // #ifndef PRODUCT
|
|
2044
|
|
2045 #ifdef ASSERT
|
|
2046 if (VerifyParallelOldWithMarkSweep &&
|
|
2047 (PSParallelCompact::total_invocations() %
|
|
2048 VerifyParallelOldWithMarkSweepInterval) == 0) {
|
|
2049 gclog_or_tty->print_cr("Verify marking with mark_sweep_phase1()");
|
|
2050 if (PrintGCDetails && Verbose) {
|
|
2051 gclog_or_tty->print_cr("mark_sweep_phase1:");
|
|
2052 }
|
|
2053 // Clear the discovered lists so that discovered objects
|
|
2054 // don't look like they have been discovered twice.
|
|
2055 ref_processor()->clear_discovered_references();
|
|
2056
|
|
2057 PSMarkSweep::allocate_stacks();
|
|
2058 MemRegion mr = Universe::heap()->reserved_region();
|
|
2059 PSMarkSweep::ref_processor()->enable_discovery();
|
|
2060 PSMarkSweep::mark_sweep_phase1(maximum_heap_compaction);
|
|
2061 }
|
|
2062 #endif
|
|
2063
|
|
2064 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc;
|
|
2065 summary_phase(serial_CM, maximum_heap_compaction || max_on_system_gc);
|
|
2066
|
|
2067 #ifdef ASSERT
|
|
2068 if (VerifyParallelOldWithMarkSweep &&
|
|
2069 (PSParallelCompact::total_invocations() %
|
|
2070 VerifyParallelOldWithMarkSweepInterval) == 0) {
|
|
2071 if (PrintGCDetails && Verbose) {
|
|
2072 gclog_or_tty->print_cr("mark_sweep_phase2:");
|
|
2073 }
|
|
2074 PSMarkSweep::mark_sweep_phase2();
|
|
2075 }
|
|
2076 #endif
|
|
2077
|
|
2078 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
|
|
2079 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
|
|
2080
|
|
2081 // adjust_roots() updates Universe::_intArrayKlassObj which is
|
|
2082 // needed by the compaction for filling holes in the dense prefix.
|
|
2083 adjust_roots();
|
|
2084
|
|
2085 #ifdef ASSERT
|
|
2086 if (VerifyParallelOldWithMarkSweep &&
|
|
2087 (PSParallelCompact::total_invocations() %
|
|
2088 VerifyParallelOldWithMarkSweepInterval) == 0) {
|
|
2089 // Do a separate verify phase so that the verify
|
|
2090 // code can use the the forwarding pointers to
|
|
2091 // check the new pointer calculation. The restore_marks()
|
|
2092 // has to be done before the real compact.
|
|
2093 serial_CM->set_action(ParCompactionManager::VerifyUpdate);
|
|
2094 compact_perm(serial_CM);
|
|
2095 compact_serial(serial_CM);
|
|
2096 serial_CM->set_action(ParCompactionManager::ResetObjects);
|
|
2097 compact_perm(serial_CM);
|
|
2098 compact_serial(serial_CM);
|
|
2099 serial_CM->set_action(ParCompactionManager::UpdateAndCopy);
|
|
2100
|
|
2101 // For debugging only
|
|
2102 PSMarkSweep::restore_marks();
|
|
2103 PSMarkSweep::deallocate_stacks();
|
|
2104 }
|
|
2105 #endif
|
|
2106
|
|
2107 compaction_start.update();
|
|
2108 // Does the perm gen always have to be done serially because
|
|
2109 // klasses are used in the update of an object?
|
|
2110 compact_perm(serial_CM);
|
|
2111
|
|
2112 if (UseParallelOldGCCompacting) {
|
|
2113 compact();
|
|
2114 } else {
|
|
2115 compact_serial(serial_CM);
|
|
2116 }
|
|
2117
|
|
2118 delete serial_CM;
|
|
2119
|
|
2120 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
|
|
2121 // done before resizing.
|
|
2122 post_compact();
|
|
2123
|
|
2124 // Let the size policy know we're done
|
|
2125 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
|
|
2126
|
|
2127 if (UseAdaptiveSizePolicy) {
|
|
2128 if (PrintAdaptiveSizePolicy) {
|
|
2129 gclog_or_tty->print("AdaptiveSizeStart: ");
|
|
2130 gclog_or_tty->stamp();
|
|
2131 gclog_or_tty->print_cr(" collection: %d ",
|
|
2132 heap->total_collections());
|
|
2133 if (Verbose) {
|
|
2134 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
|
|
2135 " perm_gen_capacity: %d ",
|
|
2136 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
|
|
2137 perm_gen->capacity_in_bytes());
|
|
2138 }
|
|
2139 }
|
|
2140
|
|
2141 // Don't check if the size_policy is ready here. Let
|
|
2142 // the size_policy check that internally.
|
|
2143 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
|
|
2144 ((gc_cause != GCCause::_java_lang_system_gc) ||
|
|
2145 UseAdaptiveSizePolicyWithSystemGC)) {
|
|
2146 // Calculate optimal free space amounts
|
|
2147 assert(young_gen->max_size() >
|
|
2148 young_gen->from_space()->capacity_in_bytes() +
|
|
2149 young_gen->to_space()->capacity_in_bytes(),
|
|
2150 "Sizes of space in young gen are out-of-bounds");
|
|
2151 size_t max_eden_size = young_gen->max_size() -
|
|
2152 young_gen->from_space()->capacity_in_bytes() -
|
|
2153 young_gen->to_space()->capacity_in_bytes();
|
|
2154 size_policy->compute_generation_free_space(young_gen->used_in_bytes(),
|
|
2155 young_gen->eden_space()->used_in_bytes(),
|
|
2156 old_gen->used_in_bytes(),
|
|
2157 perm_gen->used_in_bytes(),
|
|
2158 young_gen->eden_space()->capacity_in_bytes(),
|
|
2159 old_gen->max_gen_size(),
|
|
2160 max_eden_size,
|
|
2161 true /* full gc*/,
|
|
2162 gc_cause);
|
|
2163
|
|
2164 heap->resize_old_gen(size_policy->calculated_old_free_size_in_bytes());
|
|
2165
|
|
2166 // Don't resize the young generation at an major collection. A
|
|
2167 // desired young generation size may have been calculated but
|
|
2168 // resizing the young generation complicates the code because the
|
|
2169 // resizing of the old generation may have moved the boundary
|
|
2170 // between the young generation and the old generation. Let the
|
|
2171 // young generation resizing happen at the minor collections.
|
|
2172 }
|
|
2173 if (PrintAdaptiveSizePolicy) {
|
|
2174 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
|
|
2175 heap->total_collections());
|
|
2176 }
|
|
2177 }
|
|
2178
|
|
2179 if (UsePerfData) {
|
|
2180 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
|
|
2181 counters->update_counters();
|
|
2182 counters->update_old_capacity(old_gen->capacity_in_bytes());
|
|
2183 counters->update_young_capacity(young_gen->capacity_in_bytes());
|
|
2184 }
|
|
2185
|
|
2186 heap->resize_all_tlabs();
|
|
2187
|
|
2188 // We collected the perm gen, so we'll resize it here.
|
|
2189 perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
|
|
2190
|
|
2191 if (TraceGen1Time) accumulated_time()->stop();
|
|
2192
|
|
2193 if (PrintGC) {
|
|
2194 if (PrintGCDetails) {
|
|
2195 // No GC timestamp here. This is after GC so it would be confusing.
|
|
2196 young_gen->print_used_change(pre_gc_values.young_gen_used());
|
|
2197 old_gen->print_used_change(pre_gc_values.old_gen_used());
|
|
2198 heap->print_heap_change(pre_gc_values.heap_used());
|
|
2199 // Print perm gen last (print_heap_change() excludes the perm gen).
|
|
2200 perm_gen->print_used_change(pre_gc_values.perm_gen_used());
|
|
2201 } else {
|
|
2202 heap->print_heap_change(pre_gc_values.heap_used());
|
|
2203 }
|
|
2204 }
|
|
2205
|
|
2206 // Track memory usage and detect low memory
|
|
2207 MemoryService::track_memory_usage();
|
|
2208 heap->update_counters();
|
|
2209
|
|
2210 if (PrintGCDetails) {
|
|
2211 if (size_policy->print_gc_time_limit_would_be_exceeded()) {
|
|
2212 if (size_policy->gc_time_limit_exceeded()) {
|
|
2213 gclog_or_tty->print_cr(" GC time is exceeding GCTimeLimit "
|
|
2214 "of %d%%", GCTimeLimit);
|
|
2215 } else {
|
|
2216 gclog_or_tty->print_cr(" GC time would exceed GCTimeLimit "
|
|
2217 "of %d%%", GCTimeLimit);
|
|
2218 }
|
|
2219 }
|
|
2220 size_policy->set_print_gc_time_limit_would_be_exceeded(false);
|
|
2221 }
|
|
2222 }
|
|
2223
|
|
2224 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
|
|
2225 HandleMark hm; // Discard invalid handles created during verification
|
|
2226 gclog_or_tty->print(" VerifyAfterGC:");
|
|
2227 Universe::verify(false);
|
|
2228 }
|
|
2229
|
|
2230 // Re-verify object start arrays
|
|
2231 if (VerifyObjectStartArray &&
|
|
2232 VerifyAfterGC) {
|
|
2233 old_gen->verify_object_start_array();
|
|
2234 perm_gen->verify_object_start_array();
|
|
2235 }
|
|
2236
|
|
2237 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
|
|
2238
|
|
2239 collection_exit.update();
|
|
2240
|
|
2241 if (PrintHeapAtGC) {
|
|
2242 Universe::print_heap_after_gc();
|
|
2243 }
|
|
2244 if (PrintGCTaskTimeStamps) {
|
|
2245 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
|
|
2246 INT64_FORMAT,
|
|
2247 marking_start.ticks(), compaction_start.ticks(),
|
|
2248 collection_exit.ticks());
|
|
2249 gc_task_manager()->print_task_time_stamps();
|
|
2250 }
|
|
2251 }
|
|
2252
|
|
2253 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
|
|
2254 PSYoungGen* young_gen,
|
|
2255 PSOldGen* old_gen) {
|
|
2256 MutableSpace* const eden_space = young_gen->eden_space();
|
|
2257 assert(!eden_space->is_empty(), "eden must be non-empty");
|
|
2258 assert(young_gen->virtual_space()->alignment() ==
|
|
2259 old_gen->virtual_space()->alignment(), "alignments do not match");
|
|
2260
|
|
2261 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
|
|
2262 return false;
|
|
2263 }
|
|
2264
|
|
2265 // Both generations must be completely committed.
|
|
2266 if (young_gen->virtual_space()->uncommitted_size() != 0) {
|
|
2267 return false;
|
|
2268 }
|
|
2269 if (old_gen->virtual_space()->uncommitted_size() != 0) {
|
|
2270 return false;
|
|
2271 }
|
|
2272
|
|
2273 // Figure out how much to take from eden. Include the average amount promoted
|
|
2274 // in the total; otherwise the next young gen GC will simply bail out to a
|
|
2275 // full GC.
|
|
2276 const size_t alignment = old_gen->virtual_space()->alignment();
|
|
2277 const size_t eden_used = eden_space->used_in_bytes();
|
|
2278 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
|
|
2279 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
|
|
2280 const size_t eden_capacity = eden_space->capacity_in_bytes();
|
|
2281
|
|
2282 if (absorb_size >= eden_capacity) {
|
|
2283 return false; // Must leave some space in eden.
|
|
2284 }
|
|
2285
|
|
2286 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
|
|
2287 if (new_young_size < young_gen->min_gen_size()) {
|
|
2288 return false; // Respect young gen minimum size.
|
|
2289 }
|
|
2290
|
|
2291 if (TraceAdaptiveGCBoundary && Verbose) {
|
|
2292 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
|
|
2293 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
|
|
2294 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
|
|
2295 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
|
|
2296 absorb_size / K,
|
|
2297 eden_capacity / K, (eden_capacity - absorb_size) / K,
|
|
2298 young_gen->from_space()->used_in_bytes() / K,
|
|
2299 young_gen->to_space()->used_in_bytes() / K,
|
|
2300 young_gen->capacity_in_bytes() / K, new_young_size / K);
|
|
2301 }
|
|
2302
|
|
2303 // Fill the unused part of the old gen.
|
|
2304 MutableSpace* const old_space = old_gen->object_space();
|
|
2305 MemRegion old_gen_unused(old_space->top(), old_space->end());
|
|
2306 if (!old_gen_unused.is_empty()) {
|
|
2307 SharedHeap::fill_region_with_object(old_gen_unused);
|
|
2308 }
|
|
2309
|
|
2310 // Take the live data from eden and set both top and end in the old gen to
|
|
2311 // eden top. (Need to set end because reset_after_change() mangles the region
|
|
2312 // from end to virtual_space->high() in debug builds).
|
|
2313 HeapWord* const new_top = eden_space->top();
|
|
2314 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
|
|
2315 absorb_size);
|
|
2316 young_gen->reset_after_change();
|
|
2317 old_space->set_top(new_top);
|
|
2318 old_space->set_end(new_top);
|
|
2319 old_gen->reset_after_change();
|
|
2320
|
|
2321 // Update the object start array for the filler object and the data from eden.
|
|
2322 ObjectStartArray* const start_array = old_gen->start_array();
|
|
2323 HeapWord* const start = old_gen_unused.start();
|
|
2324 for (HeapWord* addr = start; addr < new_top; addr += oop(addr)->size()) {
|
|
2325 start_array->allocate_block(addr);
|
|
2326 }
|
|
2327
|
|
2328 // Could update the promoted average here, but it is not typically updated at
|
|
2329 // full GCs and the value to use is unclear. Something like
|
|
2330 //
|
|
2331 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
|
|
2332
|
|
2333 size_policy->set_bytes_absorbed_from_eden(absorb_size);
|
|
2334 return true;
|
|
2335 }
|
|
2336
|
|
2337 GCTaskManager* const PSParallelCompact::gc_task_manager() {
|
|
2338 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
|
|
2339 "shouldn't return NULL");
|
|
2340 return ParallelScavengeHeap::gc_task_manager();
|
|
2341 }
|
|
2342
|
|
2343 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
|
|
2344 bool maximum_heap_compaction) {
|
|
2345 // Recursively traverse all live objects and mark them
|
|
2346 EventMark m("1 mark object");
|
|
2347 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
|
|
2348
|
|
2349 ParallelScavengeHeap* heap = gc_heap();
|
|
2350 uint parallel_gc_threads = heap->gc_task_manager()->workers();
|
|
2351 TaskQueueSetSuper* qset = ParCompactionManager::chunk_array();
|
|
2352 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
|
|
2353
|
|
2354 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
|
|
2355 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
|
|
2356
|
|
2357 {
|
|
2358 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
|
|
2359
|
|
2360 GCTaskQueue* q = GCTaskQueue::create();
|
|
2361
|
|
2362 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
|
|
2363 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
|
|
2364 // We scan the thread roots in parallel
|
|
2365 Threads::create_thread_roots_marking_tasks(q);
|
|
2366 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
|
|
2367 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
|
|
2368 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
|
|
2369 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
|
|
2370 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
|
|
2371 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols));
|
|
2372
|
|
2373 if (parallel_gc_threads > 1) {
|
|
2374 for (uint j = 0; j < parallel_gc_threads; j++) {
|
|
2375 q->enqueue(new StealMarkingTask(&terminator));
|
|
2376 }
|
|
2377 }
|
|
2378
|
|
2379 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
|
|
2380 q->enqueue(fin);
|
|
2381
|
|
2382 gc_task_manager()->add_list(q);
|
|
2383
|
|
2384 fin->wait_for();
|
|
2385
|
|
2386 // We have to release the barrier tasks!
|
|
2387 WaitForBarrierGCTask::destroy(fin);
|
|
2388 }
|
|
2389
|
|
2390 // Process reference objects found during marking
|
|
2391 {
|
|
2392 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
|
|
2393 ReferencePolicy *soft_ref_policy;
|
|
2394 if (maximum_heap_compaction) {
|
|
2395 soft_ref_policy = new AlwaysClearPolicy();
|
|
2396 } else {
|
|
2397 #ifdef COMPILER2
|
|
2398 soft_ref_policy = new LRUMaxHeapPolicy();
|
|
2399 #else
|
|
2400 soft_ref_policy = new LRUCurrentHeapPolicy();
|
|
2401 #endif // COMPILER2
|
|
2402 }
|
|
2403 assert(soft_ref_policy != NULL, "No soft reference policy");
|
|
2404 if (ref_processor()->processing_is_mt()) {
|
|
2405 RefProcTaskExecutor task_executor;
|
|
2406 ref_processor()->process_discovered_references(
|
|
2407 soft_ref_policy, is_alive_closure(), &mark_and_push_closure,
|
|
2408 &follow_stack_closure, &task_executor);
|
|
2409 } else {
|
|
2410 ref_processor()->process_discovered_references(
|
|
2411 soft_ref_policy, is_alive_closure(), &mark_and_push_closure,
|
|
2412 &follow_stack_closure, NULL);
|
|
2413 }
|
|
2414 }
|
|
2415
|
|
2416 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
|
|
2417 // Follow system dictionary roots and unload classes.
|
|
2418 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
|
|
2419
|
|
2420 // Follow code cache roots.
|
|
2421 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
|
|
2422 purged_class);
|
|
2423 follow_stack(cm); // Flush marking stack.
|
|
2424
|
|
2425 // Update subklass/sibling/implementor links of live klasses
|
|
2426 // revisit_klass_stack is used in follow_weak_klass_links().
|
|
2427 follow_weak_klass_links(cm);
|
|
2428
|
|
2429 // Visit symbol and interned string tables and delete unmarked oops
|
|
2430 SymbolTable::unlink(is_alive_closure());
|
|
2431 StringTable::unlink(is_alive_closure());
|
|
2432
|
|
2433 assert(cm->marking_stack()->size() == 0, "stack should be empty by now");
|
|
2434 assert(cm->overflow_stack()->is_empty(), "stack should be empty by now");
|
|
2435 }
|
|
2436
|
|
2437 // This should be moved to the shared markSweep code!
|
|
2438 class PSAlwaysTrueClosure: public BoolObjectClosure {
|
|
2439 public:
|
|
2440 void do_object(oop p) { ShouldNotReachHere(); }
|
|
2441 bool do_object_b(oop p) { return true; }
|
|
2442 };
|
|
2443 static PSAlwaysTrueClosure always_true;
|
|
2444
|
|
2445 void PSParallelCompact::adjust_roots() {
|
|
2446 // Adjust the pointers to reflect the new locations
|
|
2447 EventMark m("3 adjust roots");
|
|
2448 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
|
|
2449
|
|
2450 // General strong roots.
|
|
2451 Universe::oops_do(adjust_root_pointer_closure());
|
|
2452 ReferenceProcessor::oops_do(adjust_root_pointer_closure());
|
|
2453 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles
|
|
2454 Threads::oops_do(adjust_root_pointer_closure());
|
|
2455 ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
|
|
2456 FlatProfiler::oops_do(adjust_root_pointer_closure());
|
|
2457 Management::oops_do(adjust_root_pointer_closure());
|
|
2458 JvmtiExport::oops_do(adjust_root_pointer_closure());
|
|
2459 // SO_AllClasses
|
|
2460 SystemDictionary::oops_do(adjust_root_pointer_closure());
|
|
2461 vmSymbols::oops_do(adjust_root_pointer_closure());
|
|
2462
|
|
2463 // Now adjust pointers in remaining weak roots. (All of which should
|
|
2464 // have been cleared if they pointed to non-surviving objects.)
|
|
2465 // Global (weak) JNI handles
|
|
2466 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
|
|
2467
|
|
2468 CodeCache::oops_do(adjust_pointer_closure());
|
|
2469 SymbolTable::oops_do(adjust_root_pointer_closure());
|
|
2470 StringTable::oops_do(adjust_root_pointer_closure());
|
|
2471 ref_processor()->weak_oops_do(adjust_root_pointer_closure());
|
|
2472 // Roots were visited so references into the young gen in roots
|
|
2473 // may have been scanned. Process them also.
|
|
2474 // Should the reference processor have a span that excludes
|
|
2475 // young gen objects?
|
|
2476 PSScavenge::reference_processor()->weak_oops_do(
|
|
2477 adjust_root_pointer_closure());
|
|
2478 }
|
|
2479
|
|
2480 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
|
|
2481 EventMark m("4 compact perm");
|
|
2482 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty);
|
|
2483 // trace("4");
|
|
2484
|
|
2485 gc_heap()->perm_gen()->start_array()->reset();
|
|
2486 move_and_update(cm, perm_space_id);
|
|
2487 }
|
|
2488
|
|
2489 void PSParallelCompact::enqueue_chunk_draining_tasks(GCTaskQueue* q,
|
|
2490 uint parallel_gc_threads) {
|
|
2491 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
|
|
2492
|
|
2493 const unsigned int task_count = MAX2(parallel_gc_threads, 1U);
|
|
2494 for (unsigned int j = 0; j < task_count; j++) {
|
|
2495 q->enqueue(new DrainStacksCompactionTask());
|
|
2496 }
|
|
2497
|
|
2498 // Find all chunks that are available (can be filled immediately) and
|
|
2499 // distribute them to the thread stacks. The iteration is done in reverse
|
|
2500 // order (high to low) so the chunks will be removed in ascending order.
|
|
2501
|
|
2502 const ParallelCompactData& sd = PSParallelCompact::summary_data();
|
|
2503
|
|
2504 size_t fillable_chunks = 0; // A count for diagnostic purposes.
|
|
2505 unsigned int which = 0; // The worker thread number.
|
|
2506
|
|
2507 for (unsigned int id = to_space_id; id > perm_space_id; --id) {
|
|
2508 SpaceInfo* const space_info = _space_info + id;
|
|
2509 MutableSpace* const space = space_info->space();
|
|
2510 HeapWord* const new_top = space_info->new_top();
|
|
2511
|
|
2512 const size_t beg_chunk = sd.addr_to_chunk_idx(space_info->dense_prefix());
|
|
2513 const size_t end_chunk = sd.addr_to_chunk_idx(sd.chunk_align_up(new_top));
|
|
2514 assert(end_chunk > 0, "perm gen cannot be empty");
|
|
2515
|
|
2516 for (size_t cur = end_chunk - 1; cur >= beg_chunk; --cur) {
|
|
2517 if (sd.chunk(cur)->claim_unsafe()) {
|
|
2518 ParCompactionManager* cm = ParCompactionManager::manager_array(which);
|
|
2519 cm->save_for_processing(cur);
|
|
2520
|
|
2521 if (TraceParallelOldGCCompactionPhase && Verbose) {
|
|
2522 const size_t count_mod_8 = fillable_chunks & 7;
|
|
2523 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
|
|
2524 gclog_or_tty->print(" " SIZE_FORMAT_W("7"), cur);
|
|
2525 if (count_mod_8 == 7) gclog_or_tty->cr();
|
|
2526 }
|
|
2527
|
|
2528 NOT_PRODUCT(++fillable_chunks;)
|
|
2529
|
|
2530 // Assign chunks to threads in round-robin fashion.
|
|
2531 if (++which == task_count) {
|
|
2532 which = 0;
|
|
2533 }
|
|
2534 }
|
|
2535 }
|
|
2536 }
|
|
2537
|
|
2538 if (TraceParallelOldGCCompactionPhase) {
|
|
2539 if (Verbose && (fillable_chunks & 7) != 0) gclog_or_tty->cr();
|
|
2540 gclog_or_tty->print_cr("%u initially fillable chunks", fillable_chunks);
|
|
2541 }
|
|
2542 }
|
|
2543
|
|
2544 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
|
|
2545
|
|
2546 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
|
|
2547 uint parallel_gc_threads) {
|
|
2548 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
|
|
2549
|
|
2550 ParallelCompactData& sd = PSParallelCompact::summary_data();
|
|
2551
|
|
2552 // Iterate over all the spaces adding tasks for updating
|
|
2553 // chunks in the dense prefix. Assume that 1 gc thread
|
|
2554 // will work on opening the gaps and the remaining gc threads
|
|
2555 // will work on the dense prefix.
|
|
2556 SpaceId space_id = old_space_id;
|
|
2557 while (space_id != last_space_id) {
|
|
2558 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
|
|
2559 const MutableSpace* const space = _space_info[space_id].space();
|
|
2560
|
|
2561 if (dense_prefix_end == space->bottom()) {
|
|
2562 // There is no dense prefix for this space.
|
|
2563 space_id = next_compaction_space_id(space_id);
|
|
2564 continue;
|
|
2565 }
|
|
2566
|
|
2567 // The dense prefix is before this chunk.
|
|
2568 size_t chunk_index_end_dense_prefix =
|
|
2569 sd.addr_to_chunk_idx(dense_prefix_end);
|
|
2570 ChunkData* const dense_prefix_cp = sd.chunk(chunk_index_end_dense_prefix);
|
|
2571 assert(dense_prefix_end == space->end() ||
|
|
2572 dense_prefix_cp->available() ||
|
|
2573 dense_prefix_cp->claimed(),
|
|
2574 "The chunk after the dense prefix should always be ready to fill");
|
|
2575
|
|
2576 size_t chunk_index_start = sd.addr_to_chunk_idx(space->bottom());
|
|
2577
|
|
2578 // Is there dense prefix work?
|
|
2579 size_t total_dense_prefix_chunks =
|
|
2580 chunk_index_end_dense_prefix - chunk_index_start;
|
|
2581 // How many chunks of the dense prefix should be given to
|
|
2582 // each thread?
|
|
2583 if (total_dense_prefix_chunks > 0) {
|
|
2584 uint tasks_for_dense_prefix = 1;
|
|
2585 if (UseParallelDensePrefixUpdate) {
|
|
2586 if (total_dense_prefix_chunks <=
|
|
2587 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
|
|
2588 // Don't over partition. This assumes that
|
|
2589 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
|
|
2590 // so there are not many chunks to process.
|
|
2591 tasks_for_dense_prefix = parallel_gc_threads;
|
|
2592 } else {
|
|
2593 // Over partition
|
|
2594 tasks_for_dense_prefix = parallel_gc_threads *
|
|
2595 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
|
|
2596 }
|
|
2597 }
|
|
2598 size_t chunks_per_thread = total_dense_prefix_chunks /
|
|
2599 tasks_for_dense_prefix;
|
|
2600 // Give each thread at least 1 chunk.
|
|
2601 if (chunks_per_thread == 0) {
|
|
2602 chunks_per_thread = 1;
|
|
2603 }
|
|
2604
|
|
2605 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
|
|
2606 if (chunk_index_start >= chunk_index_end_dense_prefix) {
|
|
2607 break;
|
|
2608 }
|
|
2609 // chunk_index_end is not processed
|
|
2610 size_t chunk_index_end = MIN2(chunk_index_start + chunks_per_thread,
|
|
2611 chunk_index_end_dense_prefix);
|
|
2612 q->enqueue(new UpdateDensePrefixTask(
|
|
2613 space_id,
|
|
2614 chunk_index_start,
|
|
2615 chunk_index_end));
|
|
2616 chunk_index_start = chunk_index_end;
|
|
2617 }
|
|
2618 }
|
|
2619 // This gets any part of the dense prefix that did not
|
|
2620 // fit evenly.
|
|
2621 if (chunk_index_start < chunk_index_end_dense_prefix) {
|
|
2622 q->enqueue(new UpdateDensePrefixTask(
|
|
2623 space_id,
|
|
2624 chunk_index_start,
|
|
2625 chunk_index_end_dense_prefix));
|
|
2626 }
|
|
2627 space_id = next_compaction_space_id(space_id);
|
|
2628 } // End tasks for dense prefix
|
|
2629 }
|
|
2630
|
|
2631 void PSParallelCompact::enqueue_chunk_stealing_tasks(
|
|
2632 GCTaskQueue* q,
|
|
2633 ParallelTaskTerminator* terminator_ptr,
|
|
2634 uint parallel_gc_threads) {
|
|
2635 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
|
|
2636
|
|
2637 // Once a thread has drained it's stack, it should try to steal chunks from
|
|
2638 // other threads.
|
|
2639 if (parallel_gc_threads > 1) {
|
|
2640 for (uint j = 0; j < parallel_gc_threads; j++) {
|
|
2641 q->enqueue(new StealChunkCompactionTask(terminator_ptr));
|
|
2642 }
|
|
2643 }
|
|
2644 }
|
|
2645
|
|
2646 void PSParallelCompact::compact() {
|
|
2647 EventMark m("5 compact");
|
|
2648 // trace("5");
|
|
2649 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
|
|
2650
|
|
2651 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
|
|
2652 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
|
|
2653 PSOldGen* old_gen = heap->old_gen();
|
|
2654 old_gen->start_array()->reset();
|
|
2655 uint parallel_gc_threads = heap->gc_task_manager()->workers();
|
|
2656 TaskQueueSetSuper* qset = ParCompactionManager::chunk_array();
|
|
2657 ParallelTaskTerminator terminator(parallel_gc_threads, qset);
|
|
2658
|
|
2659 GCTaskQueue* q = GCTaskQueue::create();
|
|
2660 enqueue_chunk_draining_tasks(q, parallel_gc_threads);
|
|
2661 enqueue_dense_prefix_tasks(q, parallel_gc_threads);
|
|
2662 enqueue_chunk_stealing_tasks(q, &terminator, parallel_gc_threads);
|
|
2663
|
|
2664 {
|
|
2665 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
|
|
2666
|
|
2667 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
|
|
2668 q->enqueue(fin);
|
|
2669
|
|
2670 gc_task_manager()->add_list(q);
|
|
2671
|
|
2672 fin->wait_for();
|
|
2673
|
|
2674 // We have to release the barrier tasks!
|
|
2675 WaitForBarrierGCTask::destroy(fin);
|
|
2676
|
|
2677 #ifdef ASSERT
|
|
2678 // Verify that all chunks have been processed before the deferred updates.
|
|
2679 // Note that perm_space_id is skipped; this type of verification is not
|
|
2680 // valid until the perm gen is compacted by chunks.
|
|
2681 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
|
|
2682 verify_complete(SpaceId(id));
|
|
2683 }
|
|
2684 #endif
|
|
2685 }
|
|
2686
|
|
2687 {
|
|
2688 // Update the deferred objects, if any. Any compaction manager can be used.
|
|
2689 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
|
|
2690 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
|
|
2691 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
|
|
2692 update_deferred_objects(cm, SpaceId(id));
|
|
2693 }
|
|
2694 }
|
|
2695 }
|
|
2696
|
|
2697 #ifdef ASSERT
|
|
2698 void PSParallelCompact::verify_complete(SpaceId space_id) {
|
|
2699 // All Chunks between space bottom() to new_top() should be marked as filled
|
|
2700 // and all Chunks between new_top() and top() should be available (i.e.,
|
|
2701 // should have been emptied).
|
|
2702 ParallelCompactData& sd = summary_data();
|
|
2703 SpaceInfo si = _space_info[space_id];
|
|
2704 HeapWord* new_top_addr = sd.chunk_align_up(si.new_top());
|
|
2705 HeapWord* old_top_addr = sd.chunk_align_up(si.space()->top());
|
|
2706 const size_t beg_chunk = sd.addr_to_chunk_idx(si.space()->bottom());
|
|
2707 const size_t new_top_chunk = sd.addr_to_chunk_idx(new_top_addr);
|
|
2708 const size_t old_top_chunk = sd.addr_to_chunk_idx(old_top_addr);
|
|
2709
|
|
2710 bool issued_a_warning = false;
|
|
2711
|
|
2712 size_t cur_chunk;
|
|
2713 for (cur_chunk = beg_chunk; cur_chunk < new_top_chunk; ++cur_chunk) {
|
|
2714 const ChunkData* const c = sd.chunk(cur_chunk);
|
|
2715 if (!c->completed()) {
|
|
2716 warning("chunk " SIZE_FORMAT " not filled: "
|
|
2717 "destination_count=" SIZE_FORMAT,
|
|
2718 cur_chunk, c->destination_count());
|
|
2719 issued_a_warning = true;
|
|
2720 }
|
|
2721 }
|
|
2722
|
|
2723 for (cur_chunk = new_top_chunk; cur_chunk < old_top_chunk; ++cur_chunk) {
|
|
2724 const ChunkData* const c = sd.chunk(cur_chunk);
|
|
2725 if (!c->available()) {
|
|
2726 warning("chunk " SIZE_FORMAT " not empty: "
|
|
2727 "destination_count=" SIZE_FORMAT,
|
|
2728 cur_chunk, c->destination_count());
|
|
2729 issued_a_warning = true;
|
|
2730 }
|
|
2731 }
|
|
2732
|
|
2733 if (issued_a_warning) {
|
|
2734 print_chunk_ranges();
|
|
2735 }
|
|
2736 }
|
|
2737 #endif // #ifdef ASSERT
|
|
2738
|
|
2739 void PSParallelCompact::compact_serial(ParCompactionManager* cm) {
|
|
2740 EventMark m("5 compact serial");
|
|
2741 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty);
|
|
2742
|
|
2743 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
|
|
2744 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
|
|
2745
|
|
2746 PSYoungGen* young_gen = heap->young_gen();
|
|
2747 PSOldGen* old_gen = heap->old_gen();
|
|
2748
|
|
2749 old_gen->start_array()->reset();
|
|
2750 old_gen->move_and_update(cm);
|
|
2751 young_gen->move_and_update(cm);
|
|
2752 }
|
|
2753
|
|
2754 void PSParallelCompact::follow_root(ParCompactionManager* cm, oop* p) {
|
|
2755 assert(!Universe::heap()->is_in_reserved(p),
|
|
2756 "roots shouldn't be things within the heap");
|
|
2757 #ifdef VALIDATE_MARK_SWEEP
|
|
2758 if (ValidateMarkSweep) {
|
|
2759 guarantee(!_root_refs_stack->contains(p), "should only be in here once");
|
|
2760 _root_refs_stack->push(p);
|
|
2761 }
|
|
2762 #endif
|
|
2763 oop m = *p;
|
|
2764 if (m != NULL && mark_bitmap()->is_unmarked(m)) {
|
|
2765 if (mark_obj(m)) {
|
|
2766 m->follow_contents(cm); // Follow contents of the marked object
|
|
2767 }
|
|
2768 }
|
|
2769 follow_stack(cm);
|
|
2770 }
|
|
2771
|
|
2772 void PSParallelCompact::follow_stack(ParCompactionManager* cm) {
|
|
2773 while(!cm->overflow_stack()->is_empty()) {
|
|
2774 oop obj = cm->overflow_stack()->pop();
|
|
2775 obj->follow_contents(cm);
|
|
2776 }
|
|
2777
|
|
2778 oop obj;
|
|
2779 // obj is a reference!!!
|
|
2780 while (cm->marking_stack()->pop_local(obj)) {
|
|
2781 // It would be nice to assert about the type of objects we might
|
|
2782 // pop, but they can come from anywhere, unfortunately.
|
|
2783 obj->follow_contents(cm);
|
|
2784 }
|
|
2785 }
|
|
2786
|
|
2787 void
|
|
2788 PSParallelCompact::follow_weak_klass_links(ParCompactionManager* serial_cm) {
|
|
2789 // All klasses on the revisit stack are marked at this point.
|
|
2790 // Update and follow all subklass, sibling and implementor links.
|
|
2791 for (uint i = 0; i < ParallelGCThreads+1; i++) {
|
|
2792 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
|
|
2793 KeepAliveClosure keep_alive_closure(cm);
|
|
2794 for (int i = 0; i < cm->revisit_klass_stack()->length(); i++) {
|
|
2795 cm->revisit_klass_stack()->at(i)->follow_weak_klass_links(
|
|
2796 is_alive_closure(),
|
|
2797 &keep_alive_closure);
|
|
2798 }
|
|
2799 follow_stack(cm);
|
|
2800 }
|
|
2801 }
|
|
2802
|
|
2803 void
|
|
2804 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
|
|
2805 cm->revisit_klass_stack()->push(k);
|
|
2806 }
|
|
2807
|
|
2808 #ifdef VALIDATE_MARK_SWEEP
|
|
2809
|
|
2810 void PSParallelCompact::track_adjusted_pointer(oop* p, oop newobj, bool isroot) {
|
|
2811 if (!ValidateMarkSweep)
|
|
2812 return;
|
|
2813
|
|
2814 if (!isroot) {
|
|
2815 if (_pointer_tracking) {
|
|
2816 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
|
|
2817 _adjusted_pointers->remove(p);
|
|
2818 }
|
|
2819 } else {
|
|
2820 ptrdiff_t index = _root_refs_stack->find(p);
|
|
2821 if (index != -1) {
|
|
2822 int l = _root_refs_stack->length();
|
|
2823 if (l > 0 && l - 1 != index) {
|
|
2824 oop* last = _root_refs_stack->pop();
|
|
2825 assert(last != p, "should be different");
|
|
2826 _root_refs_stack->at_put(index, last);
|
|
2827 } else {
|
|
2828 _root_refs_stack->remove(p);
|
|
2829 }
|
|
2830 }
|
|
2831 }
|
|
2832 }
|
|
2833
|
|
2834
|
|
2835 void PSParallelCompact::check_adjust_pointer(oop* p) {
|
|
2836 _adjusted_pointers->push(p);
|
|
2837 }
|
|
2838
|
|
2839
|
|
2840 class AdjusterTracker: public OopClosure {
|
|
2841 public:
|
|
2842 AdjusterTracker() {};
|
|
2843 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); }
|
|
2844 };
|
|
2845
|
|
2846
|
|
2847 void PSParallelCompact::track_interior_pointers(oop obj) {
|
|
2848 if (ValidateMarkSweep) {
|
|
2849 _adjusted_pointers->clear();
|
|
2850 _pointer_tracking = true;
|
|
2851
|
|
2852 AdjusterTracker checker;
|
|
2853 obj->oop_iterate(&checker);
|
|
2854 }
|
|
2855 }
|
|
2856
|
|
2857
|
|
2858 void PSParallelCompact::check_interior_pointers() {
|
|
2859 if (ValidateMarkSweep) {
|
|
2860 _pointer_tracking = false;
|
|
2861 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
|
|
2862 }
|
|
2863 }
|
|
2864
|
|
2865
|
|
2866 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
|
|
2867 if (ValidateMarkSweep) {
|
|
2868 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
|
|
2869 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
|
|
2870 }
|
|
2871 }
|
|
2872
|
|
2873
|
|
2874 void PSParallelCompact::register_live_oop(oop p, size_t size) {
|
|
2875 if (ValidateMarkSweep) {
|
|
2876 _live_oops->push(p);
|
|
2877 _live_oops_size->push(size);
|
|
2878 _live_oops_index++;
|
|
2879 }
|
|
2880 }
|
|
2881
|
|
2882 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
|
|
2883 if (ValidateMarkSweep) {
|
|
2884 oop obj = _live_oops->at((int)_live_oops_index);
|
|
2885 guarantee(obj == p, "should be the same object");
|
|
2886 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
|
|
2887 _live_oops_index++;
|
|
2888 }
|
|
2889 }
|
|
2890
|
|
2891 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
|
|
2892 HeapWord* compaction_top) {
|
|
2893 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
|
|
2894 "should be moved to forwarded location");
|
|
2895 if (ValidateMarkSweep) {
|
|
2896 PSParallelCompact::validate_live_oop(oop(q), size);
|
|
2897 _live_oops_moved_to->push(oop(compaction_top));
|
|
2898 }
|
|
2899 if (RecordMarkSweepCompaction) {
|
|
2900 _cur_gc_live_oops->push(q);
|
|
2901 _cur_gc_live_oops_moved_to->push(compaction_top);
|
|
2902 _cur_gc_live_oops_size->push(size);
|
|
2903 }
|
|
2904 }
|
|
2905
|
|
2906
|
|
2907 void PSParallelCompact::compaction_complete() {
|
|
2908 if (RecordMarkSweepCompaction) {
|
|
2909 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops;
|
|
2910 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
|
|
2911 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size;
|
|
2912
|
|
2913 _cur_gc_live_oops = _last_gc_live_oops;
|
|
2914 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to;
|
|
2915 _cur_gc_live_oops_size = _last_gc_live_oops_size;
|
|
2916 _last_gc_live_oops = _tmp_live_oops;
|
|
2917 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
|
|
2918 _last_gc_live_oops_size = _tmp_live_oops_size;
|
|
2919 }
|
|
2920 }
|
|
2921
|
|
2922
|
|
2923 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
|
|
2924 if (!RecordMarkSweepCompaction) {
|
|
2925 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
|
|
2926 return;
|
|
2927 }
|
|
2928
|
|
2929 if (_last_gc_live_oops == NULL) {
|
|
2930 tty->print_cr("No compaction information gathered yet");
|
|
2931 return;
|
|
2932 }
|
|
2933
|
|
2934 for (int i = 0; i < _last_gc_live_oops->length(); i++) {
|
|
2935 HeapWord* old_oop = _last_gc_live_oops->at(i);
|
|
2936 size_t sz = _last_gc_live_oops_size->at(i);
|
|
2937 if (old_oop <= q && q < (old_oop + sz)) {
|
|
2938 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
|
|
2939 size_t offset = (q - old_oop);
|
|
2940 tty->print_cr("Address " PTR_FORMAT, q);
|
|
2941 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
|
|
2942 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
|
|
2943 return;
|
|
2944 }
|
|
2945 }
|
|
2946
|
|
2947 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
|
|
2948 }
|
|
2949 #endif //VALIDATE_MARK_SWEEP
|
|
2950
|
|
2951 void PSParallelCompact::adjust_pointer(oop* p, bool isroot) {
|
|
2952 oop obj = *p;
|
|
2953 VALIDATE_MARK_SWEEP_ONLY(oop saved_new_pointer = NULL);
|
|
2954 if (obj != NULL) {
|
|
2955 oop new_pointer = (oop) summary_data().calc_new_pointer(obj);
|
|
2956 assert(new_pointer != NULL || // is forwarding ptr?
|
|
2957 obj->is_shared(), // never forwarded?
|
|
2958 "should have a new location");
|
|
2959 // Just always do the update unconditionally?
|
|
2960 if (new_pointer != NULL) {
|
|
2961 *p = new_pointer;
|
|
2962 assert(Universe::heap()->is_in_reserved(new_pointer),
|
|
2963 "should be in object space");
|
|
2964 VALIDATE_MARK_SWEEP_ONLY(saved_new_pointer = new_pointer);
|
|
2965 }
|
|
2966 }
|
|
2967 VALIDATE_MARK_SWEEP_ONLY(track_adjusted_pointer(p, saved_new_pointer, isroot));
|
|
2968 }
|
|
2969
|
|
2970 // Update interior oops in the ranges of chunks [beg_chunk, end_chunk).
|
|
2971 void
|
|
2972 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
|
|
2973 SpaceId space_id,
|
|
2974 size_t beg_chunk,
|
|
2975 size_t end_chunk) {
|
|
2976 ParallelCompactData& sd = summary_data();
|
|
2977 ParMarkBitMap* const mbm = mark_bitmap();
|
|
2978
|
|
2979 HeapWord* beg_addr = sd.chunk_to_addr(beg_chunk);
|
|
2980 HeapWord* const end_addr = sd.chunk_to_addr(end_chunk);
|
|
2981 assert(beg_chunk <= end_chunk, "bad chunk range");
|
|
2982 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
|
|
2983
|
|
2984 #ifdef ASSERT
|
|
2985 // Claim the chunks to avoid triggering an assert when they are marked as
|
|
2986 // filled.
|
|
2987 for (size_t claim_chunk = beg_chunk; claim_chunk < end_chunk; ++claim_chunk) {
|
|
2988 assert(sd.chunk(claim_chunk)->claim_unsafe(), "claim() failed");
|
|
2989 }
|
|
2990 #endif // #ifdef ASSERT
|
|
2991
|
|
2992 if (beg_addr != space(space_id)->bottom()) {
|
|
2993 // Find the first live object or block of dead space that *starts* in this
|
|
2994 // range of chunks. If a partial object crosses onto the chunk, skip it; it
|
|
2995 // will be marked for 'deferred update' when the object head is processed.
|
|
2996 // If dead space crosses onto the chunk, it is also skipped; it will be
|
|
2997 // filled when the prior chunk is processed. If neither of those apply, the
|
|
2998 // first word in the chunk is the start of a live object or dead space.
|
|
2999 assert(beg_addr > space(space_id)->bottom(), "sanity");
|
|
3000 const ChunkData* const cp = sd.chunk(beg_chunk);
|
|
3001 if (cp->partial_obj_size() != 0) {
|
|
3002 beg_addr = sd.partial_obj_end(beg_chunk);
|
|
3003 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
|
|
3004 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
|
|
3005 }
|
|
3006 }
|
|
3007
|
|
3008 if (beg_addr < end_addr) {
|
|
3009 // A live object or block of dead space starts in this range of Chunks.
|
|
3010 HeapWord* const dense_prefix_end = dense_prefix(space_id);
|
|
3011
|
|
3012 // Create closures and iterate.
|
|
3013 UpdateOnlyClosure update_closure(mbm, cm, space_id);
|
|
3014 FillClosure fill_closure(cm, space_id);
|
|
3015 ParMarkBitMap::IterationStatus status;
|
|
3016 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
|
|
3017 dense_prefix_end);
|
|
3018 if (status == ParMarkBitMap::incomplete) {
|
|
3019 update_closure.do_addr(update_closure.source());
|
|
3020 }
|
|
3021 }
|
|
3022
|
|
3023 // Mark the chunks as filled.
|
|
3024 ChunkData* const beg_cp = sd.chunk(beg_chunk);
|
|
3025 ChunkData* const end_cp = sd.chunk(end_chunk);
|
|
3026 for (ChunkData* cp = beg_cp; cp < end_cp; ++cp) {
|
|
3027 cp->set_completed();
|
|
3028 }
|
|
3029 }
|
|
3030
|
|
3031 // Return the SpaceId for the space containing addr. If addr is not in the
|
|
3032 // heap, last_space_id is returned. In debug mode it expects the address to be
|
|
3033 // in the heap and asserts such.
|
|
3034 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
|
|
3035 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
|
|
3036
|
|
3037 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
|
|
3038 if (_space_info[id].space()->contains(addr)) {
|
|
3039 return SpaceId(id);
|
|
3040 }
|
|
3041 }
|
|
3042
|
|
3043 assert(false, "no space contains the addr");
|
|
3044 return last_space_id;
|
|
3045 }
|
|
3046
|
|
3047 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
|
|
3048 SpaceId id) {
|
|
3049 assert(id < last_space_id, "bad space id");
|
|
3050
|
|
3051 ParallelCompactData& sd = summary_data();
|
|
3052 const SpaceInfo* const space_info = _space_info + id;
|
|
3053 ObjectStartArray* const start_array = space_info->start_array();
|
|
3054
|
|
3055 const MutableSpace* const space = space_info->space();
|
|
3056 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
|
|
3057 HeapWord* const beg_addr = space_info->dense_prefix();
|
|
3058 HeapWord* const end_addr = sd.chunk_align_up(space_info->new_top());
|
|
3059
|
|
3060 const ChunkData* const beg_chunk = sd.addr_to_chunk_ptr(beg_addr);
|
|
3061 const ChunkData* const end_chunk = sd.addr_to_chunk_ptr(end_addr);
|
|
3062 const ChunkData* cur_chunk;
|
|
3063 for (cur_chunk = beg_chunk; cur_chunk < end_chunk; ++cur_chunk) {
|
|
3064 HeapWord* const addr = cur_chunk->deferred_obj_addr();
|
|
3065 if (addr != NULL) {
|
|
3066 if (start_array != NULL) {
|
|
3067 start_array->allocate_block(addr);
|
|
3068 }
|
|
3069 oop(addr)->update_contents(cm);
|
|
3070 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
|
|
3071 }
|
|
3072 }
|
|
3073 }
|
|
3074
|
|
3075 // Skip over count live words starting from beg, and return the address of the
|
|
3076 // next live word. Unless marked, the word corresponding to beg is assumed to
|
|
3077 // be dead. Callers must either ensure beg does not correspond to the middle of
|
|
3078 // an object, or account for those live words in some other way. Callers must
|
|
3079 // also ensure that there are enough live words in the range [beg, end) to skip.
|
|
3080 HeapWord*
|
|
3081 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
|
|
3082 {
|
|
3083 assert(count > 0, "sanity");
|
|
3084
|
|
3085 ParMarkBitMap* m = mark_bitmap();
|
|
3086 idx_t bits_to_skip = m->words_to_bits(count);
|
|
3087 idx_t cur_beg = m->addr_to_bit(beg);
|
|
3088 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
|
|
3089
|
|
3090 do {
|
|
3091 cur_beg = m->find_obj_beg(cur_beg, search_end);
|
|
3092 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
|
|
3093 const size_t obj_bits = cur_end - cur_beg + 1;
|
|
3094 if (obj_bits > bits_to_skip) {
|
|
3095 return m->bit_to_addr(cur_beg + bits_to_skip);
|
|
3096 }
|
|
3097 bits_to_skip -= obj_bits;
|
|
3098 cur_beg = cur_end + 1;
|
|
3099 } while (bits_to_skip > 0);
|
|
3100
|
|
3101 // Skipping the desired number of words landed just past the end of an object.
|
|
3102 // Find the start of the next object.
|
|
3103 cur_beg = m->find_obj_beg(cur_beg, search_end);
|
|
3104 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
|
|
3105 return m->bit_to_addr(cur_beg);
|
|
3106 }
|
|
3107
|
|
3108 HeapWord*
|
|
3109 PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
|
|
3110 size_t src_chunk_idx)
|
|
3111 {
|
|
3112 ParMarkBitMap* const bitmap = mark_bitmap();
|
|
3113 const ParallelCompactData& sd = summary_data();
|
|
3114 const size_t ChunkSize = ParallelCompactData::ChunkSize;
|
|
3115
|
|
3116 assert(sd.is_chunk_aligned(dest_addr), "not aligned");
|
|
3117
|
|
3118 const ChunkData* const src_chunk_ptr = sd.chunk(src_chunk_idx);
|
|
3119 const size_t partial_obj_size = src_chunk_ptr->partial_obj_size();
|
|
3120 HeapWord* const src_chunk_destination = src_chunk_ptr->destination();
|
|
3121
|
|
3122 assert(dest_addr >= src_chunk_destination, "wrong src chunk");
|
|
3123 assert(src_chunk_ptr->data_size() > 0, "src chunk cannot be empty");
|
|
3124
|
|
3125 HeapWord* const src_chunk_beg = sd.chunk_to_addr(src_chunk_idx);
|
|
3126 HeapWord* const src_chunk_end = src_chunk_beg + ChunkSize;
|
|
3127
|
|
3128 HeapWord* addr = src_chunk_beg;
|
|
3129 if (dest_addr == src_chunk_destination) {
|
|
3130 // Return the first live word in the source chunk.
|
|
3131 if (partial_obj_size == 0) {
|
|
3132 addr = bitmap->find_obj_beg(addr, src_chunk_end);
|
|
3133 assert(addr < src_chunk_end, "no objects start in src chunk");
|
|
3134 }
|
|
3135 return addr;
|
|
3136 }
|
|
3137
|
|
3138 // Must skip some live data.
|
|
3139 size_t words_to_skip = dest_addr - src_chunk_destination;
|
|
3140 assert(src_chunk_ptr->data_size() > words_to_skip, "wrong src chunk");
|
|
3141
|
|
3142 if (partial_obj_size >= words_to_skip) {
|
|
3143 // All the live words to skip are part of the partial object.
|
|
3144 addr += words_to_skip;
|
|
3145 if (partial_obj_size == words_to_skip) {
|
|
3146 // Find the first live word past the partial object.
|
|
3147 addr = bitmap->find_obj_beg(addr, src_chunk_end);
|
|
3148 assert(addr < src_chunk_end, "wrong src chunk");
|
|
3149 }
|
|
3150 return addr;
|
|
3151 }
|
|
3152
|
|
3153 // Skip over the partial object (if any).
|
|
3154 if (partial_obj_size != 0) {
|
|
3155 words_to_skip -= partial_obj_size;
|
|
3156 addr += partial_obj_size;
|
|
3157 }
|
|
3158
|
|
3159 // Skip over live words due to objects that start in the chunk.
|
|
3160 addr = skip_live_words(addr, src_chunk_end, words_to_skip);
|
|
3161 assert(addr < src_chunk_end, "wrong src chunk");
|
|
3162 return addr;
|
|
3163 }
|
|
3164
|
|
3165 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
|
|
3166 size_t beg_chunk,
|
|
3167 HeapWord* end_addr)
|
|
3168 {
|
|
3169 ParallelCompactData& sd = summary_data();
|
|
3170 ChunkData* const beg = sd.chunk(beg_chunk);
|
|
3171 HeapWord* const end_addr_aligned_up = sd.chunk_align_up(end_addr);
|
|
3172 ChunkData* const end = sd.addr_to_chunk_ptr(end_addr_aligned_up);
|
|
3173 size_t cur_idx = beg_chunk;
|
|
3174 for (ChunkData* cur = beg; cur < end; ++cur, ++cur_idx) {
|
|
3175 assert(cur->data_size() > 0, "chunk must have live data");
|
|
3176 cur->decrement_destination_count();
|
|
3177 if (cur_idx <= cur->source_chunk() && cur->available() && cur->claim()) {
|
|
3178 cm->save_for_processing(cur_idx);
|
|
3179 }
|
|
3180 }
|
|
3181 }
|
|
3182
|
|
3183 size_t PSParallelCompact::next_src_chunk(MoveAndUpdateClosure& closure,
|
|
3184 SpaceId& src_space_id,
|
|
3185 HeapWord*& src_space_top,
|
|
3186 HeapWord* end_addr)
|
|
3187 {
|
|
3188 typedef ParallelCompactData::ChunkData ChunkData;
|
|
3189
|
|
3190 ParallelCompactData& sd = PSParallelCompact::summary_data();
|
|
3191 const size_t chunk_size = ParallelCompactData::ChunkSize;
|
|
3192
|
|
3193 size_t src_chunk_idx = 0;
|
|
3194
|
|
3195 // Skip empty chunks (if any) up to the top of the space.
|
|
3196 HeapWord* const src_aligned_up = sd.chunk_align_up(end_addr);
|
|
3197 ChunkData* src_chunk_ptr = sd.addr_to_chunk_ptr(src_aligned_up);
|
|
3198 HeapWord* const top_aligned_up = sd.chunk_align_up(src_space_top);
|
|
3199 const ChunkData* const top_chunk_ptr = sd.addr_to_chunk_ptr(top_aligned_up);
|
|
3200 while (src_chunk_ptr < top_chunk_ptr && src_chunk_ptr->data_size() == 0) {
|
|
3201 ++src_chunk_ptr;
|
|
3202 }
|
|
3203
|
|
3204 if (src_chunk_ptr < top_chunk_ptr) {
|
|
3205 // The next source chunk is in the current space. Update src_chunk_idx and
|
|
3206 // the source address to match src_chunk_ptr.
|
|
3207 src_chunk_idx = sd.chunk(src_chunk_ptr);
|
|
3208 HeapWord* const src_chunk_addr = sd.chunk_to_addr(src_chunk_idx);
|
|
3209 if (src_chunk_addr > closure.source()) {
|
|
3210 closure.set_source(src_chunk_addr);
|
|
3211 }
|
|
3212 return src_chunk_idx;
|
|
3213 }
|
|
3214
|
|
3215 // Switch to a new source space and find the first non-empty chunk.
|
|
3216 unsigned int space_id = src_space_id + 1;
|
|
3217 assert(space_id < last_space_id, "not enough spaces");
|
|
3218
|
|
3219 HeapWord* const destination = closure.destination();
|
|
3220
|
|
3221 do {
|
|
3222 MutableSpace* space = _space_info[space_id].space();
|
|
3223 HeapWord* const bottom = space->bottom();
|
|
3224 const ChunkData* const bottom_cp = sd.addr_to_chunk_ptr(bottom);
|
|
3225
|
|
3226 // Iterate over the spaces that do not compact into themselves.
|
|
3227 if (bottom_cp->destination() != bottom) {
|
|
3228 HeapWord* const top_aligned_up = sd.chunk_align_up(space->top());
|
|
3229 const ChunkData* const top_cp = sd.addr_to_chunk_ptr(top_aligned_up);
|
|
3230
|
|
3231 for (const ChunkData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
|
|
3232 if (src_cp->live_obj_size() > 0) {
|
|
3233 // Found it.
|
|
3234 assert(src_cp->destination() == destination,
|
|
3235 "first live obj in the space must match the destination");
|
|
3236 assert(src_cp->partial_obj_size() == 0,
|
|
3237 "a space cannot begin with a partial obj");
|
|
3238
|
|
3239 src_space_id = SpaceId(space_id);
|
|
3240 src_space_top = space->top();
|
|
3241 const size_t src_chunk_idx = sd.chunk(src_cp);
|
|
3242 closure.set_source(sd.chunk_to_addr(src_chunk_idx));
|
|
3243 return src_chunk_idx;
|
|
3244 } else {
|
|
3245 assert(src_cp->data_size() == 0, "sanity");
|
|
3246 }
|
|
3247 }
|
|
3248 }
|
|
3249 } while (++space_id < last_space_id);
|
|
3250
|
|
3251 assert(false, "no source chunk was found");
|
|
3252 return 0;
|
|
3253 }
|
|
3254
|
|
3255 void PSParallelCompact::fill_chunk(ParCompactionManager* cm, size_t chunk_idx)
|
|
3256 {
|
|
3257 typedef ParMarkBitMap::IterationStatus IterationStatus;
|
|
3258 const size_t ChunkSize = ParallelCompactData::ChunkSize;
|
|
3259 ParMarkBitMap* const bitmap = mark_bitmap();
|
|
3260 ParallelCompactData& sd = summary_data();
|
|
3261 ChunkData* const chunk_ptr = sd.chunk(chunk_idx);
|
|
3262
|
|
3263 // Get the items needed to construct the closure.
|
|
3264 HeapWord* dest_addr = sd.chunk_to_addr(chunk_idx);
|
|
3265 SpaceId dest_space_id = space_id(dest_addr);
|
|
3266 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
|
|
3267 HeapWord* new_top = _space_info[dest_space_id].new_top();
|
|
3268 assert(dest_addr < new_top, "sanity");
|
|
3269 const size_t words = MIN2(pointer_delta(new_top, dest_addr), ChunkSize);
|
|
3270
|
|
3271 // Get the source chunk and related info.
|
|
3272 size_t src_chunk_idx = chunk_ptr->source_chunk();
|
|
3273 SpaceId src_space_id = space_id(sd.chunk_to_addr(src_chunk_idx));
|
|
3274 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
|
|
3275
|
|
3276 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
|
|
3277 closure.set_source(first_src_addr(dest_addr, src_chunk_idx));
|
|
3278
|
|
3279 // Adjust src_chunk_idx to prepare for decrementing destination counts (the
|
|
3280 // destination count is not decremented when a chunk is copied to itself).
|
|
3281 if (src_chunk_idx == chunk_idx) {
|
|
3282 src_chunk_idx += 1;
|
|
3283 }
|
|
3284
|
|
3285 if (bitmap->is_unmarked(closure.source())) {
|
|
3286 // The first source word is in the middle of an object; copy the remainder
|
|
3287 // of the object or as much as will fit. The fact that pointer updates were
|
|
3288 // deferred will be noted when the object header is processed.
|
|
3289 HeapWord* const old_src_addr = closure.source();
|
|
3290 closure.copy_partial_obj();
|
|
3291 if (closure.is_full()) {
|
|
3292 decrement_destination_counts(cm, src_chunk_idx, closure.source());
|
|
3293 chunk_ptr->set_deferred_obj_addr(NULL);
|
|
3294 chunk_ptr->set_completed();
|
|
3295 return;
|
|
3296 }
|
|
3297
|
|
3298 HeapWord* const end_addr = sd.chunk_align_down(closure.source());
|
|
3299 if (sd.chunk_align_down(old_src_addr) != end_addr) {
|
|
3300 // The partial object was copied from more than one source chunk.
|
|
3301 decrement_destination_counts(cm, src_chunk_idx, end_addr);
|
|
3302
|
|
3303 // Move to the next source chunk, possibly switching spaces as well. All
|
|
3304 // args except end_addr may be modified.
|
|
3305 src_chunk_idx = next_src_chunk(closure, src_space_id, src_space_top,
|
|
3306 end_addr);
|
|
3307 }
|
|
3308 }
|
|
3309
|
|
3310 do {
|
|
3311 HeapWord* const cur_addr = closure.source();
|
|
3312 HeapWord* const end_addr = MIN2(sd.chunk_align_up(cur_addr + 1),
|
|
3313 src_space_top);
|
|
3314 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
|
|
3315
|
|
3316 if (status == ParMarkBitMap::incomplete) {
|
|
3317 // The last obj that starts in the source chunk does not end in the chunk.
|
|
3318 assert(closure.source() < end_addr, "sanity")
|
|
3319 HeapWord* const obj_beg = closure.source();
|
|
3320 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
|
|
3321 src_space_top);
|
|
3322 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
|
|
3323 if (obj_end < range_end) {
|
|
3324 // The end was found; the entire object will fit.
|
|
3325 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
|
|
3326 assert(status != ParMarkBitMap::would_overflow, "sanity");
|
|
3327 } else {
|
|
3328 // The end was not found; the object will not fit.
|
|
3329 assert(range_end < src_space_top, "obj cannot cross space boundary");
|
|
3330 status = ParMarkBitMap::would_overflow;
|
|
3331 }
|
|
3332 }
|
|
3333
|
|
3334 if (status == ParMarkBitMap::would_overflow) {
|
|
3335 // The last object did not fit. Note that interior oop updates were
|
|
3336 // deferred, then copy enough of the object to fill the chunk.
|
|
3337 chunk_ptr->set_deferred_obj_addr(closure.destination());
|
|
3338 status = closure.copy_until_full(); // copies from closure.source()
|
|
3339
|
|
3340 decrement_destination_counts(cm, src_chunk_idx, closure.source());
|
|
3341 chunk_ptr->set_completed();
|
|
3342 return;
|
|
3343 }
|
|
3344
|
|
3345 if (status == ParMarkBitMap::full) {
|
|
3346 decrement_destination_counts(cm, src_chunk_idx, closure.source());
|
|
3347 chunk_ptr->set_deferred_obj_addr(NULL);
|
|
3348 chunk_ptr->set_completed();
|
|
3349 return;
|
|
3350 }
|
|
3351
|
|
3352 decrement_destination_counts(cm, src_chunk_idx, end_addr);
|
|
3353
|
|
3354 // Move to the next source chunk, possibly switching spaces as well. All
|
|
3355 // args except end_addr may be modified.
|
|
3356 src_chunk_idx = next_src_chunk(closure, src_space_id, src_space_top,
|
|
3357 end_addr);
|
|
3358 } while (true);
|
|
3359 }
|
|
3360
|
|
3361 void
|
|
3362 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
|
|
3363 const MutableSpace* sp = space(space_id);
|
|
3364 if (sp->is_empty()) {
|
|
3365 return;
|
|
3366 }
|
|
3367
|
|
3368 ParallelCompactData& sd = PSParallelCompact::summary_data();
|
|
3369 ParMarkBitMap* const bitmap = mark_bitmap();
|
|
3370 HeapWord* const dp_addr = dense_prefix(space_id);
|
|
3371 HeapWord* beg_addr = sp->bottom();
|
|
3372 HeapWord* end_addr = sp->top();
|
|
3373
|
|
3374 #ifdef ASSERT
|
|
3375 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
|
|
3376 if (cm->should_verify_only()) {
|
|
3377 VerifyUpdateClosure verify_update(cm, sp);
|
|
3378 bitmap->iterate(&verify_update, beg_addr, end_addr);
|
|
3379 return;
|
|
3380 }
|
|
3381
|
|
3382 if (cm->should_reset_only()) {
|
|
3383 ResetObjectsClosure reset_objects(cm);
|
|
3384 bitmap->iterate(&reset_objects, beg_addr, end_addr);
|
|
3385 return;
|
|
3386 }
|
|
3387 #endif
|
|
3388
|
|
3389 const size_t beg_chunk = sd.addr_to_chunk_idx(beg_addr);
|
|
3390 const size_t dp_chunk = sd.addr_to_chunk_idx(dp_addr);
|
|
3391 if (beg_chunk < dp_chunk) {
|
|
3392 update_and_deadwood_in_dense_prefix(cm, space_id, beg_chunk, dp_chunk);
|
|
3393 }
|
|
3394
|
|
3395 // The destination of the first live object that starts in the chunk is one
|
|
3396 // past the end of the partial object entering the chunk (if any).
|
|
3397 HeapWord* const dest_addr = sd.partial_obj_end(dp_chunk);
|
|
3398 HeapWord* const new_top = _space_info[space_id].new_top();
|
|
3399 assert(new_top >= dest_addr, "bad new_top value");
|
|
3400 const size_t words = pointer_delta(new_top, dest_addr);
|
|
3401
|
|
3402 if (words > 0) {
|
|
3403 ObjectStartArray* start_array = _space_info[space_id].start_array();
|
|
3404 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
|
|
3405
|
|
3406 ParMarkBitMap::IterationStatus status;
|
|
3407 status = bitmap->iterate(&closure, dest_addr, end_addr);
|
|
3408 assert(status == ParMarkBitMap::full, "iteration not complete");
|
|
3409 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
|
|
3410 "live objects skipped because closure is full");
|
|
3411 }
|
|
3412 }
|
|
3413
|
|
3414 jlong PSParallelCompact::millis_since_last_gc() {
|
|
3415 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc;
|
|
3416 // XXX See note in genCollectedHeap::millis_since_last_gc().
|
|
3417 if (ret_val < 0) {
|
|
3418 NOT_PRODUCT(warning("time warp: %d", ret_val);)
|
|
3419 return 0;
|
|
3420 }
|
|
3421 return ret_val;
|
|
3422 }
|
|
3423
|
|
3424 void PSParallelCompact::reset_millis_since_last_gc() {
|
|
3425 _time_of_last_gc = os::javaTimeMillis();
|
|
3426 }
|
|
3427
|
|
3428 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
|
|
3429 {
|
|
3430 if (source() != destination()) {
|
|
3431 assert(source() > destination(), "must copy to the left");
|
|
3432 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
|
|
3433 }
|
|
3434 update_state(words_remaining());
|
|
3435 assert(is_full(), "sanity");
|
|
3436 return ParMarkBitMap::full;
|
|
3437 }
|
|
3438
|
|
3439 void MoveAndUpdateClosure::copy_partial_obj()
|
|
3440 {
|
|
3441 size_t words = words_remaining();
|
|
3442
|
|
3443 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
|
|
3444 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
|
|
3445 if (end_addr < range_end) {
|
|
3446 words = bitmap()->obj_size(source(), end_addr);
|
|
3447 }
|
|
3448
|
|
3449 // This test is necessary; if omitted, the pointer updates to a partial object
|
|
3450 // that crosses the dense prefix boundary could be overwritten.
|
|
3451 if (source() != destination()) {
|
|
3452 assert(source() > destination(), "must copy to the left");
|
|
3453 Copy::aligned_conjoint_words(source(), destination(), words);
|
|
3454 }
|
|
3455 update_state(words);
|
|
3456 }
|
|
3457
|
|
3458 ParMarkBitMapClosure::IterationStatus
|
|
3459 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
|
|
3460 assert(destination() != NULL, "sanity");
|
|
3461 assert(bitmap()->obj_size(addr) == words, "bad size");
|
|
3462
|
|
3463 _source = addr;
|
|
3464 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
|
|
3465 destination(), "wrong destination");
|
|
3466
|
|
3467 if (words > words_remaining()) {
|
|
3468 return ParMarkBitMap::would_overflow;
|
|
3469 }
|
|
3470
|
|
3471 // The start_array must be updated even if the object is not moving.
|
|
3472 if (_start_array != NULL) {
|
|
3473 _start_array->allocate_block(destination());
|
|
3474 }
|
|
3475
|
|
3476 if (destination() != source()) {
|
|
3477 assert(destination() < source(), "must copy to the left");
|
|
3478 Copy::aligned_conjoint_words(source(), destination(), words);
|
|
3479 }
|
|
3480
|
|
3481 oop moved_oop = (oop) destination();
|
|
3482 moved_oop->update_contents(compaction_manager());
|
|
3483 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
|
|
3484
|
|
3485 update_state(words);
|
|
3486 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
|
|
3487 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
|
|
3488 }
|
|
3489
|
|
3490 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
|
|
3491 ParCompactionManager* cm,
|
|
3492 PSParallelCompact::SpaceId space_id) :
|
|
3493 ParMarkBitMapClosure(mbm, cm),
|
|
3494 _space_id(space_id),
|
|
3495 _start_array(PSParallelCompact::start_array(space_id))
|
|
3496 {
|
|
3497 }
|
|
3498
|
|
3499 // Updates the references in the object to their new values.
|
|
3500 ParMarkBitMapClosure::IterationStatus
|
|
3501 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
|
|
3502 do_addr(addr);
|
|
3503 return ParMarkBitMap::incomplete;
|
|
3504 }
|
|
3505
|
|
3506 BitBlockUpdateClosure::BitBlockUpdateClosure(ParMarkBitMap* mbm,
|
|
3507 ParCompactionManager* cm,
|
|
3508 size_t chunk_index) :
|
|
3509 ParMarkBitMapClosure(mbm, cm),
|
|
3510 _live_data_left(0),
|
|
3511 _cur_block(0) {
|
|
3512 _chunk_start =
|
|
3513 PSParallelCompact::summary_data().chunk_to_addr(chunk_index);
|
|
3514 _chunk_end =
|
|
3515 PSParallelCompact::summary_data().chunk_to_addr(chunk_index) +
|
|
3516 ParallelCompactData::ChunkSize;
|
|
3517 _chunk_index = chunk_index;
|
|
3518 _cur_block =
|
|
3519 PSParallelCompact::summary_data().addr_to_block_idx(_chunk_start);
|
|
3520 }
|
|
3521
|
|
3522 bool BitBlockUpdateClosure::chunk_contains_cur_block() {
|
|
3523 return ParallelCompactData::chunk_contains_block(_chunk_index, _cur_block);
|
|
3524 }
|
|
3525
|
|
3526 void BitBlockUpdateClosure::reset_chunk(size_t chunk_index) {
|
|
3527 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(7);)
|
|
3528 ParallelCompactData& sd = PSParallelCompact::summary_data();
|
|
3529 _chunk_index = chunk_index;
|
|
3530 _live_data_left = 0;
|
|
3531 _chunk_start = sd.chunk_to_addr(chunk_index);
|
|
3532 _chunk_end = sd.chunk_to_addr(chunk_index) + ParallelCompactData::ChunkSize;
|
|
3533
|
|
3534 // The first block in this chunk
|
|
3535 size_t first_block = sd.addr_to_block_idx(_chunk_start);
|
|
3536 size_t partial_live_size = sd.chunk(chunk_index)->partial_obj_size();
|
|
3537
|
|
3538 // Set the offset to 0. By definition it should have that value
|
|
3539 // but it may have been written while processing an earlier chunk.
|
|
3540 if (partial_live_size == 0) {
|
|
3541 // No live object extends onto the chunk. The first bit
|
|
3542 // in the bit map for the first chunk must be a start bit.
|
|
3543 // Although there may not be any marked bits, it is safe
|
|
3544 // to set it as a start bit.
|
|
3545 sd.block(first_block)->set_start_bit_offset(0);
|
|
3546 sd.block(first_block)->set_first_is_start_bit(true);
|
|
3547 } else if (sd.partial_obj_ends_in_block(first_block)) {
|
|
3548 sd.block(first_block)->set_end_bit_offset(0);
|
|
3549 sd.block(first_block)->set_first_is_start_bit(false);
|
|
3550 } else {
|
|
3551 // The partial object extends beyond the first block.
|
|
3552 // There is no object starting in the first block
|
|
3553 // so the offset and bit parity are not needed.
|
|
3554 // Set the the bit parity to start bit so assertions
|
|
3555 // work when not bit is found.
|
|
3556 sd.block(first_block)->set_end_bit_offset(0);
|
|
3557 sd.block(first_block)->set_first_is_start_bit(false);
|
|
3558 }
|
|
3559 _cur_block = first_block;
|
|
3560 #ifdef ASSERT
|
|
3561 if (sd.block(first_block)->first_is_start_bit()) {
|
|
3562 assert(!sd.partial_obj_ends_in_block(first_block),
|
|
3563 "Partial object cannot end in first block");
|
|
3564 }
|
|
3565
|
|
3566 if (PrintGCDetails && Verbose) {
|
|
3567 if (partial_live_size == 1) {
|
|
3568 gclog_or_tty->print_cr("first_block " PTR_FORMAT
|
|
3569 " _offset " PTR_FORMAT
|
|
3570 " _first_is_start_bit %d",
|
|
3571 first_block,
|
|
3572 sd.block(first_block)->raw_offset(),
|
|
3573 sd.block(first_block)->first_is_start_bit());
|
|
3574 }
|
|
3575 }
|
|
3576 #endif
|
|
3577 DEBUG_ONLY(ParallelCompactData::BlockData::set_cur_phase(17);)
|
|
3578 }
|
|
3579
|
|
3580 // This method is called when a object has been found (both beginning
|
|
3581 // and end of the object) in the range of iteration. This method is
|
|
3582 // calculating the words of live data to the left of a block. That live
|
|
3583 // data includes any object starting to the left of the block (i.e.,
|
|
3584 // the live-data-to-the-left of block AAA will include the full size
|
|
3585 // of any object entering AAA).
|
|
3586
|
|
3587 ParMarkBitMapClosure::IterationStatus
|
|
3588 BitBlockUpdateClosure::do_addr(HeapWord* addr, size_t words) {
|
|
3589 // add the size to the block data.
|
|
3590 HeapWord* obj = addr;
|
|
3591 ParallelCompactData& sd = PSParallelCompact::summary_data();
|
|
3592
|
|
3593 assert(bitmap()->obj_size(obj) == words, "bad size");
|
|
3594 assert(_chunk_start <= obj, "object is not in chunk");
|
|
3595 assert(obj + words <= _chunk_end, "object is not in chunk");
|
|
3596
|
|
3597 // Update the live data to the left
|
|
3598 size_t prev_live_data_left = _live_data_left;
|
|
3599 _live_data_left = _live_data_left + words;
|
|
3600
|
|
3601 // Is this object in the current block.
|
|
3602 size_t block_of_obj = sd.addr_to_block_idx(obj);
|
|
3603 size_t block_of_obj_last = sd.addr_to_block_idx(obj + words - 1);
|
|
3604 HeapWord* block_of_obj_last_addr = sd.block_to_addr(block_of_obj_last);
|
|
3605 if (_cur_block < block_of_obj) {
|
|
3606
|
|
3607 //
|
|
3608 // No object crossed the block boundary and this object was found
|
|
3609 // on the other side of the block boundary. Update the offset for
|
|
3610 // the new block with the data size that does not include this object.
|
|
3611 //
|
|
3612 // The first bit in block_of_obj is a start bit except in the
|
|
3613 // case where the partial object for the chunk extends into
|
|
3614 // this block.
|
|
3615 if (sd.partial_obj_ends_in_block(block_of_obj)) {
|
|
3616 sd.block(block_of_obj)->set_end_bit_offset(prev_live_data_left);
|
|
3617 } else {
|
|
3618 sd.block(block_of_obj)->set_start_bit_offset(prev_live_data_left);
|
|
3619 }
|
|
3620
|
|
3621 // Does this object pass beyond the its block?
|
|
3622 if (block_of_obj < block_of_obj_last) {
|
|
3623 // Object crosses block boundary. Two blocks need to be udpated:
|
|
3624 // the current block where the object started
|
|
3625 // the block where the object ends
|
|
3626 //
|
|
3627 // The offset for blocks with no objects starting in them
|
|
3628 // (e.g., blocks between _cur_block and block_of_obj_last)
|
|
3629 // should not be needed.
|
|
3630 // Note that block_of_obj_last may be in another chunk. If so,
|
|
3631 // it should be overwritten later. This is a problem (writting
|
|
3632 // into a block in a later chunk) for parallel execution.
|
|
3633 assert(obj < block_of_obj_last_addr,
|
|
3634 "Object should start in previous block");
|
|
3635
|
|
3636 // obj is crossing into block_of_obj_last so the first bit
|
|
3637 // is and end bit.
|
|
3638 sd.block(block_of_obj_last)->set_end_bit_offset(_live_data_left);
|
|
3639
|
|
3640 _cur_block = block_of_obj_last;
|
|
3641 } else {
|
|
3642 // _first_is_start_bit has already been set correctly
|
|
3643 // in the if-then-else above so don't reset it here.
|
|
3644 _cur_block = block_of_obj;
|
|
3645 }
|
|
3646 } else {
|
|
3647 // The current block only changes if the object extends beyound
|
|
3648 // the block it starts in.
|
|
3649 //
|
|
3650 // The object starts in the current block.
|
|
3651 // Does this object pass beyond the end of it?
|
|
3652 if (block_of_obj < block_of_obj_last) {
|
|
3653 // Object crosses block boundary.
|
|
3654 // See note above on possible blocks between block_of_obj and
|
|
3655 // block_of_obj_last
|
|
3656 assert(obj < block_of_obj_last_addr,
|
|
3657 "Object should start in previous block");
|
|
3658
|
|
3659 sd.block(block_of_obj_last)->set_end_bit_offset(_live_data_left);
|
|
3660
|
|
3661 _cur_block = block_of_obj_last;
|
|
3662 }
|
|
3663 }
|
|
3664
|
|
3665 // Return incomplete if there are more blocks to be done.
|
|
3666 if (chunk_contains_cur_block()) {
|
|
3667 return ParMarkBitMap::incomplete;
|
|
3668 }
|
|
3669 return ParMarkBitMap::complete;
|
|
3670 }
|
|
3671
|
|
3672 // Verify the new location using the forwarding pointer
|
|
3673 // from MarkSweep::mark_sweep_phase2(). Set the mark_word
|
|
3674 // to the initial value.
|
|
3675 ParMarkBitMapClosure::IterationStatus
|
|
3676 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) {
|
|
3677 // The second arg (words) is not used.
|
|
3678 oop obj = (oop) addr;
|
|
3679 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer();
|
|
3680 HeapWord* new_pointer = summary_data().calc_new_pointer(obj);
|
|
3681 if (forwarding_ptr == NULL) {
|
|
3682 // The object is dead or not moving.
|
|
3683 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj),
|
|
3684 "Object liveness is wrong.");
|
|
3685 return ParMarkBitMap::incomplete;
|
|
3686 }
|
|
3687 assert(UseParallelOldGCDensePrefix ||
|
|
3688 (HeapMaximumCompactionInterval > 1) ||
|
|
3689 (MarkSweepAlwaysCompactCount > 1) ||
|
|
3690 (forwarding_ptr == new_pointer),
|
|
3691 "Calculation of new location is incorrect");
|
|
3692 return ParMarkBitMap::incomplete;
|
|
3693 }
|
|
3694
|
|
3695 // Reset objects modified for debug checking.
|
|
3696 ParMarkBitMapClosure::IterationStatus
|
|
3697 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) {
|
|
3698 // The second arg (words) is not used.
|
|
3699 oop obj = (oop) addr;
|
|
3700 obj->init_mark();
|
|
3701 return ParMarkBitMap::incomplete;
|
|
3702 }
|
|
3703
|
|
3704 // Prepare for compaction. This method is executed once
|
|
3705 // (i.e., by a single thread) before compaction.
|
|
3706 // Save the updated location of the intArrayKlassObj for
|
|
3707 // filling holes in the dense prefix.
|
|
3708 void PSParallelCompact::compact_prologue() {
|
|
3709 _updated_int_array_klass_obj = (klassOop)
|
|
3710 summary_data().calc_new_pointer(Universe::intArrayKlassObj());
|
|
3711 }
|
|
3712
|
|
3713 // The initial implementation of this method created a field
|
|
3714 // _next_compaction_space_id in SpaceInfo and initialized
|
|
3715 // that field in SpaceInfo::initialize_space_info(). That
|
|
3716 // required that _next_compaction_space_id be declared a
|
|
3717 // SpaceId in SpaceInfo and that would have required that
|
|
3718 // either SpaceId be declared in a separate class or that
|
|
3719 // it be declared in SpaceInfo. It didn't seem consistent
|
|
3720 // to declare it in SpaceInfo (didn't really fit logically).
|
|
3721 // Alternatively, defining a separate class to define SpaceId
|
|
3722 // seem excessive. This implementation is simple and localizes
|
|
3723 // the knowledge.
|
|
3724
|
|
3725 PSParallelCompact::SpaceId
|
|
3726 PSParallelCompact::next_compaction_space_id(SpaceId id) {
|
|
3727 assert(id < last_space_id, "id out of range");
|
|
3728 switch (id) {
|
|
3729 case perm_space_id :
|
|
3730 return last_space_id;
|
|
3731 case old_space_id :
|
|
3732 return eden_space_id;
|
|
3733 case eden_space_id :
|
|
3734 return from_space_id;
|
|
3735 case from_space_id :
|
|
3736 return to_space_id;
|
|
3737 case to_space_id :
|
|
3738 return last_space_id;
|
|
3739 default:
|
|
3740 assert(false, "Bad space id");
|
|
3741 return last_space_id;
|
|
3742 }
|
|
3743 }
|
|
3744
|
|
3745 // Here temporarily for debugging
|
|
3746 #ifdef ASSERT
|
|
3747 size_t ParallelCompactData::block_idx(BlockData* block) {
|
|
3748 size_t index = pointer_delta(block,
|
|
3749 PSParallelCompact::summary_data()._block_data, sizeof(BlockData));
|
|
3750 return index;
|
|
3751 }
|
|
3752 #endif
|