Mercurial > hg > truffle
annotate src/share/vm/gc_implementation/parNew/parGCAllocBuffer.cpp @ 113:ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
Summary: Compressed oops in instances, arrays, and headers. Code contributors are coleenp, phh, never, swamyv
Reviewed-by: jmasa, kamg, acorn, tbell, kvn, rasbold
| author | coleenp |
|---|---|
| date | Sun, 13 Apr 2008 17:43:42 -0400 |
| parents | a61af66fc99e |
| children | d1605aabd0a1 |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 2 * Copyright 2001-2006 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/_parGCAllocBuffer.cpp.incl" | |
| 27 | |
| 28 ParGCAllocBuffer::ParGCAllocBuffer(size_t desired_plab_sz_) : | |
| 29 _word_sz(desired_plab_sz_), _bottom(NULL), _top(NULL), | |
| 30 _end(NULL), _hard_end(NULL), | |
| 31 _retained(false), _retained_filler(), | |
| 32 _allocated(0), _wasted(0) | |
| 33 { | |
| 34 assert (min_size() > AlignmentReserve, "Inconsistency!"); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
0
diff
changeset
|
35 // arrayOopDesc::header_size depends on command line initialization. |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
0
diff
changeset
|
36 FillerHeaderSize = align_object_size(arrayOopDesc::header_size(T_INT)); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
0
diff
changeset
|
37 AlignmentReserve = oopDesc::header_size() > MinObjAlignment ? FillerHeaderSize : 0; |
| 0 | 38 } |
| 39 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
0
diff
changeset
|
40 size_t ParGCAllocBuffer::FillerHeaderSize; |
| 0 | 41 |
| 42 // If the minimum object size is greater than MinObjAlignment, we can | |
| 43 // end up with a shard at the end of the buffer that's smaller than | |
| 44 // the smallest object. We can't allow that because the buffer must | |
| 45 // look like it's full of objects when we retire it, so we make | |
| 46 // sure we have enough space for a filler int array object. | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
0
diff
changeset
|
47 size_t ParGCAllocBuffer::AlignmentReserve; |
| 0 | 48 |
| 49 void ParGCAllocBuffer::retire(bool end_of_gc, bool retain) { | |
| 50 assert(!retain || end_of_gc, "Can only retain at GC end."); | |
| 51 if (_retained) { | |
| 52 // If the buffer had been retained shorten the previous filler object. | |
| 53 assert(_retained_filler.end() <= _top, "INVARIANT"); | |
| 54 SharedHeap::fill_region_with_object(_retained_filler); | |
| 55 // Wasted space book-keeping, otherwise (normally) done in invalidate() | |
| 56 _wasted += _retained_filler.word_size(); | |
| 57 _retained = false; | |
| 58 } | |
| 59 assert(!end_of_gc || !_retained, "At this point, end_of_gc ==> !_retained."); | |
| 60 if (_top < _hard_end) { | |
| 61 SharedHeap::fill_region_with_object(MemRegion(_top, _hard_end)); | |
| 62 if (!retain) { | |
| 63 invalidate(); | |
| 64 } else { | |
| 65 // Is there wasted space we'd like to retain for the next GC? | |
| 66 if (pointer_delta(_end, _top) > FillerHeaderSize) { | |
| 67 _retained = true; | |
| 68 _retained_filler = MemRegion(_top, FillerHeaderSize); | |
| 69 _top = _top + FillerHeaderSize; | |
| 70 } else { | |
| 71 invalidate(); | |
| 72 } | |
| 73 } | |
| 74 } | |
| 75 } | |
| 76 | |
| 77 void ParGCAllocBuffer::flush_stats(PLABStats* stats) { | |
| 78 assert(ResizePLAB, "Wasted work"); | |
| 79 stats->add_allocated(_allocated); | |
| 80 stats->add_wasted(_wasted); | |
| 81 stats->add_unused(pointer_delta(_end, _top)); | |
| 82 } | |
| 83 | |
| 84 // Compute desired plab size and latch result for later | |
| 85 // use. This should be called once at the end of parallel | |
| 86 // scavenge; it clears the sensor accumulators. | |
| 87 void PLABStats::adjust_desired_plab_sz() { | |
| 88 assert(ResizePLAB, "Not set"); | |
| 89 if (_allocated == 0) { | |
| 90 assert(_unused == 0, "Inconsistency in PLAB stats"); | |
| 91 _allocated = 1; | |
| 92 } | |
| 93 double wasted_frac = (double)_unused/(double)_allocated; | |
| 94 size_t target_refills = (size_t)((wasted_frac*TargetSurvivorRatio)/ | |
| 95 TargetPLABWastePct); | |
| 96 if (target_refills == 0) { | |
| 97 target_refills = 1; | |
| 98 } | |
| 99 _used = _allocated - _wasted - _unused; | |
| 100 size_t plab_sz = _used/(target_refills*ParallelGCThreads); | |
| 101 if (PrintPLAB) gclog_or_tty->print(" (plab_sz = %d ", plab_sz); | |
| 102 // Take historical weighted average | |
| 103 _filter.sample(plab_sz); | |
| 104 // Clip from above and below, and align to object boundary | |
| 105 plab_sz = MAX2(min_size(), (size_t)_filter.average()); | |
| 106 plab_sz = MIN2(max_size(), plab_sz); | |
| 107 plab_sz = align_object_size(plab_sz); | |
| 108 // Latch the result | |
| 109 if (PrintPLAB) gclog_or_tty->print(" desired_plab_sz = %d) ", plab_sz); | |
| 110 if (ResizePLAB) { | |
| 111 _desired_plab_sz = plab_sz; | |
| 112 } | |
| 113 // Now clear the accumulators for next round: | |
| 114 // note this needs to be fixed in the case where we | |
| 115 // are retaining across scavenges. FIX ME !!! XXX | |
| 116 _allocated = 0; | |
| 117 _wasted = 0; | |
| 118 _unused = 0; | |
| 119 } | |
| 120 | |
| 121 #ifndef PRODUCT | |
| 122 void ParGCAllocBuffer::print() { | |
| 123 gclog_or_tty->print("parGCAllocBuffer: _bottom: %p _top: %p _end: %p _hard_end: %p" | |
| 124 "_retained: %c _retained_filler: [%p,%p)\n", | |
| 125 _bottom, _top, _end, _hard_end, | |
| 126 "FT"[_retained], _retained_filler.start(), _retained_filler.end()); | |
| 127 } | |
| 128 #endif // !PRODUCT | |
| 129 | |
| 130 const size_t ParGCAllocBufferWithBOT::ChunkSizeInWords = | |
| 131 MIN2(CardTableModRefBS::par_chunk_heapword_alignment(), | |
| 132 ((size_t)Generation::GenGrain)/HeapWordSize); | |
| 133 const size_t ParGCAllocBufferWithBOT::ChunkSizeInBytes = | |
| 134 MIN2(CardTableModRefBS::par_chunk_heapword_alignment() * HeapWordSize, | |
| 135 (size_t)Generation::GenGrain); | |
| 136 | |
| 137 ParGCAllocBufferWithBOT::ParGCAllocBufferWithBOT(size_t word_sz, | |
| 138 BlockOffsetSharedArray* bsa) : | |
| 139 ParGCAllocBuffer(word_sz), | |
| 140 _bsa(bsa), | |
| 141 _bt(bsa, MemRegion(_bottom, _hard_end)), | |
| 142 _true_end(_hard_end) | |
| 143 {} | |
| 144 | |
| 145 // The buffer comes with its own BOT, with a shared (obviously) underlying | |
| 146 // BlockOffsetSharedArray. We manipulate this BOT in the normal way | |
| 147 // as we would for any contiguous space. However, on accasion we | |
| 148 // need to do some buffer surgery at the extremities before we | |
| 149 // start using the body of the buffer for allocations. Such surgery | |
| 150 // (as explained elsewhere) is to prevent allocation on a card that | |
| 151 // is in the process of being walked concurrently by another GC thread. | |
| 152 // When such surgery happens at a point that is far removed (to the | |
| 153 // right of the current allocation point, top), we use the "contig" | |
| 154 // parameter below to directly manipulate the shared array without | |
| 155 // modifying the _next_threshold state in the BOT. | |
| 156 void ParGCAllocBufferWithBOT::fill_region_with_block(MemRegion mr, | |
| 157 bool contig) { | |
| 158 SharedHeap::fill_region_with_object(mr); | |
| 159 if (contig) { | |
| 160 _bt.alloc_block(mr.start(), mr.end()); | |
| 161 } else { | |
| 162 _bt.BlockOffsetArray::alloc_block(mr.start(), mr.end()); | |
| 163 } | |
| 164 } | |
| 165 | |
| 166 HeapWord* ParGCAllocBufferWithBOT::allocate_slow(size_t word_sz) { | |
| 167 HeapWord* res = NULL; | |
| 168 if (_true_end > _hard_end) { | |
| 169 assert((HeapWord*)align_size_down(intptr_t(_hard_end), | |
| 170 ChunkSizeInBytes) == _hard_end, | |
| 171 "or else _true_end should be equal to _hard_end"); | |
| 172 assert(_retained, "or else _true_end should be equal to _hard_end"); | |
| 173 assert(_retained_filler.end() <= _top, "INVARIANT"); | |
| 174 SharedHeap::fill_region_with_object(_retained_filler); | |
| 175 if (_top < _hard_end) { | |
| 176 fill_region_with_block(MemRegion(_top, _hard_end), true); | |
| 177 } | |
| 178 HeapWord* next_hard_end = MIN2(_true_end, _hard_end + ChunkSizeInWords); | |
| 179 _retained_filler = MemRegion(_hard_end, FillerHeaderSize); | |
| 180 _bt.alloc_block(_retained_filler.start(), _retained_filler.word_size()); | |
| 181 _top = _retained_filler.end(); | |
| 182 _hard_end = next_hard_end; | |
| 183 _end = _hard_end - AlignmentReserve; | |
| 184 res = ParGCAllocBuffer::allocate(word_sz); | |
| 185 if (res != NULL) { | |
| 186 _bt.alloc_block(res, word_sz); | |
| 187 } | |
| 188 } | |
| 189 return res; | |
| 190 } | |
| 191 | |
| 192 void | |
| 193 ParGCAllocBufferWithBOT::undo_allocation(HeapWord* obj, size_t word_sz) { | |
| 194 ParGCAllocBuffer::undo_allocation(obj, word_sz); | |
| 195 // This may back us up beyond the previous threshold, so reset. | |
| 196 _bt.set_region(MemRegion(_top, _hard_end)); | |
| 197 _bt.initialize_threshold(); | |
| 198 } | |
| 199 | |
| 200 void ParGCAllocBufferWithBOT::retire(bool end_of_gc, bool retain) { | |
| 201 assert(!retain || end_of_gc, "Can only retain at GC end."); | |
| 202 if (_retained) { | |
| 203 // We're about to make the retained_filler into a block. | |
| 204 _bt.BlockOffsetArray::alloc_block(_retained_filler.start(), | |
| 205 _retained_filler.end()); | |
| 206 } | |
| 207 // Reset _hard_end to _true_end (and update _end) | |
| 208 if (retain && _hard_end != NULL) { | |
| 209 assert(_hard_end <= _true_end, "Invariant."); | |
| 210 _hard_end = _true_end; | |
| 211 _end = MAX2(_top, _hard_end - AlignmentReserve); | |
| 212 assert(_end <= _hard_end, "Invariant."); | |
| 213 } | |
| 214 _true_end = _hard_end; | |
| 215 HeapWord* pre_top = _top; | |
| 216 | |
| 217 ParGCAllocBuffer::retire(end_of_gc, retain); | |
| 218 // Now any old _retained_filler is cut back to size, the free part is | |
| 219 // filled with a filler object, and top is past the header of that | |
| 220 // object. | |
| 221 | |
| 222 if (retain && _top < _end) { | |
| 223 assert(end_of_gc && retain, "Or else retain should be false."); | |
| 224 // If the lab does not start on a card boundary, we don't want to | |
| 225 // allocate onto that card, since that might lead to concurrent | |
| 226 // allocation and card scanning, which we don't support. So we fill | |
| 227 // the first card with a garbage object. | |
| 228 size_t first_card_index = _bsa->index_for(pre_top); | |
| 229 HeapWord* first_card_start = _bsa->address_for_index(first_card_index); | |
| 230 if (first_card_start < pre_top) { | |
| 231 HeapWord* second_card_start = | |
| 232 _bsa->address_for_index(first_card_index + 1); | |
| 233 | |
| 234 // Ensure enough room to fill with the smallest block | |
| 235 second_card_start = MAX2(second_card_start, pre_top + AlignmentReserve); | |
| 236 | |
| 237 // If the end is already in the first card, don't go beyond it! | |
| 238 // Or if the remainder is too small for a filler object, gobble it up. | |
| 239 if (_hard_end < second_card_start || | |
| 240 pointer_delta(_hard_end, second_card_start) < AlignmentReserve) { | |
| 241 second_card_start = _hard_end; | |
| 242 } | |
| 243 if (pre_top < second_card_start) { | |
| 244 MemRegion first_card_suffix(pre_top, second_card_start); | |
| 245 fill_region_with_block(first_card_suffix, true); | |
| 246 } | |
| 247 pre_top = second_card_start; | |
| 248 _top = pre_top; | |
| 249 _end = MAX2(_top, _hard_end - AlignmentReserve); | |
| 250 } | |
| 251 | |
| 252 // If the lab does not end on a card boundary, we don't want to | |
| 253 // allocate onto that card, since that might lead to concurrent | |
| 254 // allocation and card scanning, which we don't support. So we fill | |
| 255 // the last card with a garbage object. | |
| 256 size_t last_card_index = _bsa->index_for(_hard_end); | |
| 257 HeapWord* last_card_start = _bsa->address_for_index(last_card_index); | |
| 258 if (last_card_start < _hard_end) { | |
| 259 | |
| 260 // Ensure enough room to fill with the smallest block | |
| 261 last_card_start = MIN2(last_card_start, _hard_end - AlignmentReserve); | |
| 262 | |
| 263 // If the top is already in the last card, don't go back beyond it! | |
| 264 // Or if the remainder is too small for a filler object, gobble it up. | |
| 265 if (_top > last_card_start || | |
| 266 pointer_delta(last_card_start, _top) < AlignmentReserve) { | |
| 267 last_card_start = _top; | |
| 268 } | |
| 269 if (last_card_start < _hard_end) { | |
| 270 MemRegion last_card_prefix(last_card_start, _hard_end); | |
| 271 fill_region_with_block(last_card_prefix, false); | |
| 272 } | |
| 273 _hard_end = last_card_start; | |
| 274 _end = MAX2(_top, _hard_end - AlignmentReserve); | |
| 275 _true_end = _hard_end; | |
| 276 assert(_end <= _hard_end, "Invariant."); | |
| 277 } | |
| 278 | |
| 279 // At this point: | |
| 280 // 1) we had a filler object from the original top to hard_end. | |
| 281 // 2) We've filled in any partial cards at the front and back. | |
| 282 if (pre_top < _hard_end) { | |
| 283 // Now we can reset the _bt to do allocation in the given area. | |
| 284 MemRegion new_filler(pre_top, _hard_end); | |
| 285 fill_region_with_block(new_filler, false); | |
| 286 _top = pre_top + ParGCAllocBuffer::FillerHeaderSize; | |
| 287 // If there's no space left, don't retain. | |
| 288 if (_top >= _end) { | |
| 289 _retained = false; | |
| 290 invalidate(); | |
| 291 return; | |
| 292 } | |
| 293 _retained_filler = MemRegion(pre_top, _top); | |
| 294 _bt.set_region(MemRegion(_top, _hard_end)); | |
| 295 _bt.initialize_threshold(); | |
| 296 assert(_bt.threshold() > _top, "initialize_threshold failed!"); | |
| 297 | |
| 298 // There may be other reasons for queries into the middle of the | |
| 299 // filler object. When such queries are done in parallel with | |
| 300 // allocation, bad things can happen, if the query involves object | |
| 301 // iteration. So we ensure that such queries do not involve object | |
| 302 // iteration, by putting another filler object on the boundaries of | |
| 303 // such queries. One such is the object spanning a parallel card | |
| 304 // chunk boundary. | |
| 305 | |
| 306 // "chunk_boundary" is the address of the first chunk boundary less | |
| 307 // than "hard_end". | |
| 308 HeapWord* chunk_boundary = | |
| 309 (HeapWord*)align_size_down(intptr_t(_hard_end-1), ChunkSizeInBytes); | |
| 310 assert(chunk_boundary < _hard_end, "Or else above did not work."); | |
| 311 assert(pointer_delta(_true_end, chunk_boundary) >= AlignmentReserve, | |
| 312 "Consequence of last card handling above."); | |
| 313 | |
| 314 if (_top <= chunk_boundary) { | |
| 315 assert(_true_end == _hard_end, "Invariant."); | |
| 316 while (_top <= chunk_boundary) { | |
| 317 assert(pointer_delta(_hard_end, chunk_boundary) >= AlignmentReserve, | |
| 318 "Consequence of last card handling above."); | |
| 319 MemRegion chunk_portion(chunk_boundary, _hard_end); | |
| 320 _bt.BlockOffsetArray::alloc_block(chunk_portion.start(), | |
| 321 chunk_portion.end()); | |
| 322 SharedHeap::fill_region_with_object(chunk_portion); | |
| 323 _hard_end = chunk_portion.start(); | |
| 324 chunk_boundary -= ChunkSizeInWords; | |
| 325 } | |
| 326 _end = _hard_end - AlignmentReserve; | |
| 327 assert(_top <= _end, "Invariant."); | |
| 328 // Now reset the initial filler chunk so it doesn't overlap with | |
| 329 // the one(s) inserted above. | |
| 330 MemRegion new_filler(pre_top, _hard_end); | |
| 331 fill_region_with_block(new_filler, false); | |
| 332 } | |
| 333 } else { | |
| 334 _retained = false; | |
| 335 invalidate(); | |
| 336 } | |
| 337 } else { | |
| 338 assert(!end_of_gc || | |
| 339 (!_retained && _true_end == _hard_end), "Checking."); | |
| 340 } | |
| 341 assert(_end <= _hard_end, "Invariant."); | |
| 342 assert(_top < _end || _top == _hard_end, "Invariant"); | |
| 343 } |