< prev index next >

src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp

Print this page
rev 7471 : 8060025: Object copy time regressions after JDK-8031323 and JDK-8057536
Summary: Evaluate and improve object copy time by micro-optimizations and splitting out slow and fast paths aggressively.
Reviewed-by:
Contributed-by: Tony Printezis <tprintezis@twitter.com>, Thomas Schatzl <thomas.schatzl@oracle.com>
rev 7473 : imported patch mikael-refactor-cset-state


  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
  27 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
  28 #include "gc_implementation/g1/g1ParScanThreadState.inline.hpp"
  29 #include "oops/oop.inline.hpp"
  30 #include "oops/oop.pcgc.inline.hpp"
  31 #include "runtime/prefetch.inline.hpp"
  32 
  33 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num, ReferenceProcessor* rp)
  34   : _g1h(g1h),
  35     _refs(g1h->task_queue(queue_num)),
  36     _dcq(&g1h->dirty_card_queue_set()),
  37     _ct_bs(g1h->g1_barrier_set()),
  38     _g1_rem(g1h->g1_rem_set()),
  39     _hash_seed(17), _queue_num(queue_num),
  40     _term_attempts(0),

  41     _age_table(false), _scanner(g1h, rp),
  42     _strong_roots_time(0), _term_time(0) {
  43   _scanner.set_par_scan_thread_state(this);
  44   // we allocate G1YoungSurvRateNumRegions plus one entries, since
  45   // we "sacrifice" entry 0 to keep track of surviving bytes for
  46   // non-young regions (where the age is -1)
  47   // We also add a few elements at the beginning and at the end in
  48   // an attempt to eliminate cache contention
  49   uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
  50   uint array_length = PADDING_ELEM_NUM +
  51                       real_length +
  52                       PADDING_ELEM_NUM;
  53   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
  54   if (_surviving_young_words_base == NULL)
  55     vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR,
  56                           "Not enough space for young surv histo.");
  57   _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
  58   memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
  59 
  60   _g1_par_allocator = G1ParGCAllocator::create_allocator(_g1h);
  61 






  62   _start = os::elapsedTime();
  63 }
  64 
  65 G1ParScanThreadState::~G1ParScanThreadState() {
  66   _g1_par_allocator->retire_alloc_buffers();
  67   delete _g1_par_allocator;
  68   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base, mtGC);
  69 }
  70 
  71 void
  72 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
  73 {
  74   st->print_raw_cr("GC Termination Stats");
  75   st->print_raw_cr("     elapsed  --strong roots-- -------termination-------"
  76                    " ------waste (KiB)------");
  77   st->print_raw_cr("thr     ms        ms      %        ms      %    attempts"
  78                    "  total   alloc    undo");
  79   st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
  80                    " ------- ------- -------");
  81 }


 133   }
 134 }
 135 #endif // ASSERT
 136 
 137 void G1ParScanThreadState::trim_queue() {
 138   assert(_evac_failure_cl != NULL, "not set");
 139 
 140   StarTask ref;
 141   do {
 142     // Drain the overflow stack first, so other threads can steal.
 143     while (_refs->pop_overflow(ref)) {
 144       dispatch_reference(ref);
 145     }
 146 
 147     while (_refs->pop_local(ref)) {
 148       dispatch_reference(ref);
 149     }
 150   } while (!_refs->is_empty());
 151 }
 152 
 153 oop G1ParScanThreadState::copy_to_survivor_space(oop const old,








































 154                                                  markOop const old_mark) {
 155   size_t word_sz = old->size();
 156   HeapRegion* from_region = _g1h->heap_region_containing_raw(old);
 157   // +1 to make the -1 indexes valid...
 158   int       young_index = from_region->young_index_in_cset()+1;
 159   assert( (from_region->is_young() && young_index >  0) ||
 160          (!from_region->is_young() && young_index == 0), "invariant" );
 161   G1CollectorPolicy* g1p = _g1h->g1_policy();
 162   uint age = old_mark->has_displaced_mark_helper() ? old_mark->displaced_mark_helper()->age()
 163                                                    : old_mark->age();
 164   GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
 165                                                              word_sz);
 166   AllocationContext_t context = from_region->allocation_context();
 167   HeapWord* obj_ptr = _g1_par_allocator->allocate(alloc_purpose, word_sz, context);
 168 #ifndef PRODUCT
 169   // Should this evacuation fail?
 170   if (_g1h->evacuation_should_fail()) {
 171     if (obj_ptr != NULL) {
 172       _g1_par_allocator->undo_allocation(alloc_purpose, obj_ptr, word_sz, context);
 173       obj_ptr = NULL;
 174     }
 175   }
 176 #endif // !PRODUCT
 177 






 178   if (obj_ptr == NULL) {
 179     // This will either forward-to-self, or detect that someone else has
 180     // installed a forwarding pointer.
 181     return _g1h->handle_evacuation_failure_par(this, old);
 182   }


 183 
 184   oop obj = oop(obj_ptr);









 185 
 186   // We're going to allocate linearly, so might as well prefetch ahead.
 187   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
 188 
 189   oop forward_ptr = old->forward_to_atomic(obj);

 190   if (forward_ptr == NULL) {
 191     Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
 192 
 193     // alloc_purpose is just a hint to allocate() above, recheck the type of region
 194     // we actually allocated from and update alloc_purpose accordingly
 195     HeapRegion* to_region = _g1h->heap_region_containing_raw(obj_ptr);
 196     alloc_purpose = to_region->is_young() ? GCAllocForSurvived : GCAllocForTenured;
 197 
 198     if (g1p->track_object_age(alloc_purpose)) {
 199       if (age < markOopDesc::max_age) {
 200         age++;
 201       }
 202       if (old_mark->has_displaced_mark_helper()) {
 203         // In this case, we have to install the mark word first,
 204         // otherwise obj looks to be forwarded (the old mark word,
 205         // which contains the forward pointer, was copied)
 206         obj->set_mark(old_mark);
 207         markOop new_mark = old_mark->displaced_mark_helper()->set_age(age);
 208         old_mark->set_displaced_mark_helper(new_mark);
 209       } else {
 210         obj->set_mark(old_mark->set_age(age));
 211       }
 212       age_table()->add(age, word_sz);
 213     } else {
 214       obj->set_mark(old_mark);
 215     }
 216 
 217     if (G1StringDedup::is_enabled()) {
 218       G1StringDedup::enqueue_from_evacuation(from_region->is_young(),
 219                                              to_region->is_young(),






 220                                              queue_num(),
 221                                              obj);
 222     }
 223 
 224     size_t* surv_young_words = surviving_young_words();
 225     surv_young_words[young_index] += word_sz;
 226 
 227     if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
 228       // We keep track of the next start index in the length field of
 229       // the to-space object. The actual length can be found in the
 230       // length field of the from-space object.
 231       arrayOop(obj)->set_length(0);
 232       oop* old_p = set_partial_array_mask(old);
 233       push_on_queue(old_p);
 234     } else {
 235       // No point in using the slower heap_region_containing() method,
 236       // given that we know obj is in the heap.
 237       _scanner.set_region(_g1h->heap_region_containing_raw(obj));
 238       obj->oop_iterate_backwards(&_scanner);
 239     }

 240   } else {
 241     _g1_par_allocator->undo_allocation(alloc_purpose, obj_ptr, word_sz, context);
 242     obj = forward_ptr;
 243   }
 244   return obj;
 245 }


  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
  27 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
  28 #include "gc_implementation/g1/g1ParScanThreadState.inline.hpp"
  29 #include "oops/oop.inline.hpp"
  30 #include "oops/oop.pcgc.inline.hpp"
  31 #include "runtime/prefetch.inline.hpp"
  32 
  33 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num, ReferenceProcessor* rp)
  34   : _g1h(g1h),
  35     _refs(g1h->task_queue(queue_num)),
  36     _dcq(&g1h->dirty_card_queue_set()),
  37     _ct_bs(g1h->g1_barrier_set()),
  38     _g1_rem(g1h->g1_rem_set()),
  39     _hash_seed(17), _queue_num(queue_num),
  40     _term_attempts(0),
  41     _tenuring_threshold(g1h->g1_policy()->tenuring_threshold()),
  42     _age_table(false), _scanner(g1h, rp),
  43     _strong_roots_time(0), _term_time(0) {
  44   _scanner.set_par_scan_thread_state(this);
  45   // we allocate G1YoungSurvRateNumRegions plus one entries, since
  46   // we "sacrifice" entry 0 to keep track of surviving bytes for
  47   // non-young regions (where the age is -1)
  48   // We also add a few elements at the beginning and at the end in
  49   // an attempt to eliminate cache contention
  50   uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
  51   uint array_length = PADDING_ELEM_NUM +
  52                       real_length +
  53                       PADDING_ELEM_NUM;
  54   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
  55   if (_surviving_young_words_base == NULL)
  56     vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR,
  57                           "Not enough space for young surv histo.");
  58   _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
  59   memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
  60 
  61   _g1_par_allocator = G1ParGCAllocator::create_allocator(_g1h);
  62 
  63   _dest[InCSetState::NotInCSet]    = InCSetState::NotInCSet;
  64   // The dest for Young is used when the objects are aged enough to
  65   // need to be moved to the next space.
  66   _dest[InCSetState::Young]        = InCSetState::Old;
  67   _dest[InCSetState::Old]          = InCSetState::Old;
  68 
  69   _start = os::elapsedTime();
  70 }
  71 
  72 G1ParScanThreadState::~G1ParScanThreadState() {
  73   _g1_par_allocator->retire_alloc_buffers();
  74   delete _g1_par_allocator;
  75   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base, mtGC);
  76 }
  77 
  78 void
  79 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
  80 {
  81   st->print_raw_cr("GC Termination Stats");
  82   st->print_raw_cr("     elapsed  --strong roots-- -------termination-------"
  83                    " ------waste (KiB)------");
  84   st->print_raw_cr("thr     ms        ms      %        ms      %    attempts"
  85                    "  total   alloc    undo");
  86   st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
  87                    " ------- ------- -------");
  88 }


 140   }
 141 }
 142 #endif // ASSERT
 143 
 144 void G1ParScanThreadState::trim_queue() {
 145   assert(_evac_failure_cl != NULL, "not set");
 146 
 147   StarTask ref;
 148   do {
 149     // Drain the overflow stack first, so other threads can steal.
 150     while (_refs->pop_overflow(ref)) {
 151       dispatch_reference(ref);
 152     }
 153 
 154     while (_refs->pop_local(ref)) {
 155       dispatch_reference(ref);
 156     }
 157   } while (!_refs->is_empty());
 158 }
 159 
 160 HeapWord* G1ParScanThreadState::allocate_in_next_plab(InCSetState const state,
 161                                                       InCSetState* dest,
 162                                                       size_t word_sz,
 163                                                       AllocationContext_t const context) {
 164   assert(state.is_in_cset_or_humongous(), err_msg("Unexpected state: %u", state.value()));
 165   assert(dest->is_in_cset_or_humongous(), err_msg("Unexpected dest: %u", dest->value()));
 166 
 167   // Right now we only have two types of regions (young / old) so
 168   // let's keep the logic here simple. We can generalize it when necessary.
 169   if (dest->is_young()) {
 170     HeapWord* const obj_ptr = _g1_par_allocator->allocate(InCSetState::Old,
 171                                                           word_sz, context);
 172     if (obj_ptr == NULL) {
 173       return NULL;
 174     }
 175     // Make sure that we won't attempt to copy any other objects out
 176     // of a survivor region (given that apparently we cannot allocate
 177     // any new ones) to avoid coming into this slow path.
 178     _tenuring_threshold = 0;
 179     dest->set_old();
 180     return obj_ptr;
 181   } else {
 182     assert(dest->is_old(), err_msg("Unexpected dest: %u", dest->value()));
 183     // no other space to try.
 184     return NULL;
 185   }
 186 }
 187 
 188 InCSetState G1ParScanThreadState::next_state(InCSetState const state, markOop const m, uint& age) {
 189   if (state.is_young()) {
 190     age = !m->has_displaced_mark_helper() ? m->age()
 191                                           : m->displaced_mark_helper()->age();
 192     if (age < _tenuring_threshold) {
 193       return state;
 194     }
 195   }
 196   return dest(state);
 197 }
 198 
 199 oop G1ParScanThreadState::copy_to_survivor_space(InCSetState const state,
 200                                                  oop const old,
 201                                                  markOop const old_mark) {
 202   const size_t word_sz = old->size();
 203   HeapRegion* const from_region = _g1h->heap_region_containing_raw(old);
 204   // +1 to make the -1 indexes valid...
 205   const int young_index = from_region->young_index_in_cset()+1;
 206   assert( (from_region->is_young() && young_index >  0) ||
 207          (!from_region->is_young() && young_index == 0), "invariant" );
 208   const AllocationContext_t context = from_region->allocation_context();
 209 
 210   uint age = 0;
 211   InCSetState dest_state = next_state(state, old_mark, age);
 212   HeapWord* obj_ptr = _g1_par_allocator->plab_allocate(dest_state, word_sz, context);











 213 
 214   // PLAB allocations should succeed most of the time, so we'll
 215   // normally check against NULL once and that's it.
 216   if (obj_ptr == NULL) {
 217     obj_ptr = _g1_par_allocator->allocate_direct_or_new_plab(dest_state, word_sz, context);
 218     if (obj_ptr == NULL) {
 219       obj_ptr = allocate_in_next_plab(state, &dest_state, word_sz, context);
 220       if (obj_ptr == NULL) {
 221         // This will either forward-to-self, or detect that someone else has
 222         // installed a forwarding pointer.
 223         return _g1h->handle_evacuation_failure_par(this, old);
 224       }
 225     }
 226   }
 227 
 228   assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
 229 #ifndef PRODUCT
 230   // Should this evacuation fail?
 231   if (_g1h->evacuation_should_fail()) {
 232     // Doing this after all the allocation attempts also tests the
 233     // undo_allocation() method too.
 234     _g1_par_allocator->undo_allocation(dest_state, obj_ptr, word_sz, context);
 235     return _g1h->handle_evacuation_failure_par(this, old);
 236   }
 237 #endif // !PRODUCT
 238 
 239   // We're going to allocate linearly, so might as well prefetch ahead.
 240   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
 241 
 242   const oop obj = oop(obj_ptr);
 243   const oop forward_ptr = old->forward_to_atomic(obj);
 244   if (forward_ptr == NULL) {
 245     Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
 246 
 247     if (dest_state.is_young()) {





 248       if (age < markOopDesc::max_age) {
 249         age++;
 250       }
 251       if (old_mark->has_displaced_mark_helper()) {
 252         // In this case, we have to install the mark word first,
 253         // otherwise obj looks to be forwarded (the old mark word,
 254         // which contains the forward pointer, was copied)
 255         obj->set_mark(old_mark);
 256         markOop new_mark = old_mark->displaced_mark_helper()->set_age(age);
 257         old_mark->set_displaced_mark_helper(new_mark);
 258       } else {
 259         obj->set_mark(old_mark->set_age(age));
 260       }
 261       age_table()->add(age, word_sz);
 262     } else {
 263       obj->set_mark(old_mark);
 264     }
 265 
 266     if (G1StringDedup::is_enabled()) {
 267       const bool is_from_young = state.is_young();
 268       const bool is_to_young = dest_state.is_young();
 269       assert(is_from_young == _g1h->heap_region_containing_raw(old)->is_young(),
 270              "sanity");
 271       assert(is_to_young == _g1h->heap_region_containing_raw(obj)->is_young(),
 272              "sanity");
 273       G1StringDedup::enqueue_from_evacuation(is_from_young,
 274                                              is_to_young,
 275                                              queue_num(),
 276                                              obj);
 277     }
 278 
 279     size_t* const surv_young_words = surviving_young_words();
 280     surv_young_words[young_index] += word_sz;
 281 
 282     if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
 283       // We keep track of the next start index in the length field of
 284       // the to-space object. The actual length can be found in the
 285       // length field of the from-space object.
 286       arrayOop(obj)->set_length(0);
 287       oop* old_p = set_partial_array_mask(old);
 288       push_on_queue(old_p);
 289     } else {
 290       HeapRegion* const to_region = _g1h->heap_region_containing_raw(obj_ptr);
 291       _scanner.set_region(to_region);

 292       obj->oop_iterate_backwards(&_scanner);
 293     }
 294     return obj;
 295   } else {
 296     _g1_par_allocator->undo_allocation(dest_state, obj_ptr, word_sz, context);
 297     return forward_ptr;
 298   }

 299 }
< prev index next >