< 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


 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(in_cset_state_t const state,
 161                                                       in_cset_state_t* dest,
 162                                                       size_t word_sz,
 163                                                       AllocationContext_t const context) {
 164   assert(state != InCSetState::NotInCSet, err_msg("Unexpected state: %u", state));
 165   assert(*dest != InCSetState::NotInCSet, err_msg("Unexpected dest: %u", *dest));
 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 == InCSetState::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 = InCSetState::Old;
 180     return obj_ptr;
 181   } else {
 182     assert(*dest == InCSetState::Old, err_msg("Unexpected dest: %u", *dest));
 183     // no other space to try.
 184     return NULL;
 185   }
 186 }
 187 
 188 in_cset_state_t G1ParScanThreadState::next_state(in_cset_state_t const state, markOop const m, uint& age) {
 189   if (state == InCSetState::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(in_cset_state_t 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   in_cset_state_t 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 == InCSetState::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 == InCSetState::Young;
 268       const bool is_to_young = dest_state == InCSetState::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);


 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);
< prev index next >