duke@435: /* tamao@5161: * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: stefank@2314: #ifndef SHARE_VM_GC_IMPLEMENTATION_PARALLELSCAVENGE_PSPARALLELCOMPACT_HPP stefank@2314: #define SHARE_VM_GC_IMPLEMENTATION_PARALLELSCAVENGE_PSPARALLELCOMPACT_HPP stefank@2314: stefank@2314: #include "gc_implementation/parallelScavenge/objectStartArray.hpp" stefank@2314: #include "gc_implementation/parallelScavenge/parMarkBitMap.hpp" stefank@2314: #include "gc_implementation/parallelScavenge/psCompactionManager.hpp" stefank@2314: #include "gc_implementation/shared/collectorCounters.hpp" stefank@2314: #include "gc_implementation/shared/markSweep.hpp" stefank@2314: #include "gc_implementation/shared/mutableSpace.hpp" stefank@2314: #include "memory/sharedHeap.hpp" stefank@2314: #include "oops/oop.hpp" stefank@2314: duke@435: class ParallelScavengeHeap; duke@435: class PSAdaptiveSizePolicy; duke@435: class PSYoungGen; duke@435: class PSOldGen; duke@435: class ParCompactionManager; duke@435: class ParallelTaskTerminator; duke@435: class PSParallelCompact; duke@435: class GCTaskManager; duke@435: class GCTaskQueue; duke@435: class PreGCValues; duke@435: class MoveAndUpdateClosure; duke@435: class RefProcTaskExecutor; sla@5237: class ParallelOldTracer; sla@5237: class STWGCTimer; duke@435: jcoomes@917: // The SplitInfo class holds the information needed to 'split' a source region jcoomes@917: // so that the live data can be copied to two destination *spaces*. Normally, jcoomes@917: // all the live data in a region is copied to a single destination space (e.g., jcoomes@917: // everything live in a region in eden is copied entirely into the old gen). jcoomes@917: // However, when the heap is nearly full, all the live data in eden may not fit jcoomes@917: // into the old gen. Copying only some of the regions from eden to old gen jcoomes@917: // requires finding a region that does not contain a partial object (i.e., no jcoomes@917: // live object crosses the region boundary) somewhere near the last object that jcoomes@917: // does fit into the old gen. Since it's not always possible to find such a jcoomes@917: // region, splitting is necessary for predictable behavior. jcoomes@917: // jcoomes@917: // A region is always split at the end of the partial object. This avoids jcoomes@917: // additional tests when calculating the new location of a pointer, which is a jcoomes@917: // very hot code path. The partial object and everything to its left will be jcoomes@917: // copied to another space (call it dest_space_1). The live data to the right jcoomes@917: // of the partial object will be copied either within the space itself, or to a jcoomes@917: // different destination space (distinct from dest_space_1). jcoomes@917: // jcoomes@917: // Split points are identified during the summary phase, when region jcoomes@917: // destinations are computed: data about the split, including the jcoomes@917: // partial_object_size, is recorded in a SplitInfo record and the jcoomes@917: // partial_object_size field in the summary data is set to zero. The zeroing is jcoomes@917: // possible (and necessary) since the partial object will move to a different jcoomes@917: // destination space than anything to its right, thus the partial object should jcoomes@917: // not affect the locations of any objects to its right. jcoomes@917: // jcoomes@917: // The recorded data is used during the compaction phase, but only rarely: when jcoomes@917: // the partial object on the split region will be copied across a destination jcoomes@917: // region boundary. This test is made once each time a region is filled, and is jcoomes@917: // a simple address comparison, so the overhead is negligible (see jcoomes@917: // PSParallelCompact::first_src_addr()). jcoomes@917: // jcoomes@917: // Notes: jcoomes@917: // jcoomes@917: // Only regions with partial objects are split; a region without a partial jcoomes@917: // object does not need any extra bookkeeping. jcoomes@917: // jcoomes@917: // At most one region is split per space, so the amount of data required is jcoomes@917: // constant. jcoomes@917: // jcoomes@917: // A region is split only when the destination space would overflow. Once that jcoomes@917: // happens, the destination space is abandoned and no other data (even from jcoomes@917: // other source spaces) is targeted to that destination space. Abandoning the jcoomes@917: // destination space may leave a somewhat large unused area at the end, if a jcoomes@917: // large object caused the overflow. jcoomes@917: // jcoomes@917: // Future work: jcoomes@917: // jcoomes@917: // More bookkeeping would be required to continue to use the destination space. jcoomes@917: // The most general solution would allow data from regions in two different jcoomes@917: // source spaces to be "joined" in a single destination region. At the very jcoomes@917: // least, additional code would be required in next_src_region() to detect the jcoomes@917: // join and skip to an out-of-order source region. If the join region was also jcoomes@917: // the last destination region to which a split region was copied (the most jcoomes@917: // likely case), then additional work would be needed to get fill_region() to jcoomes@917: // stop iteration and switch to a new source region at the right point. Basic jcoomes@917: // idea would be to use a fake value for the top of the source space. It is jcoomes@917: // doable, if a bit tricky. jcoomes@917: // jcoomes@917: // A simpler (but less general) solution would fill the remainder of the jcoomes@917: // destination region with a dummy object and continue filling the next jcoomes@917: // destination region. jcoomes@917: jcoomes@917: class SplitInfo jcoomes@917: { jcoomes@917: public: jcoomes@917: // Return true if this split info is valid (i.e., if a split has been jcoomes@917: // recorded). The very first region cannot have a partial object and thus is jcoomes@917: // never split, so 0 is the 'invalid' value. jcoomes@917: bool is_valid() const { return _src_region_idx > 0; } jcoomes@917: jcoomes@917: // Return true if this split holds data for the specified source region. jcoomes@917: inline bool is_split(size_t source_region) const; jcoomes@917: jcoomes@917: // The index of the split region, the size of the partial object on that jcoomes@917: // region and the destination of the partial object. jcoomes@917: size_t src_region_idx() const { return _src_region_idx; } jcoomes@917: size_t partial_obj_size() const { return _partial_obj_size; } jcoomes@917: HeapWord* destination() const { return _destination; } jcoomes@917: jcoomes@917: // The destination count of the partial object referenced by this split jcoomes@917: // (either 1 or 2). This must be added to the destination count of the jcoomes@917: // remainder of the source region. jcoomes@917: unsigned int destination_count() const { return _destination_count; } jcoomes@917: jcoomes@917: // If a word within the partial object will be written to the first word of a jcoomes@917: // destination region, this is the address of the destination region; jcoomes@917: // otherwise this is NULL. jcoomes@917: HeapWord* dest_region_addr() const { return _dest_region_addr; } jcoomes@917: jcoomes@917: // If a word within the partial object will be written to the first word of a jcoomes@917: // destination region, this is the address of that word within the partial jcoomes@917: // object; otherwise this is NULL. jcoomes@917: HeapWord* first_src_addr() const { return _first_src_addr; } jcoomes@917: jcoomes@917: // Record the data necessary to split the region src_region_idx. jcoomes@917: void record(size_t src_region_idx, size_t partial_obj_size, jcoomes@917: HeapWord* destination); jcoomes@917: jcoomes@917: void clear(); jcoomes@917: jcoomes@917: DEBUG_ONLY(void verify_clear();) jcoomes@917: jcoomes@917: private: jcoomes@917: size_t _src_region_idx; jcoomes@917: size_t _partial_obj_size; jcoomes@917: HeapWord* _destination; jcoomes@917: unsigned int _destination_count; jcoomes@917: HeapWord* _dest_region_addr; jcoomes@917: HeapWord* _first_src_addr; jcoomes@917: }; jcoomes@917: jcoomes@917: inline bool SplitInfo::is_split(size_t region_idx) const jcoomes@917: { jcoomes@917: return _src_region_idx == region_idx && is_valid(); jcoomes@917: } jcoomes@917: duke@435: class SpaceInfo duke@435: { duke@435: public: duke@435: MutableSpace* space() const { return _space; } duke@435: duke@435: // Where the free space will start after the collection. Valid only after the duke@435: // summary phase completes. duke@435: HeapWord* new_top() const { return _new_top; } duke@435: duke@435: // Allows new_top to be set. duke@435: HeapWord** new_top_addr() { return &_new_top; } duke@435: duke@435: // Where the smallest allowable dense prefix ends (used only for perm gen). duke@435: HeapWord* min_dense_prefix() const { return _min_dense_prefix; } duke@435: duke@435: // Where the dense prefix ends, or the compacted region begins. duke@435: HeapWord* dense_prefix() const { return _dense_prefix; } duke@435: duke@435: // The start array for the (generation containing the) space, or NULL if there duke@435: // is no start array. duke@435: ObjectStartArray* start_array() const { return _start_array; } duke@435: jcoomes@917: SplitInfo& split_info() { return _split_info; } jcoomes@917: duke@435: void set_space(MutableSpace* s) { _space = s; } duke@435: void set_new_top(HeapWord* addr) { _new_top = addr; } duke@435: void set_min_dense_prefix(HeapWord* addr) { _min_dense_prefix = addr; } duke@435: void set_dense_prefix(HeapWord* addr) { _dense_prefix = addr; } duke@435: void set_start_array(ObjectStartArray* s) { _start_array = s; } duke@435: jcoomes@917: void publish_new_top() const { _space->set_top(_new_top); } jcoomes@917: duke@435: private: duke@435: MutableSpace* _space; duke@435: HeapWord* _new_top; duke@435: HeapWord* _min_dense_prefix; duke@435: HeapWord* _dense_prefix; duke@435: ObjectStartArray* _start_array; jcoomes@917: SplitInfo _split_info; duke@435: }; duke@435: duke@435: class ParallelCompactData duke@435: { duke@435: public: duke@435: // Sizes are in HeapWords, unless indicated otherwise. jcoomes@810: static const size_t Log2RegionSize; jcoomes@810: static const size_t RegionSize; jcoomes@810: static const size_t RegionSizeBytes; duke@435: jcoomes@810: // Mask for the bits in a size_t to get an offset within a region. jcoomes@810: static const size_t RegionSizeOffsetMask; jcoomes@810: // Mask for the bits in a pointer to get an offset within a region. jcoomes@810: static const size_t RegionAddrOffsetMask; jcoomes@810: // Mask for the bits in a pointer to get the address of the start of a region. jcoomes@810: static const size_t RegionAddrMask; duke@435: jcoomes@5201: static const size_t Log2BlockSize; jcoomes@5201: static const size_t BlockSize; jcoomes@5201: static const size_t BlockSizeBytes; jcoomes@5201: jcoomes@5201: static const size_t BlockSizeOffsetMask; jcoomes@5201: static const size_t BlockAddrOffsetMask; jcoomes@5201: static const size_t BlockAddrMask; jcoomes@5201: jcoomes@5201: static const size_t BlocksPerRegion; jcoomes@5201: static const size_t Log2BlocksPerRegion; jcoomes@5201: jcoomes@810: class RegionData duke@435: { duke@435: public: jcoomes@810: // Destination address of the region. duke@435: HeapWord* destination() const { return _destination; } duke@435: jcoomes@810: // The first region containing data destined for this region. jcoomes@810: size_t source_region() const { return _source_region; } duke@435: jcoomes@810: // The object (if any) starting in this region and ending in a different jcoomes@810: // region that could not be updated during the main (parallel) compaction duke@435: // phase. This is different from _partial_obj_addr, which is an object that jcoomes@810: // extends onto a source region. However, the two uses do not overlap in duke@435: // time, so the same field is used to save space. duke@435: HeapWord* deferred_obj_addr() const { return _partial_obj_addr; } duke@435: jcoomes@810: // The starting address of the partial object extending onto the region. duke@435: HeapWord* partial_obj_addr() const { return _partial_obj_addr; } duke@435: jcoomes@810: // Size of the partial object extending onto the region (words). duke@435: size_t partial_obj_size() const { return _partial_obj_size; } duke@435: jcoomes@810: // Size of live data that lies within this region due to objects that start jcoomes@810: // in this region (words). This does not include the partial object jcoomes@810: // extending onto the region (if any), or the part of an object that extends jcoomes@810: // onto the next region (if any). duke@435: size_t live_obj_size() const { return _dc_and_los & los_mask; } duke@435: jcoomes@810: // Total live data that lies within the region (words). duke@435: size_t data_size() const { return partial_obj_size() + live_obj_size(); } duke@435: jcoomes@810: // The destination_count is the number of other regions to which data from jcoomes@810: // this region will be copied. At the end of the summary phase, the valid duke@435: // values of destination_count are duke@435: // jcoomes@810: // 0 - data from the region will be compacted completely into itself, or the jcoomes@810: // region is empty. The region can be claimed and then filled. jcoomes@810: // 1 - data from the region will be compacted into 1 other region; some jcoomes@810: // data from the region may also be compacted into the region itself. jcoomes@810: // 2 - data from the region will be copied to 2 other regions. duke@435: // jcoomes@810: // During compaction as regions are emptied, the destination_count is duke@435: // decremented (atomically) and when it reaches 0, it can be claimed and duke@435: // then filled. duke@435: // jcoomes@810: // A region is claimed for processing by atomically changing the jcoomes@810: // destination_count to the claimed value (dc_claimed). After a region has duke@435: // been filled, the destination_count should be set to the completed value duke@435: // (dc_completed). duke@435: inline uint destination_count() const; duke@435: inline uint destination_count_raw() const; duke@435: jcoomes@5201: // Whether the block table for this region has been filled. jcoomes@5201: inline bool blocks_filled() const; jcoomes@5201: jcoomes@5201: // Number of times the block table was filled. jcoomes@5201: DEBUG_ONLY(inline size_t blocks_filled_count() const;) jcoomes@5201: jcoomes@810: // The location of the java heap data that corresponds to this region. duke@435: inline HeapWord* data_location() const; duke@435: jcoomes@810: // The highest address referenced by objects in this region. duke@435: inline HeapWord* highest_ref() const; duke@435: jcoomes@810: // Whether this region is available to be claimed, has been claimed, or has duke@435: // been completed. duke@435: // jcoomes@810: // Minor subtlety: claimed() returns true if the region is marked jcoomes@810: // completed(), which is desirable since a region must be claimed before it duke@435: // can be completed. duke@435: bool available() const { return _dc_and_los < dc_one; } duke@435: bool claimed() const { return _dc_and_los >= dc_claimed; } duke@435: bool completed() const { return _dc_and_los >= dc_completed; } duke@435: duke@435: // These are not atomic. duke@435: void set_destination(HeapWord* addr) { _destination = addr; } jcoomes@810: void set_source_region(size_t region) { _source_region = region; } duke@435: void set_deferred_obj_addr(HeapWord* addr) { _partial_obj_addr = addr; } duke@435: void set_partial_obj_addr(HeapWord* addr) { _partial_obj_addr = addr; } duke@435: void set_partial_obj_size(size_t words) { jcoomes@810: _partial_obj_size = (region_sz_t) words; duke@435: } jcoomes@5201: inline void set_blocks_filled(); duke@435: duke@435: inline void set_destination_count(uint count); duke@435: inline void set_live_obj_size(size_t words); duke@435: inline void set_data_location(HeapWord* addr); duke@435: inline void set_completed(); duke@435: inline bool claim_unsafe(); duke@435: duke@435: // These are atomic. duke@435: inline void add_live_obj(size_t words); duke@435: inline void set_highest_ref(HeapWord* addr); duke@435: inline void decrement_destination_count(); duke@435: inline bool claim(); duke@435: duke@435: private: jcoomes@810: // The type used to represent object sizes within a region. jcoomes@810: typedef uint region_sz_t; duke@435: duke@435: // Constants for manipulating the _dc_and_los field, which holds both the duke@435: // destination count and live obj size. The live obj size lives at the duke@435: // least significant end so no masking is necessary when adding. jcoomes@810: static const region_sz_t dc_shift; // Shift amount. jcoomes@810: static const region_sz_t dc_mask; // Mask for destination count. jcoomes@810: static const region_sz_t dc_one; // 1, shifted appropriately. jcoomes@810: static const region_sz_t dc_claimed; // Region has been claimed. jcoomes@810: static const region_sz_t dc_completed; // Region has been completed. jcoomes@810: static const region_sz_t los_mask; // Mask for live obj size. duke@435: jcoomes@810: HeapWord* _destination; jcoomes@810: size_t _source_region; jcoomes@810: HeapWord* _partial_obj_addr; jcoomes@810: region_sz_t _partial_obj_size; jcoomes@810: region_sz_t volatile _dc_and_los; jcoomes@5201: bool _blocks_filled; jcoomes@5201: duke@435: #ifdef ASSERT jcoomes@5201: size_t _blocks_filled_count; // Number of block table fills. jcoomes@5201: duke@435: // These enable optimizations that are only partially implemented. Use duke@435: // debug builds to prevent the code fragments from breaking. jcoomes@810: HeapWord* _data_location; jcoomes@810: HeapWord* _highest_ref; duke@435: #endif // #ifdef ASSERT duke@435: duke@435: #ifdef ASSERT duke@435: public: jcoomes@5201: uint _pushed; // 0 until region is pushed onto a stack duke@435: private: duke@435: #endif duke@435: }; duke@435: jcoomes@5201: // "Blocks" allow shorter sections of the bitmap to be searched. Each Block jcoomes@5201: // holds an offset, which is the amount of live data in the Region to the left jcoomes@5201: // of the first live object that starts in the Block. jcoomes@5201: class BlockData jcoomes@5201: { jcoomes@5201: public: jcoomes@5201: typedef unsigned short int blk_ofs_t; jcoomes@5201: jcoomes@5201: blk_ofs_t offset() const { return _offset; } jcoomes@5201: void set_offset(size_t val) { _offset = (blk_ofs_t)val; } jcoomes@5201: jcoomes@5201: private: jcoomes@5201: blk_ofs_t _offset; jcoomes@5201: }; jcoomes@5201: duke@435: public: duke@435: ParallelCompactData(); duke@435: bool initialize(MemRegion covered_region); duke@435: jcoomes@810: size_t region_count() const { return _region_count; } tamao@5161: size_t reserved_byte_size() const { return _reserved_byte_size; } duke@435: jcoomes@810: // Convert region indices to/from RegionData pointers. jcoomes@810: inline RegionData* region(size_t region_idx) const; jcoomes@810: inline size_t region(const RegionData* const region_ptr) const; duke@435: jcoomes@5201: size_t block_count() const { return _block_count; } jcoomes@5201: inline BlockData* block(size_t block_idx) const; jcoomes@5201: inline size_t block(const BlockData* block_ptr) const; duke@435: duke@435: void add_obj(HeapWord* addr, size_t len); duke@435: void add_obj(oop p, size_t len) { add_obj((HeapWord*)p, len); } duke@435: jcoomes@810: // Fill in the regions covering [beg, end) so that no data moves; i.e., the jcoomes@810: // destination of region n is simply the start of region n. The argument beg jcoomes@810: // must be region-aligned; end need not be. duke@435: void summarize_dense_prefix(HeapWord* beg, HeapWord* end); duke@435: jcoomes@917: HeapWord* summarize_split_space(size_t src_region, SplitInfo& split_info, jcoomes@917: HeapWord* destination, HeapWord* target_end, jcoomes@917: HeapWord** target_next); jcoomes@917: bool summarize(SplitInfo& split_info, duke@435: HeapWord* source_beg, HeapWord* source_end, jcoomes@917: HeapWord** source_next, jcoomes@917: HeapWord* target_beg, HeapWord* target_end, jcoomes@917: HeapWord** target_next); duke@435: duke@435: void clear(); jcoomes@810: void clear_range(size_t beg_region, size_t end_region); duke@435: void clear_range(HeapWord* beg, HeapWord* end) { jcoomes@810: clear_range(addr_to_region_idx(beg), addr_to_region_idx(end)); duke@435: } duke@435: jcoomes@810: // Return the number of words between addr and the start of the region duke@435: // containing addr. jcoomes@810: inline size_t region_offset(const HeapWord* addr) const; duke@435: jcoomes@810: // Convert addresses to/from a region index or region pointer. jcoomes@810: inline size_t addr_to_region_idx(const HeapWord* addr) const; jcoomes@810: inline RegionData* addr_to_region_ptr(const HeapWord* addr) const; jcoomes@810: inline HeapWord* region_to_addr(size_t region) const; jcoomes@810: inline HeapWord* region_to_addr(size_t region, size_t offset) const; jcoomes@810: inline HeapWord* region_to_addr(const RegionData* region) const; duke@435: jcoomes@810: inline HeapWord* region_align_down(HeapWord* addr) const; jcoomes@810: inline HeapWord* region_align_up(HeapWord* addr) const; jcoomes@810: inline bool is_region_aligned(HeapWord* addr) const; duke@435: jcoomes@5201: // Analogous to region_offset() for blocks. jcoomes@5201: size_t block_offset(const HeapWord* addr) const; jcoomes@5201: size_t addr_to_block_idx(const HeapWord* addr) const; jcoomes@5201: size_t addr_to_block_idx(const oop obj) const { jcoomes@5201: return addr_to_block_idx((HeapWord*) obj); jcoomes@5201: } jcoomes@5201: inline BlockData* addr_to_block_ptr(const HeapWord* addr) const; jcoomes@5201: inline HeapWord* block_to_addr(size_t block) const; jcoomes@5201: inline size_t region_to_block_idx(size_t region) const; jcoomes@5201: jcoomes@5201: inline HeapWord* block_align_down(HeapWord* addr) const; jcoomes@5201: inline HeapWord* block_align_up(HeapWord* addr) const; jcoomes@5201: inline bool is_block_aligned(HeapWord* addr) const; jcoomes@5201: duke@435: // Return the address one past the end of the partial object. jcoomes@810: HeapWord* partial_obj_end(size_t region_idx) const; duke@435: jcoomes@5201: // Return the location of the object after compaction. duke@435: HeapWord* calc_new_pointer(HeapWord* addr); duke@435: duke@435: HeapWord* calc_new_pointer(oop p) { duke@435: return calc_new_pointer((HeapWord*) p); duke@435: } duke@435: duke@435: #ifdef ASSERT duke@435: void verify_clear(const PSVirtualSpace* vspace); duke@435: void verify_clear(); duke@435: #endif // #ifdef ASSERT duke@435: duke@435: private: jcoomes@5201: bool initialize_block_data(); jcoomes@810: bool initialize_region_data(size_t region_size); duke@435: PSVirtualSpace* create_vspace(size_t count, size_t element_size); duke@435: duke@435: private: duke@435: HeapWord* _region_start; duke@435: #ifdef ASSERT duke@435: HeapWord* _region_end; duke@435: #endif // #ifdef ASSERT duke@435: jcoomes@810: PSVirtualSpace* _region_vspace; tamao@5161: size_t _reserved_byte_size; jcoomes@810: RegionData* _region_data; jcoomes@810: size_t _region_count; jcoomes@5201: jcoomes@5201: PSVirtualSpace* _block_vspace; jcoomes@5201: BlockData* _block_data; jcoomes@5201: size_t _block_count; duke@435: }; duke@435: duke@435: inline uint jcoomes@810: ParallelCompactData::RegionData::destination_count_raw() const duke@435: { duke@435: return _dc_and_los & dc_mask; duke@435: } duke@435: duke@435: inline uint jcoomes@810: ParallelCompactData::RegionData::destination_count() const duke@435: { duke@435: return destination_count_raw() >> dc_shift; duke@435: } duke@435: jcoomes@5201: inline bool jcoomes@5201: ParallelCompactData::RegionData::blocks_filled() const jcoomes@5201: { jcoomes@5201: return _blocks_filled; jcoomes@5201: } jcoomes@5201: jcoomes@5201: #ifdef ASSERT jcoomes@5201: inline size_t jcoomes@5201: ParallelCompactData::RegionData::blocks_filled_count() const jcoomes@5201: { jcoomes@5201: return _blocks_filled_count; jcoomes@5201: } jcoomes@5201: #endif // #ifdef ASSERT jcoomes@5201: jcoomes@5201: inline void jcoomes@5201: ParallelCompactData::RegionData::set_blocks_filled() jcoomes@5201: { jcoomes@5201: _blocks_filled = true; jcoomes@5201: // Debug builds count the number of times the table was filled. jcoomes@5201: DEBUG_ONLY(Atomic::inc_ptr(&_blocks_filled_count)); jcoomes@5201: } jcoomes@5201: duke@435: inline void jcoomes@810: ParallelCompactData::RegionData::set_destination_count(uint count) duke@435: { duke@435: assert(count <= (dc_completed >> dc_shift), "count too large"); jcoomes@810: const region_sz_t live_sz = (region_sz_t) live_obj_size(); duke@435: _dc_and_los = (count << dc_shift) | live_sz; duke@435: } duke@435: jcoomes@810: inline void ParallelCompactData::RegionData::set_live_obj_size(size_t words) duke@435: { duke@435: assert(words <= los_mask, "would overflow"); jcoomes@810: _dc_and_los = destination_count_raw() | (region_sz_t)words; duke@435: } duke@435: jcoomes@810: inline void ParallelCompactData::RegionData::decrement_destination_count() duke@435: { duke@435: assert(_dc_and_los < dc_claimed, "already claimed"); duke@435: assert(_dc_and_los >= dc_one, "count would go negative"); duke@435: Atomic::add((int)dc_mask, (volatile int*)&_dc_and_los); duke@435: } duke@435: jcoomes@810: inline HeapWord* ParallelCompactData::RegionData::data_location() const duke@435: { duke@435: DEBUG_ONLY(return _data_location;) duke@435: NOT_DEBUG(return NULL;) duke@435: } duke@435: jcoomes@810: inline HeapWord* ParallelCompactData::RegionData::highest_ref() const duke@435: { duke@435: DEBUG_ONLY(return _highest_ref;) duke@435: NOT_DEBUG(return NULL;) duke@435: } duke@435: jcoomes@810: inline void ParallelCompactData::RegionData::set_data_location(HeapWord* addr) duke@435: { duke@435: DEBUG_ONLY(_data_location = addr;) duke@435: } duke@435: jcoomes@810: inline void ParallelCompactData::RegionData::set_completed() duke@435: { duke@435: assert(claimed(), "must be claimed first"); jcoomes@810: _dc_and_los = dc_completed | (region_sz_t) live_obj_size(); duke@435: } duke@435: jcoomes@810: // MT-unsafe claiming of a region. Should only be used during single threaded duke@435: // execution. jcoomes@810: inline bool ParallelCompactData::RegionData::claim_unsafe() duke@435: { duke@435: if (available()) { duke@435: _dc_and_los |= dc_claimed; duke@435: return true; duke@435: } duke@435: return false; duke@435: } duke@435: jcoomes@810: inline void ParallelCompactData::RegionData::add_live_obj(size_t words) duke@435: { duke@435: assert(words <= (size_t)los_mask - live_obj_size(), "overflow"); duke@435: Atomic::add((int) words, (volatile int*) &_dc_and_los); duke@435: } duke@435: jcoomes@810: inline void ParallelCompactData::RegionData::set_highest_ref(HeapWord* addr) duke@435: { duke@435: #ifdef ASSERT duke@435: HeapWord* tmp = _highest_ref; duke@435: while (addr > tmp) { duke@435: tmp = (HeapWord*)Atomic::cmpxchg_ptr(addr, &_highest_ref, tmp); duke@435: } duke@435: #endif // #ifdef ASSERT duke@435: } duke@435: jcoomes@810: inline bool ParallelCompactData::RegionData::claim() duke@435: { duke@435: const int los = (int) live_obj_size(); duke@435: const int old = Atomic::cmpxchg(dc_claimed | los, duke@435: (volatile int*) &_dc_and_los, los); duke@435: return old == los; duke@435: } duke@435: jcoomes@810: inline ParallelCompactData::RegionData* jcoomes@810: ParallelCompactData::region(size_t region_idx) const duke@435: { jcoomes@810: assert(region_idx <= region_count(), "bad arg"); jcoomes@810: return _region_data + region_idx; duke@435: } duke@435: duke@435: inline size_t jcoomes@810: ParallelCompactData::region(const RegionData* const region_ptr) const duke@435: { jcoomes@810: assert(region_ptr >= _region_data, "bad arg"); jcoomes@810: assert(region_ptr <= _region_data + region_count(), "bad arg"); jcoomes@810: return pointer_delta(region_ptr, _region_data, sizeof(RegionData)); duke@435: } duke@435: jcoomes@5201: inline ParallelCompactData::BlockData* jcoomes@5201: ParallelCompactData::block(size_t n) const { jcoomes@5201: assert(n < block_count(), "bad arg"); jcoomes@5201: return _block_data + n; jcoomes@5201: } jcoomes@5201: duke@435: inline size_t jcoomes@810: ParallelCompactData::region_offset(const HeapWord* addr) const duke@435: { duke@435: assert(addr >= _region_start, "bad addr"); duke@435: assert(addr <= _region_end, "bad addr"); jcoomes@810: return (size_t(addr) & RegionAddrOffsetMask) >> LogHeapWordSize; duke@435: } duke@435: duke@435: inline size_t jcoomes@810: ParallelCompactData::addr_to_region_idx(const HeapWord* addr) const duke@435: { duke@435: assert(addr >= _region_start, "bad addr"); duke@435: assert(addr <= _region_end, "bad addr"); jcoomes@810: return pointer_delta(addr, _region_start) >> Log2RegionSize; duke@435: } duke@435: jcoomes@810: inline ParallelCompactData::RegionData* jcoomes@810: ParallelCompactData::addr_to_region_ptr(const HeapWord* addr) const duke@435: { jcoomes@810: return region(addr_to_region_idx(addr)); duke@435: } duke@435: duke@435: inline HeapWord* jcoomes@810: ParallelCompactData::region_to_addr(size_t region) const duke@435: { jcoomes@810: assert(region <= _region_count, "region out of range"); jcoomes@810: return _region_start + (region << Log2RegionSize); duke@435: } duke@435: duke@435: inline HeapWord* jcoomes@810: ParallelCompactData::region_to_addr(const RegionData* region) const duke@435: { jcoomes@810: return region_to_addr(pointer_delta(region, _region_data, jcoomes@810: sizeof(RegionData))); duke@435: } duke@435: duke@435: inline HeapWord* jcoomes@810: ParallelCompactData::region_to_addr(size_t region, size_t offset) const duke@435: { jcoomes@810: assert(region <= _region_count, "region out of range"); jcoomes@810: assert(offset < RegionSize, "offset too big"); // This may be too strict. jcoomes@810: return region_to_addr(region) + offset; duke@435: } duke@435: duke@435: inline HeapWord* jcoomes@810: ParallelCompactData::region_align_down(HeapWord* addr) const duke@435: { duke@435: assert(addr >= _region_start, "bad addr"); jcoomes@810: assert(addr < _region_end + RegionSize, "bad addr"); jcoomes@810: return (HeapWord*)(size_t(addr) & RegionAddrMask); duke@435: } duke@435: duke@435: inline HeapWord* jcoomes@810: ParallelCompactData::region_align_up(HeapWord* addr) const duke@435: { duke@435: assert(addr >= _region_start, "bad addr"); duke@435: assert(addr <= _region_end, "bad addr"); jcoomes@810: return region_align_down(addr + RegionSizeOffsetMask); duke@435: } duke@435: duke@435: inline bool jcoomes@810: ParallelCompactData::is_region_aligned(HeapWord* addr) const duke@435: { jcoomes@810: return region_offset(addr) == 0; duke@435: } duke@435: jcoomes@5201: inline size_t jcoomes@5201: ParallelCompactData::block_offset(const HeapWord* addr) const jcoomes@5201: { jcoomes@5201: assert(addr >= _region_start, "bad addr"); jcoomes@5201: assert(addr <= _region_end, "bad addr"); jcoomes@5201: return (size_t(addr) & BlockAddrOffsetMask) >> LogHeapWordSize; jcoomes@5201: } jcoomes@5201: jcoomes@5201: inline size_t jcoomes@5201: ParallelCompactData::addr_to_block_idx(const HeapWord* addr) const jcoomes@5201: { jcoomes@5201: assert(addr >= _region_start, "bad addr"); jcoomes@5201: assert(addr <= _region_end, "bad addr"); jcoomes@5201: return pointer_delta(addr, _region_start) >> Log2BlockSize; jcoomes@5201: } jcoomes@5201: jcoomes@5201: inline ParallelCompactData::BlockData* jcoomes@5201: ParallelCompactData::addr_to_block_ptr(const HeapWord* addr) const jcoomes@5201: { jcoomes@5201: return block(addr_to_block_idx(addr)); jcoomes@5201: } jcoomes@5201: jcoomes@5201: inline HeapWord* jcoomes@5201: ParallelCompactData::block_to_addr(size_t block) const jcoomes@5201: { jcoomes@5201: assert(block < _block_count, "block out of range"); jcoomes@5201: return _region_start + (block << Log2BlockSize); jcoomes@5201: } jcoomes@5201: jcoomes@5201: inline size_t jcoomes@5201: ParallelCompactData::region_to_block_idx(size_t region) const jcoomes@5201: { jcoomes@5201: return region << Log2BlocksPerRegion; jcoomes@5201: } jcoomes@5201: jcoomes@5201: inline HeapWord* jcoomes@5201: ParallelCompactData::block_align_down(HeapWord* addr) const jcoomes@5201: { jcoomes@5201: assert(addr >= _region_start, "bad addr"); jcoomes@5201: assert(addr < _region_end + RegionSize, "bad addr"); jcoomes@5201: return (HeapWord*)(size_t(addr) & BlockAddrMask); jcoomes@5201: } jcoomes@5201: jcoomes@5201: inline HeapWord* jcoomes@5201: ParallelCompactData::block_align_up(HeapWord* addr) const jcoomes@5201: { jcoomes@5201: assert(addr >= _region_start, "bad addr"); jcoomes@5201: assert(addr <= _region_end, "bad addr"); jcoomes@5201: return block_align_down(addr + BlockSizeOffsetMask); jcoomes@5201: } jcoomes@5201: jcoomes@5201: inline bool jcoomes@5201: ParallelCompactData::is_block_aligned(HeapWord* addr) const jcoomes@5201: { jcoomes@5201: return block_offset(addr) == 0; jcoomes@5201: } jcoomes@5201: duke@435: // Abstract closure for use with ParMarkBitMap::iterate(), which will invoke the duke@435: // do_addr() method. duke@435: // duke@435: // The closure is initialized with the number of heap words to process duke@435: // (words_remaining()), and becomes 'full' when it reaches 0. The do_addr() duke@435: // methods in subclasses should update the total as words are processed. Since duke@435: // only one subclass actually uses this mechanism to terminate iteration, the duke@435: // default initial value is > 0. The implementation is here and not in the duke@435: // single subclass that uses it to avoid making is_full() virtual, and thus duke@435: // adding a virtual call per live object. duke@435: duke@435: class ParMarkBitMapClosure: public StackObj { duke@435: public: duke@435: typedef ParMarkBitMap::idx_t idx_t; duke@435: typedef ParMarkBitMap::IterationStatus IterationStatus; duke@435: duke@435: public: duke@435: inline ParMarkBitMapClosure(ParMarkBitMap* mbm, ParCompactionManager* cm, duke@435: size_t words = max_uintx); duke@435: duke@435: inline ParCompactionManager* compaction_manager() const; duke@435: inline ParMarkBitMap* bitmap() const; duke@435: inline size_t words_remaining() const; duke@435: inline bool is_full() const; duke@435: inline HeapWord* source() const; duke@435: duke@435: inline void set_source(HeapWord* addr); duke@435: duke@435: virtual IterationStatus do_addr(HeapWord* addr, size_t words) = 0; duke@435: duke@435: protected: duke@435: inline void decrement_words_remaining(size_t words); duke@435: duke@435: private: duke@435: ParMarkBitMap* const _bitmap; duke@435: ParCompactionManager* const _compaction_manager; duke@435: DEBUG_ONLY(const size_t _initial_words_remaining;) // Useful in debugger. duke@435: size_t _words_remaining; // Words left to copy. duke@435: duke@435: protected: duke@435: HeapWord* _source; // Next addr that would be read. duke@435: }; duke@435: duke@435: inline duke@435: ParMarkBitMapClosure::ParMarkBitMapClosure(ParMarkBitMap* bitmap, duke@435: ParCompactionManager* cm, duke@435: size_t words): duke@435: _bitmap(bitmap), _compaction_manager(cm) duke@435: #ifdef ASSERT duke@435: , _initial_words_remaining(words) duke@435: #endif duke@435: { duke@435: _words_remaining = words; duke@435: _source = NULL; duke@435: } duke@435: duke@435: inline ParCompactionManager* ParMarkBitMapClosure::compaction_manager() const { duke@435: return _compaction_manager; duke@435: } duke@435: duke@435: inline ParMarkBitMap* ParMarkBitMapClosure::bitmap() const { duke@435: return _bitmap; duke@435: } duke@435: duke@435: inline size_t ParMarkBitMapClosure::words_remaining() const { duke@435: return _words_remaining; duke@435: } duke@435: duke@435: inline bool ParMarkBitMapClosure::is_full() const { duke@435: return words_remaining() == 0; duke@435: } duke@435: duke@435: inline HeapWord* ParMarkBitMapClosure::source() const { duke@435: return _source; duke@435: } duke@435: duke@435: inline void ParMarkBitMapClosure::set_source(HeapWord* addr) { duke@435: _source = addr; duke@435: } duke@435: duke@435: inline void ParMarkBitMapClosure::decrement_words_remaining(size_t words) { duke@435: assert(_words_remaining >= words, "processed too many words"); duke@435: _words_remaining -= words; duke@435: } duke@435: jcoomes@810: // The UseParallelOldGC collector is a stop-the-world garbage collector that jcoomes@810: // does parts of the collection using parallel threads. The collection includes jcoomes@810: // the tenured generation and the young generation. The permanent generation is jcoomes@810: // collected at the same time as the other two generations but the permanent jcoomes@810: // generation is collect by a single GC thread. The permanent generation is jcoomes@810: // collected serially because of the requirement that during the processing of a jcoomes@810: // klass AAA, any objects reference by AAA must already have been processed. jcoomes@810: // This requirement is enforced by a left (lower address) to right (higher jcoomes@810: // address) sliding compaction. jmasa@698: // jmasa@698: // There are four phases of the collection. jmasa@698: // jmasa@698: // - marking phase jmasa@698: // - summary phase jmasa@698: // - compacting phase jmasa@698: // - clean up phase jmasa@698: // jmasa@698: // Roughly speaking these phases correspond, respectively, to jmasa@698: // - mark all the live objects jmasa@698: // - calculate the destination of each object at the end of the collection jmasa@698: // - move the objects to their destination jmasa@698: // - update some references and reinitialize some variables jmasa@698: // jcoomes@810: // These three phases are invoked in PSParallelCompact::invoke_no_policy(). The jcoomes@810: // marking phase is implemented in PSParallelCompact::marking_phase() and does a jcoomes@810: // complete marking of the heap. The summary phase is implemented in jcoomes@810: // PSParallelCompact::summary_phase(). The move and update phase is implemented jcoomes@810: // in PSParallelCompact::compact(). jmasa@698: // jcoomes@810: // A space that is being collected is divided into regions and with each region jcoomes@810: // is associated an object of type ParallelCompactData. Each region is of a jcoomes@810: // fixed size and typically will contain more than 1 object and may have parts jcoomes@810: // of objects at the front and back of the region. jmasa@698: // jcoomes@810: // region -----+---------------------+---------- jmasa@698: // objects covered [ AAA )[ BBB )[ CCC )[ DDD ) jmasa@698: // jcoomes@810: // The marking phase does a complete marking of all live objects in the heap. jcoomes@810: // The marking also compiles the size of the data for all live objects covered jcoomes@810: // by the region. This size includes the part of any live object spanning onto jcoomes@810: // the region (part of AAA if it is live) from the front, all live objects jcoomes@810: // contained in the region (BBB and/or CCC if they are live), and the part of jcoomes@810: // any live objects covered by the region that extends off the region (part of jcoomes@810: // DDD if it is live). The marking phase uses multiple GC threads and marking jcoomes@810: // is done in a bit array of type ParMarkBitMap. The marking of the bit map is jcoomes@810: // done atomically as is the accumulation of the size of the live objects jcoomes@810: // covered by a region. jmasa@698: // jcoomes@810: // The summary phase calculates the total live data to the left of each region jcoomes@810: // XXX. Based on that total and the bottom of the space, it can calculate the jcoomes@810: // starting location of the live data in XXX. The summary phase calculates for jcoomes@810: // each region XXX quantites such as jmasa@698: // jcoomes@810: // - the amount of live data at the beginning of a region from an object jcoomes@810: // entering the region. jcoomes@810: // - the location of the first live data on the region jcoomes@810: // - a count of the number of regions receiving live data from XXX. jmasa@698: // jmasa@698: // See ParallelCompactData for precise details. The summary phase also jcoomes@810: // calculates the dense prefix for the compaction. The dense prefix is a jcoomes@810: // portion at the beginning of the space that is not moved. The objects in the jcoomes@810: // dense prefix do need to have their object references updated. See method jcoomes@810: // summarize_dense_prefix(). jmasa@698: // jmasa@698: // The summary phase is done using 1 GC thread. jmasa@698: // jcoomes@810: // The compaction phase moves objects to their new location and updates all jcoomes@810: // references in the object. jmasa@698: // jcoomes@810: // A current exception is that objects that cross a region boundary are moved jcoomes@810: // but do not have their references updated. References are not updated because jcoomes@810: // it cannot easily be determined if the klass pointer KKK for the object AAA jcoomes@810: // has been updated. KKK likely resides in a region to the left of the region jcoomes@810: // containing AAA. These AAA's have there references updated at the end in a jcoomes@810: // clean up phase. See the method PSParallelCompact::update_deferred_objects(). jcoomes@810: // An alternate strategy is being investigated for this deferral of updating. jmasa@698: // jcoomes@810: // Compaction is done on a region basis. A region that is ready to be filled is jcoomes@810: // put on a ready list and GC threads take region off the list and fill them. A jcoomes@810: // region is ready to be filled if it empty of live objects. Such a region may jcoomes@810: // have been initially empty (only contained dead objects) or may have had all jcoomes@810: // its live objects copied out already. A region that compacts into itself is jcoomes@810: // also ready for filling. The ready list is initially filled with empty jcoomes@810: // regions and regions compacting into themselves. There is always at least 1 jcoomes@810: // region that can be put on the ready list. The regions are atomically added jcoomes@810: // and removed from the ready list. jcoomes@810: duke@435: class PSParallelCompact : AllStatic { duke@435: public: duke@435: // Convenient access to type names. duke@435: typedef ParMarkBitMap::idx_t idx_t; jcoomes@810: typedef ParallelCompactData::RegionData RegionData; jcoomes@5201: typedef ParallelCompactData::BlockData BlockData; duke@435: duke@435: typedef enum { coleenp@4037: old_space_id, eden_space_id, duke@435: from_space_id, to_space_id, last_space_id duke@435: } SpaceId; duke@435: duke@435: public: coleenp@548: // Inline closure decls duke@435: // duke@435: class IsAliveClosure: public BoolObjectClosure { duke@435: public: coleenp@548: virtual bool do_object_b(oop p); duke@435: }; duke@435: duke@435: class KeepAliveClosure: public OopClosure { coleenp@548: private: coleenp@548: ParCompactionManager* _compaction_manager; coleenp@548: protected: coleenp@548: template inline void do_oop_work(T* p); coleenp@548: public: coleenp@548: KeepAliveClosure(ParCompactionManager* cm) : _compaction_manager(cm) { } coleenp@548: virtual void do_oop(oop* p); coleenp@548: virtual void do_oop(narrowOop* p); coleenp@548: }; coleenp@548: duke@435: class FollowStackClosure: public VoidClosure { coleenp@548: private: duke@435: ParCompactionManager* _compaction_manager; duke@435: public: coleenp@548: FollowStackClosure(ParCompactionManager* cm) : _compaction_manager(cm) { } coleenp@548: virtual void do_void(); duke@435: }; duke@435: coleenp@4037: class AdjustPointerClosure: public OopClosure { duke@435: public: coleenp@548: virtual void do_oop(oop* p); coleenp@548: virtual void do_oop(narrowOop* p); jrose@1424: // do not walk from thread stacks to the code cache on this phase jrose@1424: virtual void do_code_blob(CodeBlob* cb) const { } duke@435: }; duke@435: coleenp@4037: class AdjustKlassClosure : public KlassClosure { coleenp@4037: public: coleenp@4037: void do_klass(Klass* klass); coleenp@4037: }; coleenp@4037: duke@435: friend class KeepAliveClosure; duke@435: friend class FollowStackClosure; duke@435: friend class AdjustPointerClosure; coleenp@4037: friend class AdjustKlassClosure; coleenp@4037: friend class FollowKlassClosure; coleenp@4047: friend class InstanceClassLoaderKlass; duke@435: friend class RefProcTaskProxy; duke@435: duke@435: private: sla@5237: static STWGCTimer _gc_timer; sla@5237: static ParallelOldTracer _gc_tracer; duke@435: static elapsedTimer _accumulated_time; duke@435: static unsigned int _total_invocations; duke@435: static unsigned int _maximum_compaction_gc_num; duke@435: static jlong _time_of_last_gc; // ms duke@435: static CollectorCounters* _counters; duke@435: static ParMarkBitMap _mark_bitmap; duke@435: static ParallelCompactData _summary_data; duke@435: static IsAliveClosure _is_alive_closure; duke@435: static SpaceInfo _space_info[last_space_id]; duke@435: static bool _print_phases; duke@435: static AdjustPointerClosure _adjust_pointer_closure; coleenp@4037: static AdjustKlassClosure _adjust_klass_closure; duke@435: duke@435: // Reference processing (used in ...follow_contents) duke@435: static ReferenceProcessor* _ref_processor; duke@435: duke@435: // Updated location of intArrayKlassObj. coleenp@4037: static Klass* _updated_int_array_klass_obj; duke@435: duke@435: // Values computed at initialization and used by dead_wood_limiter(). duke@435: static double _dwl_mean; duke@435: static double _dwl_std_dev; duke@435: static double _dwl_first_term; duke@435: static double _dwl_adjustment; duke@435: #ifdef ASSERT duke@435: static bool _dwl_initialized; duke@435: #endif // #ifdef ASSERT duke@435: duke@435: private: duke@435: duke@435: static void initialize_space_info(); duke@435: duke@435: // Return true if details about individual phases should be printed. duke@435: static inline bool print_phases(); duke@435: duke@435: // Clear the marking bitmap and summary data that cover the specified space. duke@435: static void clear_data_covering_space(SpaceId id); duke@435: duke@435: static void pre_compact(PreGCValues* pre_gc_values); duke@435: static void post_compact(); duke@435: duke@435: // Mark live objects duke@435: static void marking_phase(ParCompactionManager* cm, sla@5237: bool maximum_heap_compaction, sla@5237: ParallelOldTracer *gc_tracer); duke@435: coleenp@548: template coleenp@548: static inline void follow_root(ParCompactionManager* cm, T* p); duke@435: duke@435: // Compute the dense prefix for the designated space. This is an experimental duke@435: // implementation currently not used in production. duke@435: static HeapWord* compute_dense_prefix_via_density(const SpaceId id, duke@435: bool maximum_compaction); duke@435: duke@435: // Methods used to compute the dense prefix. duke@435: duke@435: // Compute the value of the normal distribution at x = density. The mean and duke@435: // standard deviation are values saved by initialize_dead_wood_limiter(). duke@435: static inline double normal_distribution(double density); duke@435: duke@435: // Initialize the static vars used by dead_wood_limiter(). duke@435: static void initialize_dead_wood_limiter(); duke@435: duke@435: // Return the percentage of space that can be treated as "dead wood" (i.e., duke@435: // not reclaimed). duke@435: static double dead_wood_limiter(double density, size_t min_percent); duke@435: jcoomes@810: // Find the first (left-most) region in the range [beg, end) that has at least duke@435: // dead_words of dead space to the left. The argument beg must be the first jcoomes@810: // region in the space that is not completely live. jcoomes@810: static RegionData* dead_wood_limit_region(const RegionData* beg, jcoomes@810: const RegionData* end, jcoomes@810: size_t dead_words); duke@435: jcoomes@810: // Return a pointer to the first region in the range [beg, end) that is not duke@435: // completely full. jcoomes@810: static RegionData* first_dead_space_region(const RegionData* beg, jcoomes@810: const RegionData* end); duke@435: duke@435: // Return a value indicating the benefit or 'yield' if the compacted region duke@435: // were to start (or equivalently if the dense prefix were to end) at the jcoomes@810: // candidate region. Higher values are better. duke@435: // duke@435: // The value is based on the amount of space reclaimed vs. the costs of (a) duke@435: // updating references in the dense prefix plus (b) copying objects and duke@435: // updating references in the compacted region. jcoomes@810: static inline double reclaimed_ratio(const RegionData* const candidate, duke@435: HeapWord* const bottom, duke@435: HeapWord* const top, duke@435: HeapWord* const new_top); duke@435: duke@435: // Compute the dense prefix for the designated space. duke@435: static HeapWord* compute_dense_prefix(const SpaceId id, duke@435: bool maximum_compaction); duke@435: jcoomes@810: // Return true if dead space crosses onto the specified Region; bit must be jcoomes@810: // the bit index corresponding to the first word of the Region. jcoomes@810: static inline bool dead_space_crosses_boundary(const RegionData* region, duke@435: idx_t bit); duke@435: duke@435: // Summary phase utility routine to fill dead space (if any) at the dense duke@435: // prefix boundary. Should only be called if the the dense prefix is duke@435: // non-empty. duke@435: static void fill_dense_prefix_end(SpaceId id); duke@435: jcoomes@917: // Clear the summary data source_region field for the specified addresses. jcoomes@917: static void clear_source_region(HeapWord* beg_addr, HeapWord* end_addr); jcoomes@917: jcoomes@918: #ifndef PRODUCT jcoomes@918: // Routines to provoke splitting a young gen space (ParallelOldGCSplitALot). jcoomes@918: jcoomes@918: // Fill the region [start, start + words) with live object(s). Only usable jcoomes@918: // for the old and permanent generations. jcoomes@918: static void fill_with_live_objects(SpaceId id, HeapWord* const start, jcoomes@918: size_t words); jcoomes@918: // Include the new objects in the summary data. jcoomes@918: static void summarize_new_objects(SpaceId id, HeapWord* start); jcoomes@918: jcoomes@931: // Add live objects to a survivor space since it's rare that both survivors jcoomes@931: // are non-empty. jcoomes@931: static void provoke_split_fill_survivor(SpaceId id); jcoomes@931: jcoomes@918: // Add live objects and/or choose the dense prefix to provoke splitting. jcoomes@918: static void provoke_split(bool & maximum_compaction); jcoomes@918: #endif jcoomes@918: duke@435: static void summarize_spaces_quick(); duke@435: static void summarize_space(SpaceId id, bool maximum_compaction); duke@435: static void summary_phase(ParCompactionManager* cm, bool maximum_compaction); duke@435: duke@435: // Adjust addresses in roots. Does not adjust addresses in heap. duke@435: static void adjust_roots(); duke@435: jcoomes@5201: DEBUG_ONLY(static void write_block_fill_histogram(outputStream* const out);) jcoomes@5201: duke@435: // Move objects to new locations. duke@435: static void compact_perm(ParCompactionManager* cm); duke@435: static void compact(); duke@435: jcoomes@810: // Add available regions to the stack and draining tasks to the task queue. jcoomes@810: static void enqueue_region_draining_tasks(GCTaskQueue* q, jcoomes@810: uint parallel_gc_threads); duke@435: duke@435: // Add dense prefix update tasks to the task queue. duke@435: static void enqueue_dense_prefix_tasks(GCTaskQueue* q, duke@435: uint parallel_gc_threads); duke@435: jcoomes@810: // Add region stealing tasks to the task queue. jcoomes@810: static void enqueue_region_stealing_tasks( duke@435: GCTaskQueue* q, duke@435: ParallelTaskTerminator* terminator_ptr, duke@435: uint parallel_gc_threads); duke@435: duke@435: // If objects are left in eden after a collection, try to move the boundary duke@435: // and absorb them into the old gen. Returns true if eden was emptied. duke@435: static bool absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, duke@435: PSYoungGen* young_gen, duke@435: PSOldGen* old_gen); duke@435: duke@435: // Reset time since last full gc duke@435: static void reset_millis_since_last_gc(); duke@435: duke@435: public: duke@435: class MarkAndPushClosure: public OopClosure { coleenp@548: private: duke@435: ParCompactionManager* _compaction_manager; duke@435: public: coleenp@548: MarkAndPushClosure(ParCompactionManager* cm) : _compaction_manager(cm) { } coleenp@548: virtual void do_oop(oop* p); coleenp@548: virtual void do_oop(narrowOop* p); duke@435: }; duke@435: coleenp@4037: // The one and only place to start following the classes. coleenp@4037: // Should only be applied to the ClassLoaderData klasses list. coleenp@4037: class FollowKlassClosure : public KlassClosure { coleenp@4037: private: coleenp@4037: MarkAndPushClosure* _mark_and_push_closure; coleenp@4037: public: coleenp@4037: FollowKlassClosure(MarkAndPushClosure* mark_and_push_closure) : coleenp@4037: _mark_and_push_closure(mark_and_push_closure) { } coleenp@4037: void do_klass(Klass* klass); coleenp@4037: }; coleenp@4037: duke@435: PSParallelCompact(); duke@435: duke@435: // Convenient accessor for Universe::heap(). duke@435: static ParallelScavengeHeap* gc_heap() { duke@435: return (ParallelScavengeHeap*)Universe::heap(); duke@435: } duke@435: duke@435: static void invoke(bool maximum_heap_compaction); jcoomes@3540: static bool invoke_no_policy(bool maximum_heap_compaction); duke@435: duke@435: static void post_initialize(); duke@435: // Perform initialization for PSParallelCompact that requires duke@435: // allocations. This should be called during the VM initialization duke@435: // at a pointer where it would be appropriate to return a JNI_ENOMEM duke@435: // in the event of a failure. duke@435: static bool initialize(); duke@435: coleenp@4037: // Closure accessors coleenp@4037: static OopClosure* adjust_pointer_closure() { return (OopClosure*)&_adjust_pointer_closure; } coleenp@4037: static KlassClosure* adjust_klass_closure() { return (KlassClosure*)&_adjust_klass_closure; } coleenp@4037: static BoolObjectClosure* is_alive_closure() { return (BoolObjectClosure*)&_is_alive_closure; } coleenp@4037: duke@435: // Public accessors duke@435: static elapsedTimer* accumulated_time() { return &_accumulated_time; } duke@435: static unsigned int total_invocations() { return _total_invocations; } duke@435: static CollectorCounters* counters() { return _counters; } duke@435: duke@435: // Used to add tasks duke@435: static GCTaskManager* const gc_task_manager(); coleenp@4037: static Klass* updated_int_array_klass_obj() { duke@435: return _updated_int_array_klass_obj; duke@435: } duke@435: duke@435: // Marking support duke@435: static inline bool mark_obj(oop obj); coleenp@4037: static inline bool is_marked(oop obj); coleenp@548: // Check mark and maybe push on marking stack coleenp@548: template static inline void mark_and_push(ParCompactionManager* cm, coleenp@548: T* p); stefank@5011: template static inline void adjust_pointer(T* p); duke@435: sjohanss@6169: static inline void follow_klass(ParCompactionManager* cm, Klass* klass); coleenp@4037: coleenp@4037: static void follow_class_loader(ParCompactionManager* cm, coleenp@4037: ClassLoaderData* klass); coleenp@4037: duke@435: // Compaction support. duke@435: // Return true if p is in the range [beg_addr, end_addr). duke@435: static inline bool is_in(HeapWord* p, HeapWord* beg_addr, HeapWord* end_addr); duke@435: static inline bool is_in(oop* p, HeapWord* beg_addr, HeapWord* end_addr); duke@435: duke@435: // Convenience wrappers for per-space data kept in _space_info. duke@435: static inline MutableSpace* space(SpaceId space_id); duke@435: static inline HeapWord* new_top(SpaceId space_id); duke@435: static inline HeapWord* dense_prefix(SpaceId space_id); duke@435: static inline ObjectStartArray* start_array(SpaceId space_id); duke@435: duke@435: // Move and update the live objects in the specified space. duke@435: static void move_and_update(ParCompactionManager* cm, SpaceId space_id); duke@435: jcoomes@810: // Process the end of the given region range in the dense prefix. duke@435: // This includes saving any object not updated. jcoomes@810: static void dense_prefix_regions_epilogue(ParCompactionManager* cm, jcoomes@810: size_t region_start_index, jcoomes@810: size_t region_end_index, jcoomes@810: idx_t exiting_object_offset, jcoomes@810: idx_t region_offset_start, jcoomes@810: idx_t region_offset_end); duke@435: jcoomes@810: // Update a region in the dense prefix. For each live object jcoomes@810: // in the region, update it's interior references. For each duke@435: // dead object, fill it with deadwood. Dead space at the end jcoomes@810: // of a region range will be filled to the start of the next jcoomes@810: // live object regardless of the region_index_end. None of the duke@435: // objects in the dense prefix move and dead space is dead duke@435: // (holds only dead objects that don't need any processing), so duke@435: // dead space can be filled in any order. duke@435: static void update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, duke@435: SpaceId space_id, jcoomes@810: size_t region_index_start, jcoomes@810: size_t region_index_end); duke@435: duke@435: // Return the address of the count + 1st live word in the range [beg, end). duke@435: static HeapWord* skip_live_words(HeapWord* beg, HeapWord* end, size_t count); duke@435: duke@435: // Return the address of the word to be copied to dest_addr, which must be jcoomes@810: // aligned to a region boundary. duke@435: static HeapWord* first_src_addr(HeapWord* const dest_addr, jcoomes@917: SpaceId src_space_id, jcoomes@810: size_t src_region_idx); duke@435: jcoomes@810: // Determine the next source region, set closure.source() to the start of the jcoomes@810: // new region return the region index. Parameter end_addr is the address one duke@435: // beyond the end of source range just processed. If necessary, switch to a duke@435: // new source space and set src_space_id (in-out parameter) and src_space_top duke@435: // (out parameter) accordingly. jcoomes@810: static size_t next_src_region(MoveAndUpdateClosure& closure, jcoomes@810: SpaceId& src_space_id, jcoomes@810: HeapWord*& src_space_top, jcoomes@810: HeapWord* end_addr); duke@435: jcoomes@810: // Decrement the destination count for each non-empty source region in the jcoomes@930: // range [beg_region, region(region_align_up(end_addr))). If the destination jcoomes@930: // count for a region goes to 0 and it needs to be filled, enqueue it. duke@435: static void decrement_destination_counts(ParCompactionManager* cm, jcoomes@930: SpaceId src_space_id, jcoomes@810: size_t beg_region, duke@435: HeapWord* end_addr); duke@435: jcoomes@810: // Fill a region, copying objects from one or more source regions. jcoomes@810: static void fill_region(ParCompactionManager* cm, size_t region_idx); jcoomes@810: static void fill_and_update_region(ParCompactionManager* cm, size_t region) { jcoomes@810: fill_region(cm, region); duke@435: } duke@435: jcoomes@5201: // Fill in the block table for the specified region. jcoomes@5201: static void fill_blocks(size_t region_idx); jcoomes@5201: duke@435: // Update the deferred objects in the space. duke@435: static void update_deferred_objects(ParCompactionManager* cm, SpaceId id); duke@435: duke@435: static ParMarkBitMap* mark_bitmap() { return &_mark_bitmap; } duke@435: static ParallelCompactData& summary_data() { return _summary_data; } duke@435: duke@435: // Reference Processing duke@435: static ReferenceProcessor* const ref_processor() { return _ref_processor; } duke@435: sla@5237: static STWGCTimer* gc_timer() { return &_gc_timer; } sla@5237: duke@435: // Return the SpaceId for the given address. duke@435: static SpaceId space_id(HeapWord* addr); duke@435: duke@435: // Time since last full gc (in milliseconds). duke@435: static jlong millis_since_last_gc(); duke@435: stefank@4904: static void print_on_error(outputStream* st); stefank@4904: duke@435: #ifndef PRODUCT duke@435: // Debugging support. duke@435: static const char* space_names[last_space_id]; jcoomes@810: static void print_region_ranges(); duke@435: static void print_dense_prefix_stats(const char* const algorithm, duke@435: const SpaceId id, duke@435: const bool maximum_compaction, duke@435: HeapWord* const addr); jcoomes@917: static void summary_phase_msg(SpaceId dst_space_id, jcoomes@917: HeapWord* dst_beg, HeapWord* dst_end, jcoomes@917: SpaceId src_space_id, jcoomes@917: HeapWord* src_beg, HeapWord* src_end); duke@435: #endif // #ifndef PRODUCT duke@435: duke@435: #ifdef ASSERT jcoomes@930: // Sanity check the new location of a word in the heap. jcoomes@930: static inline void check_new_location(HeapWord* old_addr, HeapWord* new_addr); jcoomes@810: // Verify that all the regions have been emptied. duke@435: static void verify_complete(SpaceId space_id); duke@435: #endif // #ifdef ASSERT duke@435: }; duke@435: coleenp@548: inline bool PSParallelCompact::mark_obj(oop obj) { duke@435: const int obj_size = obj->size(); duke@435: if (mark_bitmap()->mark_obj(obj, obj_size)) { duke@435: _summary_data.add_obj(obj, obj_size); duke@435: return true; duke@435: } else { duke@435: return false; duke@435: } duke@435: } duke@435: coleenp@4037: inline bool PSParallelCompact::is_marked(oop obj) { coleenp@4037: return mark_bitmap()->is_marked(obj); coleenp@4037: } coleenp@4037: coleenp@548: template coleenp@548: inline void PSParallelCompact::follow_root(ParCompactionManager* cm, T* p) { coleenp@548: assert(!Universe::heap()->is_in_reserved(p), coleenp@548: "roots shouldn't be things within the heap"); johnc@4384: coleenp@548: T heap_oop = oopDesc::load_heap_oop(p); coleenp@548: if (!oopDesc::is_null(heap_oop)) { coleenp@548: oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); coleenp@548: if (mark_bitmap()->is_unmarked(obj)) { coleenp@548: if (mark_obj(obj)) { coleenp@548: obj->follow_contents(cm); coleenp@548: } coleenp@548: } coleenp@548: } jcoomes@1746: cm->follow_marking_stacks(); coleenp@548: } coleenp@548: coleenp@548: template coleenp@548: inline void PSParallelCompact::mark_and_push(ParCompactionManager* cm, T* p) { coleenp@548: T heap_oop = oopDesc::load_heap_oop(p); coleenp@548: if (!oopDesc::is_null(heap_oop)) { coleenp@548: oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); jcoomes@1993: if (mark_bitmap()->is_unmarked(obj) && mark_obj(obj)) { jcoomes@1993: cm->push(obj); coleenp@548: } coleenp@548: } coleenp@548: } coleenp@548: coleenp@548: template stefank@5011: inline void PSParallelCompact::adjust_pointer(T* p) { coleenp@548: T heap_oop = oopDesc::load_heap_oop(p); coleenp@548: if (!oopDesc::is_null(heap_oop)) { coleenp@548: oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); coleenp@548: oop new_obj = (oop)summary_data().calc_new_pointer(obj); coleenp@4037: assert(new_obj != NULL, // is forwarding ptr? coleenp@548: "should be forwarded"); coleenp@548: // Just always do the update unconditionally? coleenp@548: if (new_obj != NULL) { coleenp@548: assert(Universe::heap()->is_in_reserved(new_obj), coleenp@548: "should be in object space"); coleenp@548: oopDesc::encode_store_heap_oop_not_null(p, new_obj); coleenp@548: } coleenp@548: } coleenp@548: } coleenp@548: sjohanss@6169: inline void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) { sjohanss@6169: oop holder = klass->klass_holder(); sjohanss@6169: PSParallelCompact::mark_and_push(cm, &holder); sjohanss@6169: } sjohanss@6169: coleenp@548: template coleenp@548: inline void PSParallelCompact::KeepAliveClosure::do_oop_work(T* p) { coleenp@548: mark_and_push(_compaction_manager, p); coleenp@548: } coleenp@548: coleenp@548: inline bool PSParallelCompact::print_phases() { duke@435: return _print_phases; duke@435: } duke@435: coleenp@548: inline double PSParallelCompact::normal_distribution(double density) { duke@435: assert(_dwl_initialized, "uninitialized"); duke@435: const double squared_term = (density - _dwl_mean) / _dwl_std_dev; duke@435: return _dwl_first_term * exp(-0.5 * squared_term * squared_term); duke@435: } duke@435: duke@435: inline bool jcoomes@810: PSParallelCompact::dead_space_crosses_boundary(const RegionData* region, duke@435: idx_t bit) duke@435: { jcoomes@810: assert(bit > 0, "cannot call this for the first bit/region"); jcoomes@810: assert(_summary_data.region_to_addr(region) == _mark_bitmap.bit_to_addr(bit), duke@435: "sanity check"); duke@435: duke@435: // Dead space crosses the boundary if (1) a partial object does not extend jcoomes@810: // onto the region, (2) an object does not start at the beginning of the jcoomes@810: // region, and (3) an object does not end at the end of the prior region. jcoomes@810: return region->partial_obj_size() == 0 && duke@435: !_mark_bitmap.is_obj_beg(bit) && duke@435: !_mark_bitmap.is_obj_end(bit - 1); duke@435: } duke@435: duke@435: inline bool duke@435: PSParallelCompact::is_in(HeapWord* p, HeapWord* beg_addr, HeapWord* end_addr) { duke@435: return p >= beg_addr && p < end_addr; duke@435: } duke@435: duke@435: inline bool duke@435: PSParallelCompact::is_in(oop* p, HeapWord* beg_addr, HeapWord* end_addr) { duke@435: return is_in((HeapWord*)p, beg_addr, end_addr); duke@435: } duke@435: duke@435: inline MutableSpace* PSParallelCompact::space(SpaceId id) { duke@435: assert(id < last_space_id, "id out of range"); duke@435: return _space_info[id].space(); duke@435: } duke@435: duke@435: inline HeapWord* PSParallelCompact::new_top(SpaceId id) { duke@435: assert(id < last_space_id, "id out of range"); duke@435: return _space_info[id].new_top(); duke@435: } duke@435: duke@435: inline HeapWord* PSParallelCompact::dense_prefix(SpaceId id) { duke@435: assert(id < last_space_id, "id out of range"); duke@435: return _space_info[id].dense_prefix(); duke@435: } duke@435: duke@435: inline ObjectStartArray* PSParallelCompact::start_array(SpaceId id) { duke@435: assert(id < last_space_id, "id out of range"); duke@435: return _space_info[id].start_array(); duke@435: } duke@435: jcoomes@930: #ifdef ASSERT jcoomes@930: inline void jcoomes@930: PSParallelCompact::check_new_location(HeapWord* old_addr, HeapWord* new_addr) jcoomes@930: { jcoomes@930: assert(old_addr >= new_addr || space_id(old_addr) != space_id(new_addr), jcoomes@930: "must move left or to a different space"); kvn@1926: assert(is_object_aligned((intptr_t)old_addr) && is_object_aligned((intptr_t)new_addr), kvn@1926: "checking alignment"); jcoomes@930: } jcoomes@930: #endif // ASSERT jcoomes@930: duke@435: class MoveAndUpdateClosure: public ParMarkBitMapClosure { duke@435: public: duke@435: inline MoveAndUpdateClosure(ParMarkBitMap* bitmap, ParCompactionManager* cm, duke@435: ObjectStartArray* start_array, duke@435: HeapWord* destination, size_t words); duke@435: duke@435: // Accessors. duke@435: HeapWord* destination() const { return _destination; } duke@435: duke@435: // If the object will fit (size <= words_remaining()), copy it to the current duke@435: // destination, update the interior oops and the start array and return either duke@435: // full (if the closure is full) or incomplete. If the object will not fit, duke@435: // return would_overflow. duke@435: virtual IterationStatus do_addr(HeapWord* addr, size_t size); duke@435: duke@435: // Copy enough words to fill this closure, starting at source(). Interior duke@435: // oops and the start array are not updated. Return full. duke@435: IterationStatus copy_until_full(); duke@435: duke@435: // Copy enough words to fill this closure or to the end of an object, duke@435: // whichever is smaller, starting at source(). Interior oops and the start duke@435: // array are not updated. duke@435: void copy_partial_obj(); duke@435: duke@435: protected: duke@435: // Update variables to indicate that word_count words were processed. duke@435: inline void update_state(size_t word_count); duke@435: duke@435: protected: duke@435: ObjectStartArray* const _start_array; duke@435: HeapWord* _destination; // Next addr to be written. duke@435: }; duke@435: duke@435: inline duke@435: MoveAndUpdateClosure::MoveAndUpdateClosure(ParMarkBitMap* bitmap, duke@435: ParCompactionManager* cm, duke@435: ObjectStartArray* start_array, duke@435: HeapWord* destination, duke@435: size_t words) : duke@435: ParMarkBitMapClosure(bitmap, cm, words), _start_array(start_array) duke@435: { duke@435: _destination = destination; duke@435: } duke@435: duke@435: inline void MoveAndUpdateClosure::update_state(size_t words) duke@435: { duke@435: decrement_words_remaining(words); duke@435: _source += words; duke@435: _destination += words; duke@435: } duke@435: duke@435: class UpdateOnlyClosure: public ParMarkBitMapClosure { duke@435: private: duke@435: const PSParallelCompact::SpaceId _space_id; duke@435: ObjectStartArray* const _start_array; duke@435: duke@435: public: duke@435: UpdateOnlyClosure(ParMarkBitMap* mbm, duke@435: ParCompactionManager* cm, duke@435: PSParallelCompact::SpaceId space_id); duke@435: duke@435: // Update the object. duke@435: virtual IterationStatus do_addr(HeapWord* addr, size_t words); duke@435: duke@435: inline void do_addr(HeapWord* addr); duke@435: }; duke@435: coleenp@548: inline void UpdateOnlyClosure::do_addr(HeapWord* addr) coleenp@548: { duke@435: _start_array->allocate_block(addr); duke@435: oop(addr)->update_contents(compaction_manager()); duke@435: } duke@435: jcoomes@916: class FillClosure: public ParMarkBitMapClosure jcoomes@916: { jcoomes@916: public: coleenp@548: FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : duke@435: ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), jcoomes@916: _start_array(PSParallelCompact::start_array(space_id)) jcoomes@916: { coleenp@4037: assert(space_id == PSParallelCompact::old_space_id, duke@435: "cannot use FillClosure in the young gen"); duke@435: } duke@435: duke@435: virtual IterationStatus do_addr(HeapWord* addr, size_t size) { jcoomes@916: CollectedHeap::fill_with_objects(addr, size); jcoomes@916: HeapWord* const end = addr + size; jcoomes@916: do { jcoomes@916: _start_array->allocate_block(addr); jcoomes@916: addr += oop(addr)->size(); jcoomes@916: } while (addr < end); duke@435: return ParMarkBitMap::incomplete; duke@435: } duke@435: duke@435: private: jcoomes@916: ObjectStartArray* const _start_array; duke@435: }; stefank@2314: stefank@2314: #endif // SHARE_VM_GC_IMPLEMENTATION_PARALLELSCAVENGE_PSPARALLELCOMPACT_HPP