src/hotspot/share/services/heapDumper.cpp - master - githubmirror/galahad

src/hotspot/share/services/heapDumper.cpp
2 797 строк95 KB
Anton Artemov
8367485: os::physical_memory is broken in 32-bit JVMs when running on 64-bit OSes
30 сен 2025, 08:09
30 сен 2025, 08:092746c1a
/*
 * Copyright (c) 2005, 2025, Oracle and/or its affiliates. All rights reserved.
 * Copyright (c) 2023, Alibaba Group Holding Limited. All rights reserved.
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * This code is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License version 2 only, as
 * published by the Free Software Foundation.
 *
 * This code is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 * version 2 for more details (a copy is included in the LICENSE file that
 * accompanied this code).
 *
 * You should have received a copy of the GNU General Public License version
 * 2 along with this work; if not, write to the Free Software Foundation,
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 *
 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 * or visit www.oracle.com if you need additional information or have any
 * questions.
 *
 */

#include "classfile/classLoaderData.inline.hpp"
#include "classfile/classLoaderDataGraph.hpp"
#include "classfile/javaClasses.inline.hpp"
#include "classfile/symbolTable.hpp"
#include "classfile/vmClasses.hpp"
#include "classfile/vmSymbols.hpp"
#include "gc/shared/gcLocker.hpp"
#include "gc/shared/gcVMOperations.hpp"
#include "gc/shared/workerThread.hpp"
#include "jfr/jfrEvents.hpp"
#include "jvm.h"
#include "memory/allocation.inline.hpp"
#include "memory/resourceArea.hpp"
#include "memory/universe.hpp"
#include "oops/fieldStreams.inline.hpp"
#include "oops/klass.inline.hpp"
#include "oops/objArrayKlass.hpp"
#include "oops/objArrayOop.inline.hpp"
#include "oops/oop.inline.hpp"
#include "oops/typeArrayOop.inline.hpp"
#include "runtime/arguments.hpp"
#include "runtime/continuationWrapper.inline.hpp"
#include "runtime/frame.inline.hpp"
#include "runtime/handles.inline.hpp"
#include "runtime/javaCalls.hpp"
#include "runtime/javaThread.inline.hpp"
#include "runtime/jniHandles.hpp"
#include "runtime/os.hpp"
#include "runtime/threads.hpp"
#include "runtime/threadSMR.hpp"
#include "runtime/timerTrace.hpp"
#include "runtime/vframe.hpp"
#include "runtime/vmOperations.hpp"
#include "runtime/vmThread.hpp"
#include "services/heapDumper.hpp"
#include "services/heapDumperCompression.hpp"
#include "services/threadService.hpp"
#include "utilities/checkedCast.hpp"
#include "utilities/macros.hpp"
#include "utilities/ostream.hpp"
#ifdef LINUX
#include "os_linux.hpp"
#endif

/*
 * HPROF binary format - description copied from:
 *   src/share/demo/jvmti/hprof/hprof_io.c
 *
 *
 *  header    "JAVA PROFILE 1.0.2" (0-terminated)
 *
 *  u4        size of identifiers. Identifiers are used to represent
 *            UTF8 strings, objects, stack traces, etc. They usually
 *            have the same size as host pointers.
 * u4         high word
 * u4         low word    number of milliseconds since 0:00 GMT, 1/1/70
 * [record]*  a sequence of records.
 *
 *
 * Record format:
 *
 * u1         a TAG denoting the type of the record
 * u4         number of *microseconds* since the time stamp in the
 *            header. (wraps around in a little more than an hour)
 * u4         number of bytes *remaining* in the record. Note that
 *            this number excludes the tag and the length field itself.
 * [u1]*      BODY of the record (a sequence of bytes)
 *
 *
 * The following TAGs are supported:
 *
 * TAG           BODY       notes
 *----------------------------------------------------------
 * HPROF_UTF8               a UTF8-encoded name
 *
 *               id         name ID
 *               [u1]*      UTF8 characters (no trailing zero)
 *
 * HPROF_LOAD_CLASS         a newly loaded class
 *
 *                u4        class serial number (> 0)
 *                id        class object ID
 *                u4        stack trace serial number
 *                id        class name ID
 *
 * HPROF_UNLOAD_CLASS       an unloading class
 *
 *                u4        class serial_number
 *
 * HPROF_FRAME              a Java stack frame
 *
 *                id        stack frame ID
 *                id        method name ID
 *                id        method signature ID
 *                id        source file name ID
 *                u4        class serial number
 *                i4        line number. >0: normal
 *                                       -1: unknown
 *                                       -2: compiled method
 *                                       -3: native method
 *
 * HPROF_TRACE              a Java stack trace
 *
 *               u4         stack trace serial number
 *               u4         thread serial number
 *               u4         number of frames
 *               [id]*      stack frame IDs
 *
 *
 * HPROF_ALLOC_SITES        a set of heap allocation sites, obtained after GC
 *
 *               u2         flags 0x0001: incremental vs. complete
 *                                0x0002: sorted by allocation vs. live
 *                                0x0004: whether to force a GC
 *               u4         cutoff ratio
 *               u4         total live bytes
 *               u4         total live instances
 *               u8         total bytes allocated
 *               u8         total instances allocated
 *               u4         number of sites that follow
 *               [u1        is_array: 0:  normal object
 *                                    2:  object array
 *                                    4:  boolean array
 *                                    5:  char array
 *                                    6:  float array
 *                                    7:  double array
 *                                    8:  byte array
 *                                    9:  short array
 *                                    10: int array
 *                                    11: long array
 *                u4        class serial number (may be zero during startup)
 *                u4        stack trace serial number
 *                u4        number of bytes alive
 *                u4        number of instances alive
 *                u4        number of bytes allocated
 *                u4]*      number of instance allocated
 *
 * HPROF_START_THREAD       a newly started thread.
 *
 *               u4         thread serial number (> 0)
 *               id         thread object ID
 *               u4         stack trace serial number
 *               id         thread name ID
 *               id         thread group name ID
 *               id         thread group parent name ID
 *
 * HPROF_END_THREAD         a terminating thread.
 *
 *               u4         thread serial number
 *
 * HPROF_HEAP_SUMMARY       heap summary
 *
 *               u4         total live bytes
 *               u4         total live instances
 *               u8         total bytes allocated
 *               u8         total instances allocated
 *
 * HPROF_HEAP_DUMP          denote a heap dump
 *
 *               [heap dump sub-records]*
 *
 *                          There are four kinds of heap dump sub-records:
 *
 *               u1         sub-record type
 *
 *               HPROF_GC_ROOT_UNKNOWN         unknown root
 *
 *                          id         object ID
 *
 *               HPROF_GC_ROOT_THREAD_OBJ      thread object
 *
 *                          id         thread object ID  (may be 0 for a
 *                                     thread newly attached through JNI)
 *                          u4         thread sequence number
 *                          u4         stack trace sequence number
 *
 *               HPROF_GC_ROOT_JNI_GLOBAL      JNI global ref root
 *
 *                          id         object ID
 *                          id         JNI global ref ID
 *
 *               HPROF_GC_ROOT_JNI_LOCAL       JNI local ref
 *
 *                          id         object ID
 *                          u4         thread serial number
 *                          u4         frame # in stack trace (-1 for empty)
 *
 *               HPROF_GC_ROOT_JAVA_FRAME      Java stack frame
 *
 *                          id         object ID
 *                          u4         thread serial number
 *                          u4         frame # in stack trace (-1 for empty)
 *
 *               HPROF_GC_ROOT_NATIVE_STACK    Native stack
 *
 *                          id         object ID
 *                          u4         thread serial number
 *
 *               HPROF_GC_ROOT_STICKY_CLASS    System class
 *
 *                          id         object ID
 *
 *               HPROF_GC_ROOT_THREAD_BLOCK    Reference from thread block
 *
 *                          id         object ID
 *                          u4         thread serial number
 *
 *               HPROF_GC_ROOT_MONITOR_USED    Busy monitor
 *
 *                          id         object ID
 *
 *               HPROF_GC_CLASS_DUMP           dump of a class object
 *
 *                          id         class object ID
 *                          u4         stack trace serial number
 *                          id         super class object ID
 *                          id         class loader object ID
 *                          id         signers object ID
 *                          id         protection domain object ID
 *                          id         reserved
 *                          id         reserved
 *
 *                          u4         instance size (in bytes)
 *
 *                          u2         size of constant pool
 *                          [u2,       constant pool index,
 *                           ty,       type
 *                                     2:  object
 *                                     4:  boolean
 *                                     5:  char
 *                                     6:  float
 *                                     7:  double
 *                                     8:  byte
 *                                     9:  short
 *                                     10: int
 *                                     11: long
 *                           vl]*      and value
 *
 *                          u2         number of static fields
 *                          [id,       static field name,
 *                           ty,       type,
 *                           vl]*      and value
 *
 *                          u2         number of inst. fields (not inc. super)
 *                          [id,       instance field name,
 *                           ty]*      type
 *
 *               HPROF_GC_INSTANCE_DUMP        dump of a normal object
 *
 *                          id         object ID
 *                          u4         stack trace serial number
 *                          id         class object ID
 *                          u4         number of bytes that follow
 *                          [vl]*      instance field values (class, followed
 *                                     by super, super's super ...)
 *
 *               HPROF_GC_OBJ_ARRAY_DUMP       dump of an object array
 *
 *                          id         array object ID
 *                          u4         stack trace serial number
 *                          u4         number of elements
 *                          id         array class ID
 *                          [id]*      elements
 *
 *               HPROF_GC_PRIM_ARRAY_DUMP      dump of a primitive array
 *
 *                          id         array object ID
 *                          u4         stack trace serial number
 *                          u4         number of elements
 *                          u1         element type
 *                                     4:  boolean array
 *                                     5:  char array
 *                                     6:  float array
 *                                     7:  double array
 *                                     8:  byte array
 *                                     9:  short array
 *                                     10: int array
 *                                     11: long array
 *                          [u1]*      elements
 *
 * HPROF_CPU_SAMPLES        a set of sample traces of running threads
 *
 *                u4        total number of samples
 *                u4        # of traces
 *               [u4        # of samples
 *                u4]*      stack trace serial number
 *
 * HPROF_CONTROL_SETTINGS   the settings of on/off switches
 *
 *                u4        0x00000001: alloc traces on/off
 *                          0x00000002: cpu sampling on/off
 *                u2        stack trace depth
 *
 *
 * When the header is "JAVA PROFILE 1.0.2" a heap dump can optionally
 * be generated as a sequence of heap dump segments. This sequence is
 * terminated by an end record. The additional tags allowed by format
 * "JAVA PROFILE 1.0.2" are:
 *
 * HPROF_HEAP_DUMP_SEGMENT  denote a heap dump segment
 *
 *               [heap dump sub-records]*
 *               The same sub-record types allowed by HPROF_HEAP_DUMP
 *
 * HPROF_HEAP_DUMP_END      denotes the end of a heap dump
 *
 */


// HPROF tags

enum hprofTag : u1 {
  // top-level records
  HPROF_UTF8                    = 0x01,
  HPROF_LOAD_CLASS              = 0x02,
  HPROF_UNLOAD_CLASS            = 0x03,
  HPROF_FRAME                   = 0x04,
  HPROF_TRACE                   = 0x05,
  HPROF_ALLOC_SITES             = 0x06,
  HPROF_HEAP_SUMMARY            = 0x07,
  HPROF_START_THREAD            = 0x0A,
  HPROF_END_THREAD              = 0x0B,
  HPROF_HEAP_DUMP               = 0x0C,
  HPROF_CPU_SAMPLES             = 0x0D,
  HPROF_CONTROL_SETTINGS        = 0x0E,

  // 1.0.2 record types
  HPROF_HEAP_DUMP_SEGMENT       = 0x1C,
  HPROF_HEAP_DUMP_END           = 0x2C,

  // field types
  HPROF_ARRAY_OBJECT            = 0x01,
  HPROF_NORMAL_OBJECT           = 0x02,
  HPROF_BOOLEAN                 = 0x04,
  HPROF_CHAR                    = 0x05,
  HPROF_FLOAT                   = 0x06,
  HPROF_DOUBLE                  = 0x07,
  HPROF_BYTE                    = 0x08,
  HPROF_SHORT                   = 0x09,
  HPROF_INT                     = 0x0A,
  HPROF_LONG                    = 0x0B,

  // data-dump sub-records
  HPROF_GC_ROOT_UNKNOWN         = 0xFF,
  HPROF_GC_ROOT_JNI_GLOBAL      = 0x01,
  HPROF_GC_ROOT_JNI_LOCAL       = 0x02,
  HPROF_GC_ROOT_JAVA_FRAME      = 0x03,
  HPROF_GC_ROOT_NATIVE_STACK    = 0x04,
  HPROF_GC_ROOT_STICKY_CLASS    = 0x05,
  HPROF_GC_ROOT_THREAD_BLOCK    = 0x06,
  HPROF_GC_ROOT_MONITOR_USED    = 0x07,
  HPROF_GC_ROOT_THREAD_OBJ      = 0x08,
  HPROF_GC_CLASS_DUMP           = 0x20,
  HPROF_GC_INSTANCE_DUMP        = 0x21,
  HPROF_GC_OBJ_ARRAY_DUMP       = 0x22,
  HPROF_GC_PRIM_ARRAY_DUMP      = 0x23
};

// Default stack trace ID (used for dummy HPROF_TRACE record)
enum {
  STACK_TRACE_ID = 1,
  INITIAL_CLASS_COUNT = 200
};

// Supports I/O operations for a dump
// Base class for dump and parallel dump
class AbstractDumpWriter : public CHeapObj<mtInternal> {
 protected:
  enum {
    io_buffer_max_size = 1*M,
    dump_segment_header_size = 9
  };

  char* _buffer;    // internal buffer
  size_t _size;
  size_t _pos;

  bool _in_dump_segment; // Are we currently in a dump segment?
  bool _is_huge_sub_record; // Are we writing a sub-record larger than the buffer size?
  DEBUG_ONLY(size_t _sub_record_left;) // The bytes not written for the current sub-record.
  DEBUG_ONLY(bool _sub_record_ended;) // True if we have called the end_sub_record().

  char* buffer() const                          { return _buffer; }
  size_t buffer_size() const                    { return _size; }
  void set_position(size_t pos)                 { _pos = pos; }

  // Can be called if we have enough room in the buffer.
  void write_fast(const void* s, size_t len);

  // Returns true if we have enough room in the buffer for 'len' bytes.
  bool can_write_fast(size_t len);

  void write_address(address a);

 public:
  AbstractDumpWriter() :
    _buffer(nullptr),
    _size(io_buffer_max_size),
    _pos(0),
    _in_dump_segment(false) { }

  // Total number of bytes written to the disk
  virtual julong bytes_written() const = 0;
  // Return non-null if error occurred
  virtual char const* error() const = 0;

  size_t position() const                       { return _pos; }
  // writer functions
  virtual void write_raw(const void* s, size_t len);
  void write_u1(u1 x);
  void write_u2(u2 x);
  void write_u4(u4 x);
  void write_u8(u8 x);
  void write_objectID(oop o);
  void write_rootID(oop* p);
  void write_symbolID(Symbol* o);
  void write_classID(Klass* k);
  void write_id(u4 x);

  // Start a new sub-record. Starts a new heap dump segment if needed.
  void start_sub_record(u1 tag, u4 len);
  // Ends the current sub-record.
  void end_sub_record();
  // Finishes the current dump segment if not already finished.
  void finish_dump_segment();
  // Flush internal buffer to persistent storage
  virtual void flush() = 0;
};

void AbstractDumpWriter::write_fast(const void* s, size_t len) {
  assert(!_in_dump_segment || (_sub_record_left >= len), "sub-record too large");
  assert(buffer_size() - position() >= len, "Must fit");
  DEBUG_ONLY(_sub_record_left -= len);
  memcpy(buffer() + position(), s, len);
  set_position(position() + len);
}

bool AbstractDumpWriter::can_write_fast(size_t len) {
  return buffer_size() - position() >= len;
}

// write raw bytes
void AbstractDumpWriter::write_raw(const void* s, size_t len) {
  assert(!_in_dump_segment || (_sub_record_left >= len), "sub-record too large");
  DEBUG_ONLY(_sub_record_left -= len);

  // flush buffer to make room.
  while (len > buffer_size() - position()) {
    assert(!_in_dump_segment || _is_huge_sub_record,
           "Cannot overflow in non-huge sub-record.");
    size_t to_write = buffer_size() - position();
    memcpy(buffer() + position(), s, to_write);
    s = (void*) ((char*) s + to_write);
    len -= to_write;
    set_position(position() + to_write);
    flush();
  }

  memcpy(buffer() + position(), s, len);
  set_position(position() + len);
}

// Makes sure we inline the fast write into the write_u* functions. This is a big speedup.
#define WRITE_KNOWN_TYPE(p, len) do { if (can_write_fast((len))) write_fast((p), (len)); \
                                      else write_raw((p), (len)); } while (0)

void AbstractDumpWriter::write_u1(u1 x) {
  WRITE_KNOWN_TYPE(&x, 1);
}

void AbstractDumpWriter::write_u2(u2 x) {
  u2 v;
  Bytes::put_Java_u2((address)&v, x);
  WRITE_KNOWN_TYPE(&v, 2);
}

void AbstractDumpWriter::write_u4(u4 x) {
  u4 v;
  Bytes::put_Java_u4((address)&v, x);
  WRITE_KNOWN_TYPE(&v, 4);
}

void AbstractDumpWriter::write_u8(u8 x) {
  u8 v;
  Bytes::put_Java_u8((address)&v, x);
  WRITE_KNOWN_TYPE(&v, 8);
}

void AbstractDumpWriter::write_address(address a) {
#ifdef _LP64
  write_u8((u8)a);
#else
  write_u4((u4)a);
#endif
}

void AbstractDumpWriter::write_objectID(oop o) {
  write_address(cast_from_oop<address>(o));
}

void AbstractDumpWriter::write_rootID(oop* p) {
  write_address((address)p);
}

void AbstractDumpWriter::write_symbolID(Symbol* s) {
  write_address((address)((uintptr_t)s));
}

void AbstractDumpWriter::write_id(u4 x) {
#ifdef _LP64
  write_u8((u8) x);
#else
  write_u4(x);
#endif
}

// We use java mirror as the class ID
void AbstractDumpWriter::write_classID(Klass* k) {
  write_objectID(k->java_mirror());
}

void AbstractDumpWriter::finish_dump_segment() {
  if (_in_dump_segment) {
    assert(_sub_record_left == 0, "Last sub-record not written completely");
    assert(_sub_record_ended, "sub-record must have ended");

    // Fix up the dump segment length if we haven't written a huge sub-record last
    // (in which case the segment length was already set to the correct value initially).
    if (!_is_huge_sub_record) {
      assert(position() > dump_segment_header_size, "Dump segment should have some content");
      Bytes::put_Java_u4((address) (buffer() + 5),
                         (u4) (position() - dump_segment_header_size));
    } else {
      // Finish process huge sub record
      // Set _is_huge_sub_record to false so the parallel dump writer can flush data to file.
      _is_huge_sub_record = false;
    }

    _in_dump_segment = false;
    flush();
  }
}

void AbstractDumpWriter::start_sub_record(u1 tag, u4 len) {
  if (!_in_dump_segment) {
    if (position() > 0) {
      flush();
    }

    assert(position() == 0 && buffer_size() > dump_segment_header_size, "Must be at the start");

    write_u1(HPROF_HEAP_DUMP_SEGMENT);
    write_u4(0); // timestamp
    // Will be fixed up later if we add more sub-records.  If this is a huge sub-record,
    // this is already the correct length, since we don't add more sub-records.
    write_u4(len);
    assert(Bytes::get_Java_u4((address)(buffer() + 5)) == len, "Inconsistent size!");
    _in_dump_segment = true;
    _is_huge_sub_record = len > buffer_size() - dump_segment_header_size;
  } else if (_is_huge_sub_record || (len > buffer_size() - position())) {
    // This object will not fit in completely or the last sub-record was huge.
    // Finish the current segment and try again.
    finish_dump_segment();
    start_sub_record(tag, len);

    return;
  }

  DEBUG_ONLY(_sub_record_left = len);
  DEBUG_ONLY(_sub_record_ended = false);

  write_u1(tag);
}

void AbstractDumpWriter::end_sub_record() {
  assert(_in_dump_segment, "must be in dump segment");
  assert(_sub_record_left == 0, "sub-record not written completely");
  assert(!_sub_record_ended, "Must not have ended yet");
  DEBUG_ONLY(_sub_record_ended = true);
}

// Supports I/O operations for a dump

class DumpWriter : public AbstractDumpWriter {
private:
  FileWriter* _writer;
  AbstractCompressor* _compressor;
  size_t _bytes_written;
  char* _error;
  // Compression support
  char* _out_buffer;
  size_t _out_size;
  size_t _out_pos;
  char* _tmp_buffer;
  size_t _tmp_size;

private:
  void do_compress();

public:
  DumpWriter(const char* path, bool overwrite, AbstractCompressor* compressor);
  ~DumpWriter();
  julong bytes_written() const override        { return (julong) _bytes_written; }
  char const* error() const override           { return _error; }
  void set_error(const char* error)            { _error = (char*)error; }
  bool has_error() const                       { return _error != nullptr; }
  const char* get_file_path() const            { return _writer->get_file_path(); }
  AbstractCompressor* compressor()             { return _compressor; }
  bool is_overwrite() const                    { return _writer->is_overwrite(); }

  void flush() override;

private:
  // internals for DumpMerger
  friend class DumpMerger;
  void set_bytes_written(julong bytes_written) { _bytes_written = bytes_written; }
  int get_fd() const                           { return _writer->get_fd(); }
  void set_compressor(AbstractCompressor* p)   { _compressor = p; }
};

DumpWriter::DumpWriter(const char* path, bool overwrite, AbstractCompressor* compressor) :
  AbstractDumpWriter(),
  _writer(new (std::nothrow) FileWriter(path, overwrite)),
  _compressor(compressor),
  _bytes_written(0),
  _error(nullptr),
  _out_buffer(nullptr),
  _out_size(0),
  _out_pos(0),
  _tmp_buffer(nullptr),
  _tmp_size(0) {
  _error = (char*)_writer->open_writer();
  if (_error == nullptr) {
    _buffer = (char*)os::malloc(io_buffer_max_size, mtInternal);
    if (compressor != nullptr) {
      _error = (char*)_compressor->init(io_buffer_max_size, &_out_size, &_tmp_size);
      if (_error == nullptr) {
        if (_out_size > 0) {
          _out_buffer = (char*)os::malloc(_out_size, mtInternal);
        }
        if (_tmp_size > 0) {
          _tmp_buffer = (char*)os::malloc(_tmp_size, mtInternal);
        }
      }
    }
  }
  // initialize internal buffer
  _pos = 0;
  _size = io_buffer_max_size;
}

DumpWriter::~DumpWriter(){
  if (_buffer != nullptr) {
    os::free(_buffer);
  }
  if (_out_buffer != nullptr) {
    os::free(_out_buffer);
  }
  if (_tmp_buffer != nullptr) {
    os::free(_tmp_buffer);
  }
  if (_writer != nullptr) {
    delete _writer;
  }
  _bytes_written = -1;
}

// flush any buffered bytes to the file
void DumpWriter::flush() {
  if (_pos <= 0) {
    return;
  }
  if (has_error()) {
    _pos = 0;
    return;
  }
  char* result = nullptr;
  if (_compressor == nullptr) {
    result = (char*)_writer->write_buf(_buffer, _pos);
    _bytes_written += _pos;
  } else {
    do_compress();
    if (!has_error()) {
      result = (char*)_writer->write_buf(_out_buffer, _out_pos);
      _bytes_written += _out_pos;
    }
  }
  _pos = 0; // reset pos to make internal buffer available

  if (result != nullptr) {
    set_error(result);
  }
}

void DumpWriter::do_compress() {
  const char* msg = _compressor->compress(_buffer, _pos, _out_buffer, _out_size,
                                          _tmp_buffer, _tmp_size, &_out_pos);

  if (msg != nullptr) {
    set_error(msg);
  }
}

class DumperClassCacheTable;
class DumperClassCacheTableEntry;

// Support class with a collection of functions used when dumping the heap
class DumperSupport : AllStatic {
 public:

  // write a header of the given type
  static void write_header(AbstractDumpWriter* writer, hprofTag tag, u4 len);

  // returns hprof tag for the given type signature
  static hprofTag sig2tag(Symbol* sig);
  // returns hprof tag for the given basic type
  static hprofTag type2tag(BasicType type);
  // Returns the size of the data to write.
  static u4 sig2size(Symbol* sig);

  // returns the size of the instance of the given class
  static u4 instance_size(InstanceKlass* ik, DumperClassCacheTableEntry* class_cache_entry = nullptr);

  // dump a jfloat
  static void dump_float(AbstractDumpWriter* writer, jfloat f);
  // dump a jdouble
  static void dump_double(AbstractDumpWriter* writer, jdouble d);
  // dumps the raw value of the given field
  static void dump_field_value(AbstractDumpWriter* writer, char type, oop obj, int offset);
  // returns the size of the static fields; also counts the static fields
  static u4 get_static_fields_size(InstanceKlass* ik, u2& field_count);
  // dumps static fields of the given class
  static void dump_static_fields(AbstractDumpWriter* writer, Klass* k);
  // dump the raw values of the instance fields of the given object
  static void dump_instance_fields(AbstractDumpWriter* writer, oop o, DumperClassCacheTableEntry* class_cache_entry);
  // get the count of the instance fields for a given class
  static u2 get_instance_fields_count(InstanceKlass* ik);
  // dumps the definition of the instance fields for a given class
  static void dump_instance_field_descriptors(AbstractDumpWriter* writer, Klass* k);
  // creates HPROF_GC_INSTANCE_DUMP record for the given object
  static void dump_instance(AbstractDumpWriter* writer, oop o, DumperClassCacheTable* class_cache);
  // creates HPROF_GC_CLASS_DUMP record for the given instance class
  static void dump_instance_class(AbstractDumpWriter* writer, InstanceKlass* ik);
  // creates HPROF_GC_CLASS_DUMP record for a given array class
  static void dump_array_class(AbstractDumpWriter* writer, Klass* k);

  // creates HPROF_GC_OBJ_ARRAY_DUMP record for the given object array
  static void dump_object_array(AbstractDumpWriter* writer, objArrayOop array);
  // creates HPROF_GC_PRIM_ARRAY_DUMP record for the given type array
  static void dump_prim_array(AbstractDumpWriter* writer, typeArrayOop array);
  // create HPROF_FRAME record for the given method and bci
  static void dump_stack_frame(AbstractDumpWriter* writer, int frame_serial_num, int class_serial_num, Method* m, int bci);

  // check if we need to truncate an array
  static int calculate_array_max_length(AbstractDumpWriter* writer, arrayOop array, short header_size);

  // fixes up the current dump record and writes HPROF_HEAP_DUMP_END record
  static void end_of_dump(AbstractDumpWriter* writer);

  static oop mask_dormant_archived_object(oop o, oop ref_obj) {
    if (o != nullptr && o->klass()->java_mirror_no_keepalive() == nullptr) {
      // Ignore this object since the corresponding java mirror is not loaded.
      // Might be a dormant archive object.
      report_dormant_archived_object(o, ref_obj);
      return nullptr;
    } else {
      return o;
    }
  }

  static void report_dormant_archived_object(oop o, oop ref_obj) {
    if (log_is_enabled(Trace, aot, heap)) {
      ResourceMark rm;
      if (ref_obj != nullptr) {
        log_trace(aot, heap)("skipped dormant archived object " INTPTR_FORMAT " (%s) referenced by " INTPTR_FORMAT " (%s)",
                  p2i(o), o->klass()->external_name(),
                  p2i(ref_obj), ref_obj->klass()->external_name());
      } else {
        log_trace(aot, heap)("skipped dormant archived object " INTPTR_FORMAT " (%s)",
                  p2i(o), o->klass()->external_name());
      }
    }
  }
};

// Hash table of klasses to the klass metadata. This should greatly improve the
// hash dumping performance. This hash table is supposed to be used by a single
// thread only.
//
class DumperClassCacheTableEntry : public CHeapObj<mtServiceability> {
  friend class DumperClassCacheTable;
private:
  GrowableArray<char> _sigs_start;
  GrowableArray<int> _offsets;
  u4 _instance_size;
  int _entries;

public:
  DumperClassCacheTableEntry() : _instance_size(0), _entries(0) {};

  int field_count()             { return _entries; }
  char sig_start(int field_idx) { return _sigs_start.at(field_idx); }
  int offset(int field_idx)     { return _offsets.at(field_idx); }
  u4 instance_size()            { return _instance_size; }
};

class DumperClassCacheTable {
private:
  // HashTable SIZE is specified at compile time so we
  // use 1031 which is the first prime after 1024.
  static constexpr size_t TABLE_SIZE = 1031;

  // Maintain the cache for N classes. This limits memory footprint
  // impact, regardless of how many classes we have in the dump.
  // This also improves look up performance by keeping the statically
  // sized table from overloading.
  static constexpr int CACHE_TOP = 256;

  typedef HashTable<InstanceKlass*, DumperClassCacheTableEntry*,
                            TABLE_SIZE, AnyObj::C_HEAP, mtServiceability> PtrTable;
  PtrTable* _ptrs;

  // Single-slot cache to handle the major case of objects of the same
  // class back-to-back, e.g. from T[].
  InstanceKlass* _last_ik;
  DumperClassCacheTableEntry* _last_entry;

  void unlink_all(PtrTable* table) {
    class CleanupEntry: StackObj {
    public:
      bool do_entry(InstanceKlass*& key, DumperClassCacheTableEntry*& entry) {
        delete entry;
        return true;
      }
    } cleanup;
    table->unlink(&cleanup);
  }

public:
  DumperClassCacheTableEntry* lookup_or_create(InstanceKlass* ik) {
    if (_last_ik == ik) {
      return _last_entry;
    }

    DumperClassCacheTableEntry* entry;
    DumperClassCacheTableEntry** from_cache = _ptrs->get(ik);
    if (from_cache == nullptr) {
      entry = new DumperClassCacheTableEntry();
      for (HierarchicalFieldStream<JavaFieldStream> fld(ik); !fld.done(); fld.next()) {
        if (!fld.access_flags().is_static()) {
          Symbol* sig = fld.signature();
          entry->_sigs_start.push(sig->char_at(0));
          entry->_offsets.push(fld.offset());
          entry->_entries++;
          entry->_instance_size += DumperSupport::sig2size(sig);
        }
      }

      if (_ptrs->number_of_entries() >= CACHE_TOP) {
        // We do not track the individual hit rates for table entries.
        // Purge the entire table, and let the cache catch up with new
        // distribution.
        unlink_all(_ptrs);
      }

      _ptrs->put(ik, entry);
    } else {
      entry = *from_cache;
    }

    // Remember for single-slot cache.
    _last_ik = ik;
    _last_entry = entry;

    return entry;
  }

  DumperClassCacheTable() : _ptrs(new (mtServiceability) PtrTable), _last_ik(nullptr), _last_entry(nullptr) {}

  ~DumperClassCacheTable() {
    unlink_all(_ptrs);
    delete _ptrs;
  }
};

// write a header of the given type
void DumperSupport:: write_header(AbstractDumpWriter* writer, hprofTag tag, u4 len) {
  writer->write_u1(tag);
  writer->write_u4(0);                  // current ticks
  writer->write_u4(len);
}

// returns hprof tag for the given type signature
hprofTag DumperSupport::sig2tag(Symbol* sig) {
  switch (sig->char_at(0)) {
    case JVM_SIGNATURE_CLASS    : return HPROF_NORMAL_OBJECT;
    case JVM_SIGNATURE_ARRAY    : return HPROF_NORMAL_OBJECT;
    case JVM_SIGNATURE_BYTE     : return HPROF_BYTE;
    case JVM_SIGNATURE_CHAR     : return HPROF_CHAR;
    case JVM_SIGNATURE_FLOAT    : return HPROF_FLOAT;
    case JVM_SIGNATURE_DOUBLE   : return HPROF_DOUBLE;
    case JVM_SIGNATURE_INT      : return HPROF_INT;
    case JVM_SIGNATURE_LONG     : return HPROF_LONG;
    case JVM_SIGNATURE_SHORT    : return HPROF_SHORT;
    case JVM_SIGNATURE_BOOLEAN  : return HPROF_BOOLEAN;
    default : ShouldNotReachHere(); /* to shut up compiler */ return HPROF_BYTE;
  }
}

hprofTag DumperSupport::type2tag(BasicType type) {
  switch (type) {
    case T_BYTE     : return HPROF_BYTE;
    case T_CHAR     : return HPROF_CHAR;
    case T_FLOAT    : return HPROF_FLOAT;
    case T_DOUBLE   : return HPROF_DOUBLE;
    case T_INT      : return HPROF_INT;
    case T_LONG     : return HPROF_LONG;
    case T_SHORT    : return HPROF_SHORT;
    case T_BOOLEAN  : return HPROF_BOOLEAN;
    default : ShouldNotReachHere(); /* to shut up compiler */ return HPROF_BYTE;
  }
}

u4 DumperSupport::sig2size(Symbol* sig) {
  switch (sig->char_at(0)) {
    case JVM_SIGNATURE_CLASS:
    case JVM_SIGNATURE_ARRAY: return sizeof(address);
    case JVM_SIGNATURE_BOOLEAN:
    case JVM_SIGNATURE_BYTE: return 1;
    case JVM_SIGNATURE_SHORT:
    case JVM_SIGNATURE_CHAR: return 2;
    case JVM_SIGNATURE_INT:
    case JVM_SIGNATURE_FLOAT: return 4;
    case JVM_SIGNATURE_LONG:
    case JVM_SIGNATURE_DOUBLE: return 8;
    default: ShouldNotReachHere(); /* to shut up compiler */ return 0;
  }
}

template<typename T, typename F> T bit_cast(F from) { // replace with the real thing when we can use c++20
  T to;
  static_assert(sizeof(to) == sizeof(from), "must be of the same size");
  memcpy(&to, &from, sizeof(to));
  return to;
}

// dump a jfloat
void DumperSupport::dump_float(AbstractDumpWriter* writer, jfloat f) {
  if (g_isnan(f)) {
    writer->write_u4(0x7fc00000); // collapsing NaNs
  } else {
    writer->write_u4(bit_cast<u4>(f));
  }
}

// dump a jdouble
void DumperSupport::dump_double(AbstractDumpWriter* writer, jdouble d) {
  if (g_isnan(d)) {
    writer->write_u8(0x7ff80000ull << 32); // collapsing NaNs
  } else {
    writer->write_u8(bit_cast<u8>(d));
  }
}

// dumps the raw value of the given field
void DumperSupport::dump_field_value(AbstractDumpWriter* writer, char type, oop obj, int offset) {
  switch (type) {
    case JVM_SIGNATURE_CLASS :
    case JVM_SIGNATURE_ARRAY : {
      oop o = obj->obj_field_access<ON_UNKNOWN_OOP_REF | AS_NO_KEEPALIVE>(offset);
      o = mask_dormant_archived_object(o, obj);
      assert(oopDesc::is_oop_or_null(o), "Expected an oop or nullptr at " PTR_FORMAT, p2i(o));
      writer->write_objectID(o);
      break;
    }
    case JVM_SIGNATURE_BYTE : {
      jbyte b = obj->byte_field(offset);
      writer->write_u1(b);
      break;
    }
    case JVM_SIGNATURE_CHAR : {
      jchar c = obj->char_field(offset);
      writer->write_u2(c);
      break;
    }
    case JVM_SIGNATURE_SHORT : {
      jshort s = obj->short_field(offset);
      writer->write_u2(s);
      break;
    }
    case JVM_SIGNATURE_FLOAT : {
      jfloat f = obj->float_field(offset);
      dump_float(writer, f);
      break;
    }
    case JVM_SIGNATURE_DOUBLE : {
      jdouble d = obj->double_field(offset);
      dump_double(writer, d);
      break;
    }
    case JVM_SIGNATURE_INT : {
      jint i = obj->int_field(offset);
      writer->write_u4(i);
      break;
    }
    case JVM_SIGNATURE_LONG : {
      jlong l = obj->long_field(offset);
      writer->write_u8(l);
      break;
    }
    case JVM_SIGNATURE_BOOLEAN : {
      jboolean b = obj->bool_field(offset);
      writer->write_u1(b);
      break;
    }
    default : {
      ShouldNotReachHere();
      break;
    }
  }
}

// returns the size of the instance of the given class
u4 DumperSupport::instance_size(InstanceKlass* ik, DumperClassCacheTableEntry* class_cache_entry) {
  if (class_cache_entry != nullptr) {
    return class_cache_entry->instance_size();
  } else {
    u4 size = 0;
    for (HierarchicalFieldStream<JavaFieldStream> fld(ik); !fld.done(); fld.next()) {
      if (!fld.access_flags().is_static()) {
        size += sig2size(fld.signature());
      }
    }
    return size;
  }
}

u4 DumperSupport::get_static_fields_size(InstanceKlass* ik, u2& field_count) {
  field_count = 0;
  u4 size = 0;

  for (JavaFieldStream fldc(ik); !fldc.done(); fldc.next()) {
    if (fldc.access_flags().is_static()) {
      field_count++;
      size += sig2size(fldc.signature());
    }
  }

  // Add in resolved_references which is referenced by the cpCache
  // The resolved_references is an array per InstanceKlass holding the
  // strings and other oops resolved from the constant pool.
  oop resolved_references = ik->constants()->resolved_references_or_null();
  if (resolved_references != nullptr) {
    field_count++;
    size += sizeof(address);

    // Add in the resolved_references of the used previous versions of the class
    // in the case of RedefineClasses
    InstanceKlass* prev = ik->previous_versions();
    while (prev != nullptr && prev->constants()->resolved_references_or_null() != nullptr) {
      field_count++;
      size += sizeof(address);
      prev = prev->previous_versions();
    }
  }

  // Also provide a pointer to the init_lock if present, so there aren't unreferenced int[0]
  // arrays.
  oop init_lock = ik->init_lock();
  if (init_lock != nullptr) {
    field_count++;
    size += sizeof(address);
  }

  // We write the value itself plus a name and a one byte type tag per field.
  return checked_cast<u4>(size + field_count * (sizeof(address) + 1));
}

// dumps static fields of the given class
void DumperSupport::dump_static_fields(AbstractDumpWriter* writer, Klass* k) {
  InstanceKlass* ik = InstanceKlass::cast(k);

  // dump the field descriptors and raw values
  for (JavaFieldStream fld(ik); !fld.done(); fld.next()) {
    if (fld.access_flags().is_static()) {
      Symbol* sig = fld.signature();

      writer->write_symbolID(fld.name());   // name
      writer->write_u1(sig2tag(sig));       // type

      // value
      dump_field_value(writer, sig->char_at(0), ik->java_mirror(), fld.offset());
    }
  }

  // Add resolved_references for each class that has them
  oop resolved_references = ik->constants()->resolved_references_or_null();
  if (resolved_references != nullptr) {
    writer->write_symbolID(vmSymbols::resolved_references_name());  // name
    writer->write_u1(sig2tag(vmSymbols::object_array_signature())); // type
    writer->write_objectID(resolved_references);

    // Also write any previous versions
    InstanceKlass* prev = ik->previous_versions();
    while (prev != nullptr && prev->constants()->resolved_references_or_null() != nullptr) {
      writer->write_symbolID(vmSymbols::resolved_references_name());  // name
      writer->write_u1(sig2tag(vmSymbols::object_array_signature())); // type
      writer->write_objectID(prev->constants()->resolved_references());
      prev = prev->previous_versions();
    }
  }

  // Add init lock to the end if the class is not yet initialized
  oop init_lock = ik->init_lock();
  if (init_lock != nullptr) {
    writer->write_symbolID(vmSymbols::init_lock_name());         // name
    writer->write_u1(sig2tag(vmSymbols::int_array_signature())); // type
    writer->write_objectID(init_lock);
  }
}

// dump the raw values of the instance fields of the given object
void DumperSupport::dump_instance_fields(AbstractDumpWriter* writer, oop o, DumperClassCacheTableEntry* class_cache_entry) {
  assert(class_cache_entry != nullptr, "Pre-condition: must be provided");
  for (int idx = 0; idx < class_cache_entry->field_count(); idx++) {
    dump_field_value(writer, class_cache_entry->sig_start(idx), o, class_cache_entry->offset(idx));
  }
}

// dumps the definition of the instance fields for a given class
u2 DumperSupport::get_instance_fields_count(InstanceKlass* ik) {
  u2 field_count = 0;

  for (JavaFieldStream fldc(ik); !fldc.done(); fldc.next()) {
    if (!fldc.access_flags().is_static()) field_count++;
  }

  return field_count;
}

// dumps the definition of the instance fields for a given class
void DumperSupport::dump_instance_field_descriptors(AbstractDumpWriter* writer, Klass* k) {
  InstanceKlass* ik = InstanceKlass::cast(k);

  // dump the field descriptors
  for (JavaFieldStream fld(ik); !fld.done(); fld.next()) {
    if (!fld.access_flags().is_static()) {
      Symbol* sig = fld.signature();

      writer->write_symbolID(fld.name());   // name
      writer->write_u1(sig2tag(sig));       // type
    }
  }
}

// creates HPROF_GC_INSTANCE_DUMP record for the given object
void DumperSupport::dump_instance(AbstractDumpWriter* writer, oop o, DumperClassCacheTable* class_cache) {
  InstanceKlass* ik = InstanceKlass::cast(o->klass());

  DumperClassCacheTableEntry* cache_entry = class_cache->lookup_or_create(ik);

  u4 is = instance_size(ik, cache_entry);
  u4 size = 1 + sizeof(address) + 4 + sizeof(address) + 4 + is;

  writer->start_sub_record(HPROF_GC_INSTANCE_DUMP, size);
  writer->write_objectID(o);
  writer->write_u4(STACK_TRACE_ID);

  // class ID
  writer->write_classID(ik);

  // number of bytes that follow
  writer->write_u4(is);

  // field values
  dump_instance_fields(writer, o, cache_entry);

  writer->end_sub_record();
}

// creates HPROF_GC_CLASS_DUMP record for the given instance class
void DumperSupport::dump_instance_class(AbstractDumpWriter* writer, InstanceKlass* ik) {
  // We can safepoint and do a heap dump at a point where we have a Klass,
  // but no java mirror class has been setup for it. So we need to check
  // that the class is at least loaded, to avoid crash from a null mirror.
  if (!ik->is_loaded()) {
    return;
  }

  u2 static_fields_count = 0;
  u4 static_size = get_static_fields_size(ik, static_fields_count);
  u2 instance_fields_count = get_instance_fields_count(ik);
  u4 instance_fields_size = instance_fields_count * (sizeof(address) + 1);
  u4 size = checked_cast<u4>(1 + sizeof(address) + 4 + 6 * sizeof(address) + 4 + 2 + 2 + static_size + 2 + instance_fields_size);

  writer->start_sub_record(HPROF_GC_CLASS_DUMP, size);

  // class ID
  writer->write_classID(ik);
  writer->write_u4(STACK_TRACE_ID);

  // super class ID
  InstanceKlass* super = ik->super();
  if (super == nullptr) {
    writer->write_objectID(oop(nullptr));
  } else {
    writer->write_classID(super);
  }

  writer->write_objectID(ik->class_loader());
  writer->write_objectID(ik->signers());
  writer->write_objectID(ik->protection_domain());

  // reserved
  writer->write_objectID(oop(nullptr));
  writer->write_objectID(oop(nullptr));

  // instance size
  writer->write_u4(DumperSupport::instance_size(ik));

  // size of constant pool - ignored by HAT 1.1
  writer->write_u2(0);

  // static fields
  writer->write_u2(static_fields_count);
  dump_static_fields(writer, ik);

  // description of instance fields
  writer->write_u2(instance_fields_count);
  dump_instance_field_descriptors(writer, ik);

  writer->end_sub_record();
}

// creates HPROF_GC_CLASS_DUMP record for the given array class
void DumperSupport::dump_array_class(AbstractDumpWriter* writer, Klass* k) {
  InstanceKlass* ik = nullptr; // bottom class for object arrays, null for primitive type arrays
  if (k->is_objArray_klass()) {
    Klass *bk = ObjArrayKlass::cast(k)->bottom_klass();
    assert(bk != nullptr, "checking");
    if (bk->is_instance_klass()) {
      ik = InstanceKlass::cast(bk);
    }
  }

  u4 size = 1 + sizeof(address) + 4 + 6 * sizeof(address) + 4 + 2 + 2 + 2;
  writer->start_sub_record(HPROF_GC_CLASS_DUMP, size);
  writer->write_classID(k);
  writer->write_u4(STACK_TRACE_ID);

  // super class of array classes is java.lang.Object
  InstanceKlass* java_super = k->java_super();
  assert(java_super != nullptr, "checking");
  writer->write_classID(java_super);

  writer->write_objectID(ik == nullptr ? oop(nullptr) : ik->class_loader());
  writer->write_objectID(ik == nullptr ? oop(nullptr) : ik->signers());
  writer->write_objectID(ik == nullptr ? oop(nullptr) : ik->protection_domain());

  writer->write_objectID(oop(nullptr));    // reserved
  writer->write_objectID(oop(nullptr));
  writer->write_u4(0);             // instance size
  writer->write_u2(0);             // constant pool
  writer->write_u2(0);             // static fields
  writer->write_u2(0);             // instance fields

  writer->end_sub_record();

}

// Hprof uses an u4 as record length field,
// which means we need to truncate arrays that are too long.
int DumperSupport::calculate_array_max_length(AbstractDumpWriter* writer, arrayOop array, short header_size) {
  BasicType type = ArrayKlass::cast(array->klass())->element_type();
  assert(type >= T_BOOLEAN && type <= T_OBJECT, "invalid array element type");

  int length = array->length();

  int type_size;
  if (type == T_OBJECT) {
    type_size = sizeof(address);
  } else {
    type_size = type2aelembytes(type);
  }

  size_t length_in_bytes = (size_t)length * type_size;
  uint max_bytes = max_juint - header_size;

  if (length_in_bytes > max_bytes) {
    length = max_bytes / type_size;
    length_in_bytes = (size_t)length * type_size;

    warning("cannot dump array of type %s[] with length %d; truncating to length %d",
            type2name_tab[type], array->length(), length);
  }
  return length;
}

// creates HPROF_GC_OBJ_ARRAY_DUMP record for the given object array
void DumperSupport::dump_object_array(AbstractDumpWriter* writer, objArrayOop array) {
  // sizeof(u1) + 2 * sizeof(u4) + sizeof(objectID) + sizeof(classID)
  short header_size = 1 + 2 * 4 + 2 * sizeof(address);
  int length = calculate_array_max_length(writer, array, header_size);
  u4 size = checked_cast<u4>(header_size + length * sizeof(address));

  writer->start_sub_record(HPROF_GC_OBJ_ARRAY_DUMP, size);
  writer->write_objectID(array);
  writer->write_u4(STACK_TRACE_ID);
  writer->write_u4(length);

  // array class ID
  writer->write_classID(array->klass());

  // [id]* elements
  for (int index = 0; index < length; index++) {
    oop o = array->obj_at(index);
    o = mask_dormant_archived_object(o, array);
    writer->write_objectID(o);
  }

  writer->end_sub_record();
}

#define WRITE_ARRAY(Array, Type, Size, Length) \
  for (int i = 0; i < Length; i++) { writer->write_##Size((Size)Array->Type##_at(i)); }

// creates HPROF_GC_PRIM_ARRAY_DUMP record for the given type array
void DumperSupport::dump_prim_array(AbstractDumpWriter* writer, typeArrayOop array) {
  BasicType type = TypeArrayKlass::cast(array->klass())->element_type();
  // 2 * sizeof(u1) + 2 * sizeof(u4) + sizeof(objectID)
  short header_size = 2 * 1 + 2 * 4 + sizeof(address);

  int length = calculate_array_max_length(writer, array, header_size);
  int type_size = type2aelembytes(type);
  u4 length_in_bytes = (u4)length * type_size;
  u4 size = header_size + length_in_bytes;

  writer->start_sub_record(HPROF_GC_PRIM_ARRAY_DUMP, size);
  writer->write_objectID(array);
  writer->write_u4(STACK_TRACE_ID);
  writer->write_u4(length);
  writer->write_u1(type2tag(type));

  // nothing to copy
  if (length == 0) {
    writer->end_sub_record();
    return;
  }

  // If the byte ordering is big endian then we can copy most types directly

  switch (type) {
    case T_INT : {
      if (Endian::is_Java_byte_ordering_different()) {
        WRITE_ARRAY(array, int, u4, length);
      } else {
        writer->write_raw(array->int_at_addr(0), length_in_bytes);
      }
      break;
    }
    case T_BYTE : {
      writer->write_raw(array->byte_at_addr(0), length_in_bytes);
      break;
    }
    case T_CHAR : {
      if (Endian::is_Java_byte_ordering_different()) {
        WRITE_ARRAY(array, char, u2, length);
      } else {
        writer->write_raw(array->char_at_addr(0), length_in_bytes);
      }
      break;
    }
    case T_SHORT : {
      if (Endian::is_Java_byte_ordering_different()) {
        WRITE_ARRAY(array, short, u2, length);
      } else {
        writer->write_raw(array->short_at_addr(0), length_in_bytes);
      }
      break;
    }
    case T_BOOLEAN : {
      if (Endian::is_Java_byte_ordering_different()) {
        WRITE_ARRAY(array, bool, u1, length);
      } else {
        writer->write_raw(array->bool_at_addr(0), length_in_bytes);
      }
      break;
    }
    case T_LONG : {
      if (Endian::is_Java_byte_ordering_different()) {
        WRITE_ARRAY(array, long, u8, length);
      } else {
        writer->write_raw(array->long_at_addr(0), length_in_bytes);
      }
      break;
    }

    // handle float/doubles in a special value to ensure than NaNs are
    // written correctly. TO DO: Check if we can avoid this on processors that
    // use IEEE 754.

    case T_FLOAT : {
      for (int i = 0; i < length; i++) {
        dump_float(writer, array->float_at(i));
      }
      break;
    }
    case T_DOUBLE : {
      for (int i = 0; i < length; i++) {
        dump_double(writer, array->double_at(i));
      }
      break;
    }
    default : ShouldNotReachHere();
  }

  writer->end_sub_record();
}

// create a HPROF_FRAME record of the given Method* and bci
void DumperSupport::dump_stack_frame(AbstractDumpWriter* writer,
                                     int frame_serial_num,
                                     int class_serial_num,
                                     Method* m,
                                     int bci) {
  int line_number;
  if (m->is_native()) {
    line_number = -3;  // native frame
  } else {
    line_number = m->line_number_from_bci(bci);
  }

  write_header(writer, HPROF_FRAME, 4*oopSize + 2*sizeof(u4));
  writer->write_id(frame_serial_num);               // frame serial number
  writer->write_symbolID(m->name());                // method's name
  writer->write_symbolID(m->signature());           // method's signature

  assert(m->method_holder()->is_instance_klass(), "not InstanceKlass");
  writer->write_symbolID(m->method_holder()->source_file_name());  // source file name
  writer->write_u4(class_serial_num);               // class serial number
  writer->write_u4((u4) line_number);               // line number
}


// Support class used to generate HPROF_UTF8 records from the entries in the
// SymbolTable.

class SymbolTableDumper : public SymbolClosure {
 private:
  AbstractDumpWriter* _writer;
  AbstractDumpWriter* writer() const                { return _writer; }
 public:
  SymbolTableDumper(AbstractDumpWriter* writer)     { _writer = writer; }
  void do_symbol(Symbol** p);
};

void SymbolTableDumper::do_symbol(Symbol** p) {
  ResourceMark rm;
  Symbol* sym = *p;
  int len = sym->utf8_length();
  if (len > 0) {
    char* s = sym->as_utf8();
    DumperSupport::write_header(writer(), HPROF_UTF8, oopSize + len);
    writer()->write_symbolID(sym);
    writer()->write_raw(s, len);
  }
}

// Support class used to generate HPROF_GC_CLASS_DUMP records

class ClassDumper : public KlassClosure {
 private:
  AbstractDumpWriter* _writer;
  AbstractDumpWriter* writer() const { return _writer; }

 public:
  ClassDumper(AbstractDumpWriter* writer) : _writer(writer) {}

  void do_klass(Klass* k) {
    if (k->is_instance_klass()) {
      DumperSupport::dump_instance_class(writer(), InstanceKlass::cast(k));
    } else {
      DumperSupport::dump_array_class(writer(), k);
    }
  }
};

// Support class used to generate HPROF_LOAD_CLASS records

class LoadedClassDumper : public LockedClassesDo {
 private:
  AbstractDumpWriter* _writer;
  GrowableArray<Klass*>* _klass_map;
  u4 _class_serial_num;
  AbstractDumpWriter* writer() const { return _writer; }
  void add_class_serial_number(Klass* k, int serial_num) {
    _klass_map->at_put_grow(serial_num, k);
  }
 public:
  LoadedClassDumper(AbstractDumpWriter* writer, GrowableArray<Klass*>* klass_map)
    : _writer(writer), _klass_map(klass_map), _class_serial_num(0) {}

  void do_klass(Klass* k) {
    // len of HPROF_LOAD_CLASS record
    u4 remaining = 2 * oopSize + 2 * sizeof(u4);
    DumperSupport::write_header(writer(), HPROF_LOAD_CLASS, remaining);
    // class serial number is just a number
    writer()->write_u4(++_class_serial_num);
    // class ID
    writer()->write_classID(k);
    // add the Klass* and class serial number pair
    add_class_serial_number(k, _class_serial_num);
    writer()->write_u4(STACK_TRACE_ID);
    // class name ID
    Symbol* name = k->name();
    writer()->write_symbolID(name);
  }
};

// Support class used to generate HPROF_GC_ROOT_JNI_LOCAL records

class JNILocalsDumper : public OopClosure {
 private:
  AbstractDumpWriter* _writer;
  u4 _thread_serial_num;
  int _frame_num;
  AbstractDumpWriter* writer() const                { return _writer; }
 public:
  JNILocalsDumper(AbstractDumpWriter* writer, u4 thread_serial_num) {
    _writer = writer;
    _thread_serial_num = thread_serial_num;
    _frame_num = -1;  // default - empty stack
  }
  void set_frame_number(int n) { _frame_num = n; }
  void do_oop(oop* obj_p);
  void do_oop(narrowOop* obj_p) { ShouldNotReachHere(); }
};

void JNILocalsDumper::do_oop(oop* obj_p) {
  // ignore null handles
  oop o = *obj_p;
  if (o != nullptr) {
    u4 size = 1 + sizeof(address) + 4 + 4;
    writer()->start_sub_record(HPROF_GC_ROOT_JNI_LOCAL, size);
    writer()->write_objectID(o);
    writer()->write_u4(_thread_serial_num);
    writer()->write_u4((u4)_frame_num);
    writer()->end_sub_record();
  }
}


// Support class used to generate HPROF_GC_ROOT_JNI_GLOBAL records

class JNIGlobalsDumper : public OopClosure {
 private:
  AbstractDumpWriter* _writer;
  AbstractDumpWriter* writer() const                { return _writer; }

 public:
  JNIGlobalsDumper(AbstractDumpWriter* writer) {
    _writer = writer;
  }
  void do_oop(oop* obj_p);
  void do_oop(narrowOop* obj_p) { ShouldNotReachHere(); }
};

void JNIGlobalsDumper::do_oop(oop* obj_p) {
  oop o = NativeAccess<AS_NO_KEEPALIVE>::oop_load(obj_p);

  // ignore these
  if (o == nullptr) return;
  // we ignore global ref to symbols and other internal objects
  if (o->is_instance() || o->is_objArray() || o->is_typeArray()) {
    u4 size = 1 + 2 * sizeof(address);
    writer()->start_sub_record(HPROF_GC_ROOT_JNI_GLOBAL, size);
    writer()->write_objectID(o);
    writer()->write_rootID(obj_p);      // global ref ID
    writer()->end_sub_record();
  }
};

// Support class used to generate HPROF_GC_ROOT_STICKY_CLASS records

class StickyClassDumper : public KlassClosure {
 private:
  AbstractDumpWriter* _writer;
  AbstractDumpWriter* writer() const                { return _writer; }
 public:
  StickyClassDumper(AbstractDumpWriter* writer) {
    _writer = writer;
  }
  void do_klass(Klass* k) {
    if (k->is_instance_klass()) {
      InstanceKlass* ik = InstanceKlass::cast(k);
      u4 size = 1 + sizeof(address);
      writer()->start_sub_record(HPROF_GC_ROOT_STICKY_CLASS, size);
      writer()->write_classID(ik);
      writer()->end_sub_record();
    }
  }
};

// Support class used to generate HPROF_GC_ROOT_JAVA_FRAME records.

class JavaStackRefDumper : public StackObj {
private:
  AbstractDumpWriter* _writer;
  u4 _thread_serial_num;
  int _frame_num;
  AbstractDumpWriter* writer() const { return _writer; }
public:
  JavaStackRefDumper(AbstractDumpWriter* writer, u4 thread_serial_num)
      : _writer(writer), _thread_serial_num(thread_serial_num), _frame_num(-1) // default - empty stack
  {
  }

  void set_frame_number(int n) { _frame_num = n; }

  void dump_java_stack_refs(StackValueCollection* values);
};

void JavaStackRefDumper::dump_java_stack_refs(StackValueCollection* values) {
  for (int index = 0; index < values->size(); index++) {
    if (values->at(index)->type() == T_OBJECT) {
      oop o = values->obj_at(index)();
      if (o != nullptr) {
        u4 size = 1 + sizeof(address) + 4 + 4;
        writer()->start_sub_record(HPROF_GC_ROOT_JAVA_FRAME, size);
        writer()->write_objectID(o);
        writer()->write_u4(_thread_serial_num);
        writer()->write_u4((u4)_frame_num);
        writer()->end_sub_record();
      }
    }
  }
}

// Class to collect, store and dump thread-related data:
// - HPROF_TRACE and HPROF_FRAME records;
// - HPROF_GC_ROOT_THREAD_OBJ/HPROF_GC_ROOT_JAVA_FRAME/HPROF_GC_ROOT_JNI_LOCAL subrecords.
class ThreadDumper : public CHeapObj<mtInternal> {
public:
  enum class ThreadType { Platform, MountedVirtual, UnmountedVirtual };

private:
  ThreadType _thread_type;
  JavaThread* _java_thread;
  oop _thread_oop;

  GrowableArray<StackFrameInfo*>* _frames;
  // non-null if the thread is OOM thread
  Method* _oome_constructor;
  int _thread_serial_num;
  int _start_frame_serial_num;

  vframe* get_top_frame() const;

public:
  static bool should_dump_pthread(JavaThread* thread) {
    return thread->threadObj() != nullptr && !thread->is_exiting() && !thread->is_hidden_from_external_view();
  }

  static bool should_dump_vthread(oop vt) {
    return java_lang_VirtualThread::state(vt) != java_lang_VirtualThread::NEW
        && java_lang_VirtualThread::state(vt) != java_lang_VirtualThread::TERMINATED;
  }

  static bool is_vthread_mounted(oop vt) {
    // The code should be consistent with the "mounted virtual thread" case
    // (VM_HeapDumper::dump_stack_traces(), ThreadDumper::get_top_frame()).
    // I.e. virtual thread is mounted if its carrierThread is not null
    // and is_vthread_mounted() for the carrier thread returns true.
    oop carrier_thread = java_lang_VirtualThread::carrier_thread(vt);
    if (carrier_thread == nullptr) {
      return false;
    }
    JavaThread* java_thread = java_lang_Thread::thread(carrier_thread);
    return java_thread->is_vthread_mounted();
  }

  ThreadDumper(ThreadType thread_type, JavaThread* java_thread, oop thread_oop);
  ~ThreadDumper() {
    for (int index = 0; index < _frames->length(); index++) {
      delete _frames->at(index);
    }
    delete _frames;
  }

  // affects frame_count
  void add_oom_frame(Method* oome_constructor) {
    assert(_start_frame_serial_num == 0, "add_oom_frame cannot be called after init_serial_nums");
    _oome_constructor = oome_constructor;
  }

  void init_serial_nums(volatile int* thread_counter, volatile int* frame_counter) {
    assert(_start_frame_serial_num == 0, "already initialized");
    _thread_serial_num = AtomicAccess::fetch_then_add(thread_counter, 1);
    _start_frame_serial_num = AtomicAccess::fetch_then_add(frame_counter, frame_count());
  }

  bool oom_thread() const {
    return _oome_constructor != nullptr;
  }

  int frame_count() const {
    return _frames->length() + (oom_thread() ? 1 : 0);
  }

  u4 thread_serial_num() const {
    return (u4)_thread_serial_num;
  }

  u4 stack_trace_serial_num() const {
    return (u4)(_thread_serial_num + STACK_TRACE_ID);
  }

  // writes HPROF_TRACE and HPROF_FRAME records
  // returns number of dumped frames
  void dump_stack_traces(AbstractDumpWriter* writer, GrowableArray<Klass*>* klass_map);

  // writes HPROF_GC_ROOT_THREAD_OBJ subrecord
  void dump_thread_obj(AbstractDumpWriter* writer);

  // Walk the stack of the thread.
  // Dumps a HPROF_GC_ROOT_JAVA_FRAME subrecord for each local
  // Dumps a HPROF_GC_ROOT_JNI_LOCAL subrecord for each JNI local
  void dump_stack_refs(AbstractDumpWriter* writer);

};

ThreadDumper::ThreadDumper(ThreadType thread_type, JavaThread* java_thread, oop thread_oop)
    : _thread_type(thread_type), _java_thread(java_thread), _thread_oop(thread_oop),
      _oome_constructor(nullptr),
      _thread_serial_num(0), _start_frame_serial_num(0)
{
  // sanity checks
  if (_thread_type == ThreadType::UnmountedVirtual) {
    assert(_java_thread == nullptr, "sanity");
    assert(_thread_oop != nullptr, "sanity");
  } else {
    assert(_java_thread != nullptr, "sanity");
    assert(_thread_oop != nullptr, "sanity");
  }

  _frames = new (mtServiceability) GrowableArray<StackFrameInfo*>(10, mtServiceability);
  bool stop_at_vthread_entry = _thread_type == ThreadType::MountedVirtual;

  // vframes are resource allocated
  Thread* current_thread = Thread::current();
  ResourceMark rm(current_thread);
  HandleMark hm(current_thread);

  for (vframe* vf = get_top_frame(); vf != nullptr; vf = vf->sender()) {
    if (stop_at_vthread_entry && vf->is_vthread_entry()) {
      break;
    }
    if (vf->is_java_frame()) {
      javaVFrame* jvf = javaVFrame::cast(vf);
      _frames->append(new StackFrameInfo(jvf, false));
    } else {
      // ignore non-Java frames
    }
  }
}

void ThreadDumper::dump_stack_traces(AbstractDumpWriter* writer, GrowableArray<Klass*>* klass_map) {
  assert(_thread_serial_num != 0 && _start_frame_serial_num != 0, "serial_nums are not initialized");

  // write HPROF_FRAME records for this thread's stack trace
  int depth = _frames->length();
  int frame_serial_num = _start_frame_serial_num;

  if (oom_thread()) {
    // OOM thread
    // write fake frame that makes it look like the thread, which caused OOME,
    // is in the OutOfMemoryError zero-parameter constructor
    int oome_serial_num = klass_map->find(_oome_constructor->method_holder());
    // the class serial number starts from 1
    assert(oome_serial_num > 0, "OutOfMemoryError class not found");
    DumperSupport::dump_stack_frame(writer, ++frame_serial_num, oome_serial_num, _oome_constructor, 0);
    depth++;
  }

  for (int j = 0; j < _frames->length(); j++) {
    StackFrameInfo* frame = _frames->at(j);
    Method* m = frame->method();
    int class_serial_num = klass_map->find(m->method_holder());
    // the class serial number starts from 1
    assert(class_serial_num > 0, "class not found");
    DumperSupport::dump_stack_frame(writer, ++frame_serial_num, class_serial_num, m, frame->bci());
  }

  // write HPROF_TRACE record for the thread
  DumperSupport::write_header(writer, HPROF_TRACE, checked_cast<u4>(3 * sizeof(u4) + depth * oopSize));
  writer->write_u4(stack_trace_serial_num());   // stack trace serial number
  writer->write_u4(thread_serial_num());        // thread serial number
  writer->write_u4((u4)depth);                  // frame count (including oom frame)
  for (int j = 1; j <= depth; j++) {
    writer->write_id(_start_frame_serial_num + j);
  }
}

void ThreadDumper::dump_thread_obj(AbstractDumpWriter * writer) {
  assert(_thread_serial_num != 0 && _start_frame_serial_num != 0, "serial_num is not initialized");

  u4 size = 1 + sizeof(address) + 4 + 4;
  writer->start_sub_record(HPROF_GC_ROOT_THREAD_OBJ, size);
  writer->write_objectID(_thread_oop);
  writer->write_u4(thread_serial_num());      // thread serial number
  writer->write_u4(stack_trace_serial_num()); // stack trace serial number
  writer->end_sub_record();
}

void ThreadDumper::dump_stack_refs(AbstractDumpWriter * writer) {
  assert(_thread_serial_num != 0 && _start_frame_serial_num != 0, "serial_num is not initialized");

  JNILocalsDumper blk(writer, thread_serial_num());
  if (_thread_type == ThreadType::Platform) {
    if (!_java_thread->has_last_Java_frame()) {
      // no last java frame but there may be JNI locals
      _java_thread->active_handles()->oops_do(&blk);
      return;
    }
  }

  JavaStackRefDumper java_ref_dumper(writer, thread_serial_num());

  // vframes are resource allocated
  Thread* current_thread = Thread::current();
  ResourceMark rm(current_thread);
  HandleMark hm(current_thread);

  bool stopAtVthreadEntry = _thread_type == ThreadType::MountedVirtual;
  frame* last_entry_frame = nullptr;
  bool is_top_frame = true;
  int depth = 0;
  if (oom_thread()) {
    depth++;
  }

  for (vframe* vf = get_top_frame(); vf != nullptr; vf = vf->sender()) {
    if (stopAtVthreadEntry && vf->is_vthread_entry()) {
      break;
    }

    if (vf->is_java_frame()) {
      javaVFrame* jvf = javaVFrame::cast(vf);
      if (!(jvf->method()->is_native())) {
        java_ref_dumper.set_frame_number(depth);
        java_ref_dumper.dump_java_stack_refs(jvf->locals());
        java_ref_dumper.dump_java_stack_refs(jvf->expressions());
      } else {
        // native frame
        blk.set_frame_number(depth);
        if (is_top_frame) {
          // JNI locals for the top frame if mounted
          assert(_java_thread != nullptr || jvf->method()->is_synchronized()
                 || jvf->method()->is_object_wait0(), "impossible for unmounted vthread");
          if (_java_thread != nullptr) {
            _java_thread->active_handles()->oops_do(&blk);
          }
        } else {
          if (last_entry_frame != nullptr) {
            // JNI locals for the entry frame
            assert(last_entry_frame->is_entry_frame(), "checking");
            last_entry_frame->entry_frame_call_wrapper()->handles()->oops_do(&blk);
          }
        }
      }
      last_entry_frame = nullptr;
      // increment only for Java frames
      depth++;
    } else {
      // externalVFrame - for an entry frame then we report the JNI locals
      // when we find the corresponding javaVFrame
      frame* fr = vf->frame_pointer();
      assert(fr != nullptr, "sanity check");
      if (fr->is_entry_frame()) {
        last_entry_frame = fr;
      }
    }
  is_top_frame = false;
  }
  assert(depth == frame_count(), "total number of Java frames not matched");
}

vframe* ThreadDumper::get_top_frame() const {
  if (_thread_type == ThreadType::UnmountedVirtual) {
    ContinuationWrapper cont(java_lang_VirtualThread::continuation(_thread_oop));
    if (cont.is_empty()) {
      return nullptr;
    }
    assert(!cont.is_mounted(), "sanity check");
    stackChunkOop chunk = cont.last_nonempty_chunk();
    if (chunk == nullptr || chunk->is_empty()) {
      return nullptr;
    }

    RegisterMap reg_map(cont.continuation(), RegisterMap::UpdateMap::include);
    frame fr = chunk->top_frame(&reg_map);
    vframe* vf = vframe::new_vframe(&fr, &reg_map, nullptr); // don't need JavaThread
    return vf;
  }

  RegisterMap reg_map(_java_thread,
      RegisterMap::UpdateMap::include,
      RegisterMap::ProcessFrames::include,
      RegisterMap::WalkContinuation::skip);
  switch (_thread_type) {
  case ThreadType::Platform:
    if (!_java_thread->has_last_Java_frame()) {
      return nullptr;
    }
    return _java_thread->is_vthread_mounted()
        ? _java_thread->carrier_last_java_vframe(&reg_map)
        : _java_thread->platform_thread_last_java_vframe(&reg_map);

  case ThreadType::MountedVirtual:
    return _java_thread->last_java_vframe(&reg_map);

  default: // make compilers happy
      break;
  }
  ShouldNotReachHere();
  return nullptr;
}

// Callback to dump thread-related data for unmounted virtual threads;
// implemented by VM_HeapDumper.
class UnmountedVThreadDumper {
 public:
  virtual void dump_vthread(oop vt, AbstractDumpWriter* segment_writer) = 0;
};

// Support class used when iterating over the heap.
class HeapObjectDumper : public ObjectClosure {
 private:
  AbstractDumpWriter* _writer;
  AbstractDumpWriter* writer()                  { return _writer; }
  UnmountedVThreadDumper* _vthread_dumper;

  DumperClassCacheTable _class_cache;

 public:
  HeapObjectDumper(AbstractDumpWriter* writer, UnmountedVThreadDumper* vthread_dumper)
    : _writer(writer), _vthread_dumper(vthread_dumper) {}

  // called for each object in the heap
  void do_object(oop o);
};

void HeapObjectDumper::do_object(oop o) {
  // skip classes as these emitted as HPROF_GC_CLASS_DUMP records
  if (o->klass() == vmClasses::Class_klass()) {
    if (!java_lang_Class::is_primitive(o)) {
      return;
    }
  }

  if (DumperSupport::mask_dormant_archived_object(o, nullptr) == nullptr) {
    return;
  }

  if (o->is_instance()) {
    // create a HPROF_GC_INSTANCE record for each object
    DumperSupport::dump_instance(writer(), o, &_class_cache);
    // If we encounter an unmounted virtual thread it needs to be dumped explicitly
    // (mounted virtual threads are dumped with their carriers).
    if (java_lang_VirtualThread::is_instance(o)
        && ThreadDumper::should_dump_vthread(o) && !ThreadDumper::is_vthread_mounted(o)) {
      _vthread_dumper->dump_vthread(o, writer());
    }
  } else if (o->is_objArray()) {
    // create a HPROF_GC_OBJ_ARRAY_DUMP record for each object array
    DumperSupport::dump_object_array(writer(), objArrayOop(o));
  } else if (o->is_typeArray()) {
    // create a HPROF_GC_PRIM_ARRAY_DUMP record for each type array
    DumperSupport::dump_prim_array(writer(), typeArrayOop(o));
  }
}

// The dumper controller for parallel heap dump
class DumperController : public CHeapObj<mtInternal> {
 private:
   Monitor* _lock;
   Mutex* _global_writer_lock;

   const uint   _dumper_number;
   uint   _complete_number;

   bool   _started; // VM dumper started and acquired global writer lock

 public:
   DumperController(uint number) :
     // _lock and _global_writer_lock are used for synchronization between GC worker threads inside safepoint,
     // so we lock with _no_safepoint_check_flag.
     // signal_start() acquires _lock when global writer is locked,
     // its rank must be less than _global_writer_lock rank.
     _lock(new (std::nothrow) PaddedMonitor(Mutex::nosafepoint - 1, "DumperController_lock")),
     _global_writer_lock(new (std::nothrow) Mutex(Mutex::nosafepoint, "DumpWriter_lock")),
     _dumper_number(number),
     _complete_number(0),
     _started(false)
   {}

   ~DumperController() {
     delete _lock;
     delete _global_writer_lock;
   }

   // parallel (non VM) dumpers must wait until VM dumper acquires global writer lock
   void wait_for_start_signal() {
     MonitorLocker ml(_lock, Mutex::_no_safepoint_check_flag);
     while (_started == false) {
       ml.wait();
     }
   }

   void signal_start() {
     MonitorLocker ml(_lock, Mutex::_no_safepoint_check_flag);
     _started = true;
     ml.notify_all();
   }

   void lock_global_writer() {
     _global_writer_lock->lock_without_safepoint_check();
   }

   void unlock_global_writer() {
     _global_writer_lock->unlock();
   }

   void dumper_complete(DumpWriter* local_writer, DumpWriter* global_writer) {
     MonitorLocker ml(_lock, Mutex::_no_safepoint_check_flag);
     _complete_number++;
     // propagate local error to global if any
     if (local_writer->has_error()) {
       global_writer->set_error(local_writer->error());
     }
     ml.notify();
   }

   void wait_all_dumpers_complete() {
     MonitorLocker ml(_lock, Mutex::_no_safepoint_check_flag);
     while (_complete_number != _dumper_number) {
        ml.wait();
     }
   }
};

// DumpMerger merges separate dump files into a complete one
class DumpMerger : public StackObj {
private:
  DumpWriter* _writer;
  const char* _path;
  bool _has_error;
  int _dump_seq;

private:
  void merge_file(const char* path);
  void merge_done();
  void set_error(const char* msg);

public:
  DumpMerger(const char* path, DumpWriter* writer, int dump_seq) :
    _writer(writer),
    _path(path),
    _has_error(_writer->has_error()),
    _dump_seq(dump_seq) {}

  void do_merge();

  // returns path for the parallel DumpWriter (resource allocated)
  static char* get_writer_path(const char* base_path, int seq);

};

char* DumpMerger::get_writer_path(const char* base_path, int seq) {
  // approximate required buffer size
  size_t buf_size = strlen(base_path)
                    + 2                 // ".p"
                    + 10                // number (that's enough for 2^32 parallel dumpers)
                    + 1;                // '\0'

  char* path = NEW_RESOURCE_ARRAY(char, buf_size);
  memset(path, 0, buf_size);

  os::snprintf_checked(path, buf_size, "%s.p%d", base_path, seq);

  return path;
}


void DumpMerger::merge_done() {
  // Writes the HPROF_HEAP_DUMP_END record.
  if (!_has_error) {
    DumperSupport::end_of_dump(_writer);
    _writer->flush();
  }
  _dump_seq = 0; //reset
}

void DumpMerger::set_error(const char* msg) {
  assert(msg != nullptr, "sanity check");
  log_error(heapdump)("%s (file: %s)", msg, _path);
  _writer->set_error(msg);
  _has_error = true;
}

#ifdef LINUX
// Merge segmented heap files via sendfile, it's more efficient than the
// read+write combination, which would require transferring data to and from
// user space.
void DumpMerger::merge_file(const char* path) {
  TraceTime timer("Merge segmented heap file directly", TRACETIME_LOG(Info, heapdump));

  int segment_fd = os::open(path, O_RDONLY, 0);
  if (segment_fd == -1) {
    set_error("Can not open segmented heap file during merging");
    return;
  }

  struct stat st;
  if (os::stat(path, &st) != 0) {
    ::close(segment_fd);
    set_error("Can not get segmented heap file size during merging");
    return;
  }

  // A successful call to sendfile may write fewer bytes than requested; the
  // caller should be prepared to retry the call if there were unsent bytes.
  jlong offset = 0;
  while (offset < st.st_size) {
    int ret = os::Linux::sendfile(_writer->get_fd(), segment_fd, &offset, st.st_size);
    if (ret == -1) {
      ::close(segment_fd);
      set_error("Failed to merge segmented heap file");
      return;
    }
  }

  // As sendfile variant does not call the write method of the global writer,
  // bytes_written is also incorrect for this variant, we need to explicitly
  // accumulate bytes_written for the global writer in this case
  julong accum = _writer->bytes_written() + st.st_size;
  _writer->set_bytes_written(accum);
  ::close(segment_fd);
}
#else
// Generic implementation using read+write
void DumpMerger::merge_file(const char* path) {
  TraceTime timer("Merge segmented heap file", TRACETIME_LOG(Info, heapdump));

  fileStream segment_fs(path, "rb");
  if (!segment_fs.is_open()) {
    set_error("Can not open segmented heap file during merging");
    return;
  }

  jlong total = 0;
  size_t cnt = 0;

  // Use _writer buffer for reading.
  while ((cnt = segment_fs.read(_writer->buffer(), 1, _writer->buffer_size())) != 0) {
    _writer->set_position(cnt);
    _writer->flush();
    total += cnt;
  }

  if (segment_fs.fileSize() != total) {
    set_error("Merged heap dump is incomplete");
  }
}
#endif

void DumpMerger::do_merge() {
  TraceTime timer("Merge heap files complete", TRACETIME_LOG(Info, heapdump));

  // Since contents in segmented heap file were already zipped, we don't need to zip
  // them again during merging.
  AbstractCompressor* saved_compressor = _writer->compressor();
  _writer->set_compressor(nullptr);

  // Merge the content of the remaining files into base file. Regardless of whether
  // the merge process is successful or not, these segmented files will be deleted.
  for (int i = 0; i < _dump_seq; i++) {
    ResourceMark rm;
    const char* path = get_writer_path(_path, i);
    if (!_has_error) {
      merge_file(path);
    }
    // Delete selected segmented heap file nevertheless
    if (remove(path) != 0) {
      log_info(heapdump)("Removal of segment file (%d) failed (%d)", i, errno);
    }
  }

  // restore compressor for further use
  _writer->set_compressor(saved_compressor);
  merge_done();
}

// The VM operation that performs the heap dump
class VM_HeapDumper : public VM_GC_Operation, public WorkerTask, public UnmountedVThreadDumper {
 private:
  DumpWriter*             _writer;
  JavaThread*             _oome_thread;
  Method*                 _oome_constructor;
  bool                    _gc_before_heap_dump;
  GrowableArray<Klass*>*  _klass_map;

  ThreadDumper**          _thread_dumpers; // platform, carrier and mounted virtual threads
  int                     _thread_dumpers_count;
  volatile int            _thread_serial_num;
  volatile int            _frame_serial_num;

  volatile int            _dump_seq;
  // parallel heap dump support
  uint                    _num_dumper_threads;
  DumperController*       _dumper_controller;
  ParallelObjectIterator* _poi;

  // Dumper id of VMDumper thread.
  static const int VMDumperId = 0;
  // VM dumper dumps both heap and non-heap data, other dumpers dump heap-only data.
  static bool is_vm_dumper(int dumper_id) { return dumper_id == VMDumperId; }
  // the 1st dumper calling get_next_dumper_id becomes VM dumper
  int get_next_dumper_id() {
    return AtomicAccess::fetch_then_add(&_dump_seq, 1);
  }

  DumpWriter* writer() const { return _writer; }

  bool skip_operation() const;

  // HPROF_GC_ROOT_THREAD_OBJ records for platform and mounted virtual threads
  void dump_threads(AbstractDumpWriter* writer);

  bool is_oom_thread(JavaThread* thread) const {
    return thread == _oome_thread && _oome_constructor != nullptr;
  }

  // HPROF_TRACE and HPROF_FRAME records for platform and mounted virtual threads
  void dump_stack_traces(AbstractDumpWriter* writer);

 public:
  VM_HeapDumper(DumpWriter* writer, bool gc_before_heap_dump, bool oome, uint num_dump_threads) :
    VM_GC_Operation(0 /* total collections,      dummy, ignored */,
                    GCCause::_heap_dump /* GC Cause */,
                    0 /* total full collections, dummy, ignored */,
                    gc_before_heap_dump),
    WorkerTask("dump heap") {
    _writer = writer;
    _gc_before_heap_dump = gc_before_heap_dump;
    _klass_map = new (mtServiceability) GrowableArray<Klass*>(INITIAL_CLASS_COUNT, mtServiceability);

    _thread_dumpers = nullptr;
    _thread_dumpers_count = 0;
    _thread_serial_num = 1;
    _frame_serial_num = 1;

    _dump_seq = VMDumperId;
    _num_dumper_threads = num_dump_threads;
    _dumper_controller = nullptr;
    _poi = nullptr;
    if (oome) {
      assert(!Thread::current()->is_VM_thread(), "Dump from OutOfMemoryError cannot be called by the VMThread");
      // get OutOfMemoryError zero-parameter constructor
      InstanceKlass* oome_ik = vmClasses::OutOfMemoryError_klass();
      _oome_constructor = oome_ik->find_method(vmSymbols::object_initializer_name(),
                                                          vmSymbols::void_method_signature());
      // get thread throwing OOME when generating the heap dump at OOME
      _oome_thread = JavaThread::current();
    } else {
      _oome_thread = nullptr;
      _oome_constructor = nullptr;
    }
  }

  ~VM_HeapDumper() {
    if (_thread_dumpers != nullptr) {
      for (int i = 0; i < _thread_dumpers_count; i++) {
        delete _thread_dumpers[i];
      }
      FREE_C_HEAP_ARRAY(ThreadDumper*, _thread_dumpers);
    }

    if (_dumper_controller != nullptr) {
      delete _dumper_controller;
      _dumper_controller = nullptr;
    }
    delete _klass_map;
  }
  int dump_seq()           { return _dump_seq; }
  bool is_parallel_dump()  { return _num_dumper_threads > 1; }
  void prepare_parallel_dump(WorkerThreads* workers);

  VMOp_Type type() const { return VMOp_HeapDumper; }
  virtual bool doit_prologue();
  void doit();
  void work(uint worker_id);

  // UnmountedVThreadDumper implementation
  void dump_vthread(oop vt, AbstractDumpWriter* segment_writer);
};

bool VM_HeapDumper::skip_operation() const {
  return false;
}

// fixes up the current dump record and writes HPROF_HEAP_DUMP_END record
void DumperSupport::end_of_dump(AbstractDumpWriter* writer) {
  writer->finish_dump_segment();

  writer->write_u1(HPROF_HEAP_DUMP_END);
  writer->write_u4(0);
  writer->write_u4(0);
}

// Write a HPROF_GC_ROOT_THREAD_OBJ record for platform/carrier and mounted virtual threads.
// Then walk the stack so that locals and JNI locals are dumped.
void VM_HeapDumper::dump_threads(AbstractDumpWriter* writer) {
  for (int i = 0; i < _thread_dumpers_count; i++) {
    _thread_dumpers[i]->dump_thread_obj(writer);
    _thread_dumpers[i]->dump_stack_refs(writer);
  }
}

bool VM_HeapDumper::doit_prologue() {
  if (_gc_before_heap_dump && (UseZGC || UseShenandoahGC)) {
    // ZGC and Shenandoah cannot perform a synchronous GC cycle from within the VM thread.
    // So collect_as_vm_thread() is a noop. To respect the _gc_before_heap_dump flag a
    // synchronous GC cycle is performed from the caller thread in the prologue.
    Universe::heap()->collect(GCCause::_heap_dump);
  }
  return VM_GC_Operation::doit_prologue();
}

void VM_HeapDumper::prepare_parallel_dump(WorkerThreads* workers) {
  uint num_active_workers = workers != nullptr ? workers->active_workers() : 0;
  uint num_requested_dump_threads = _num_dumper_threads;
  // check if we can dump in parallel based on requested and active threads
  if (num_active_workers <= 1 || num_requested_dump_threads <= 1) {
    _num_dumper_threads = 1;
  } else {
    _num_dumper_threads = clamp(num_requested_dump_threads, 2U, num_active_workers);
  }
  _dumper_controller = new (std::nothrow) DumperController(_num_dumper_threads);
  bool can_parallel = _num_dumper_threads > 1;
  log_info(heapdump)("Requested dump threads %u, active dump threads %u, "
                     "actual dump threads %u, parallelism %s",
                     num_requested_dump_threads, num_active_workers,
                     _num_dumper_threads, can_parallel ? "true" : "false");
}

// The VM operation that dumps the heap. The dump consists of the following
// records:
//
//  HPROF_HEADER
//  [HPROF_UTF8]*
//  [HPROF_LOAD_CLASS]*
//  [[HPROF_FRAME]*|HPROF_TRACE]*
//  [HPROF_GC_CLASS_DUMP]*
//  [HPROF_HEAP_DUMP_SEGMENT]*
//  HPROF_HEAP_DUMP_END
//
// The HPROF_TRACE records represent the stack traces where the heap dump
// is generated and a "dummy trace" record which does not include
// any frames. The dummy trace record is used to be referenced as the
// unknown object alloc site.
//
// Each HPROF_HEAP_DUMP_SEGMENT record has a length followed by sub-records.
// To allow the heap dump be generated in a single pass we remember the position
// of the dump length and fix it up after all sub-records have been written.
// To generate the sub-records we iterate over the heap, writing
// HPROF_GC_INSTANCE_DUMP, HPROF_GC_OBJ_ARRAY_DUMP, and HPROF_GC_PRIM_ARRAY_DUMP
// records as we go. Once that is done we write records for some of the GC
// roots.

void VM_HeapDumper::doit() {

  CollectedHeap* ch = Universe::heap();

  ch->ensure_parsability(false); // must happen, even if collection does
                                 // not happen (e.g. due to GCLocker)

  if (_gc_before_heap_dump) {
    if (GCLocker::is_active()) {
      warning("GC locker is held; pre-heapdump GC was skipped");
    } else {
      ch->collect_as_vm_thread(GCCause::_heap_dump);
    }
  }

  WorkerThreads* workers = ch->safepoint_workers();
  prepare_parallel_dump(workers);

  if (!is_parallel_dump()) {
    work(VMDumperId);
  } else {
    ParallelObjectIterator poi(_num_dumper_threads);
    _poi = &poi;
    workers->run_task(this, _num_dumper_threads);
    _poi = nullptr;
  }
}

void VM_HeapDumper::work(uint worker_id) {
  // VM Dumper works on all non-heap data dumping and part of heap iteration.
  int dumper_id = get_next_dumper_id();

  if (is_vm_dumper(dumper_id)) {
    // lock global writer, it will be unlocked after VM Dumper finishes with non-heap data
    _dumper_controller->lock_global_writer();
    _dumper_controller->signal_start();
  } else {
    _dumper_controller->wait_for_start_signal();
  }

  if (is_vm_dumper(dumper_id)) {
    TraceTime timer("Dump non-objects", TRACETIME_LOG(Info, heapdump));
    // Write the file header - we always use 1.0.2
    const char* header = "JAVA PROFILE 1.0.2";

    // header is few bytes long - no chance to overflow int
    writer()->write_raw(header, strlen(header) + 1); // NUL terminated
    writer()->write_u4(oopSize);
    // timestamp is current time in ms
    writer()->write_u8(os::javaTimeMillis());
    // HPROF_UTF8 records
    SymbolTableDumper sym_dumper(writer());
    SymbolTable::symbols_do(&sym_dumper);

    // write HPROF_LOAD_CLASS records
    {
      LoadedClassDumper loaded_class_dumper(writer(), _klass_map);
      ClassLoaderDataGraph::classes_do(&loaded_class_dumper);
    }

    // write HPROF_FRAME and HPROF_TRACE records
    // this must be called after _klass_map is built when iterating the classes above.
    dump_stack_traces(writer());

    // unlock global writer, so parallel dumpers can dump stack traces of unmounted virtual threads
    _dumper_controller->unlock_global_writer();
  }

  // HPROF_HEAP_DUMP/HPROF_HEAP_DUMP_SEGMENT starts here

  ResourceMark rm;
  // share global compressor, local DumpWriter is not responsible for its life cycle
  DumpWriter segment_writer(DumpMerger::get_writer_path(writer()->get_file_path(), dumper_id),
                            writer()->is_overwrite(), writer()->compressor());
  if (!segment_writer.has_error()) {
    if (is_vm_dumper(dumper_id)) {
      // dump some non-heap subrecords to heap dump segment
      TraceTime timer("Dump non-objects (part 2)", TRACETIME_LOG(Info, heapdump));
      // Writes HPROF_GC_CLASS_DUMP records
      ClassDumper class_dumper(&segment_writer);
      ClassLoaderDataGraph::classes_do(&class_dumper);

      // HPROF_GC_ROOT_THREAD_OBJ + frames + jni locals
      dump_threads(&segment_writer);

      // HPROF_GC_ROOT_JNI_GLOBAL
      JNIGlobalsDumper jni_dumper(&segment_writer);
      JNIHandles::oops_do(&jni_dumper);
      // technically not jni roots, but global roots
      // for things like preallocated throwable backtraces
      Universe::vm_global()->oops_do(&jni_dumper);
      // HPROF_GC_ROOT_STICKY_CLASS
      // These should be classes in the null class loader data, and not all classes
      // if !ClassUnloading
      StickyClassDumper stiky_class_dumper(&segment_writer);
      ClassLoaderData::the_null_class_loader_data()->classes_do(&stiky_class_dumper);
    }

    // Heap iteration.
    // writes HPROF_GC_INSTANCE_DUMP records.
    // After each sub-record is written check_segment_length will be invoked
    // to check if the current segment exceeds a threshold. If so, a new
    // segment is started.
    // The HPROF_GC_CLASS_DUMP and HPROF_GC_INSTANCE_DUMP are the vast bulk
    // of the heap dump.

    TraceTime timer(is_parallel_dump() ? "Dump heap objects in parallel" : "Dump heap objects", TRACETIME_LOG(Info, heapdump));
    HeapObjectDumper obj_dumper(&segment_writer, this);
    if (!is_parallel_dump()) {
      Universe::heap()->object_iterate(&obj_dumper);
    } else {
      // == Parallel dump
      _poi->object_iterate(&obj_dumper, worker_id);
    }

    segment_writer.finish_dump_segment();
    segment_writer.flush();
  }

  _dumper_controller->dumper_complete(&segment_writer, writer());

  if (is_vm_dumper(dumper_id)) {
    _dumper_controller->wait_all_dumpers_complete();

    // flush global writer
    writer()->flush();

    // At this point, all fragments of the heapdump have been written to separate files.
    // We need to merge them into a complete heapdump and write HPROF_HEAP_DUMP_END at that time.
  }
}

void VM_HeapDumper::dump_stack_traces(AbstractDumpWriter* writer) {
  // write a HPROF_TRACE record without any frames to be referenced as object alloc sites
  DumperSupport::write_header(writer, HPROF_TRACE, 3 * sizeof(u4));
  writer->write_u4((u4)STACK_TRACE_ID);
  writer->write_u4(0);                    // thread number
  writer->write_u4(0);                    // frame count

  // max number if every platform thread is carrier with mounted virtual thread
  _thread_dumpers = NEW_C_HEAP_ARRAY(ThreadDumper*, Threads::number_of_threads() * 2, mtInternal);

  for (JavaThreadIteratorWithHandle jtiwh; JavaThread * thread = jtiwh.next(); ) {
    if (ThreadDumper::should_dump_pthread(thread)) {
      bool add_oom_frame = is_oom_thread(thread);

      oop mounted_vt = thread->is_vthread_mounted() ? thread->vthread() : nullptr;
      if (mounted_vt != nullptr && !ThreadDumper::should_dump_vthread(mounted_vt)) {
        mounted_vt = nullptr;
      }

      // mounted vthread (if any)
      if (mounted_vt != nullptr) {
        ThreadDumper* thread_dumper = new ThreadDumper(ThreadDumper::ThreadType::MountedVirtual, thread, mounted_vt);
        _thread_dumpers[_thread_dumpers_count++] = thread_dumper;
        if (add_oom_frame) {
          thread_dumper->add_oom_frame(_oome_constructor);
          // we add oom frame to the VT stack, don't add it to the carrier thread stack
          add_oom_frame = false;
        }
        thread_dumper->init_serial_nums(&_thread_serial_num, &_frame_serial_num);
        thread_dumper->dump_stack_traces(writer, _klass_map);
      }

      // platform or carrier thread
      ThreadDumper* thread_dumper = new ThreadDumper(ThreadDumper::ThreadType::Platform, thread, thread->threadObj());
      _thread_dumpers[_thread_dumpers_count++] = thread_dumper;
      if (add_oom_frame) {
        thread_dumper->add_oom_frame(_oome_constructor);
      }
      thread_dumper->init_serial_nums(&_thread_serial_num, &_frame_serial_num);
      thread_dumper->dump_stack_traces(writer, _klass_map);
    }
  }
}

void VM_HeapDumper::dump_vthread(oop vt, AbstractDumpWriter* segment_writer) {
  // unmounted vthread has no JavaThread
  ThreadDumper thread_dumper(ThreadDumper::ThreadType::UnmountedVirtual, nullptr, vt);
  thread_dumper.init_serial_nums(&_thread_serial_num, &_frame_serial_num);

  // write HPROF_TRACE/HPROF_FRAME records to global writer
  _dumper_controller->lock_global_writer();
  thread_dumper.dump_stack_traces(writer(), _klass_map);
  _dumper_controller->unlock_global_writer();

  // write HPROF_GC_ROOT_THREAD_OBJ/HPROF_GC_ROOT_JAVA_FRAME/HPROF_GC_ROOT_JNI_LOCAL subrecord
  // to segment writer
  thread_dumper.dump_thread_obj(segment_writer);
  thread_dumper.dump_stack_refs(segment_writer);
}

// dump the heap to given path.
int HeapDumper::dump(const char* path, outputStream* out, int compression, bool overwrite, uint num_dump_threads) {
  assert(path != nullptr && strlen(path) > 0, "path missing");

  // print message in interactive case
  if (out != nullptr) {
    out->print_cr("Dumping heap to %s ...", path);
    timer()->start();
  }

  if (_oome && num_dump_threads > 1) {
    // Each additional parallel writer requires several MB of internal memory
    // (DumpWriter buffer, DumperClassCacheTable, GZipCompressor buffers).
    // For the OOM handling we may already be limited in memory.
    // Lets ensure we have at least 20MB per thread.
    physical_memory_size_type free_memory = 0;
    // Return value ignored - defaulting to 0 on failure.
    (void)os::free_memory(free_memory);
    julong max_threads = free_memory / (20 * M);
    if (num_dump_threads > max_threads) {
      num_dump_threads = MAX2<uint>(1, (uint)max_threads);
    }
  }

  // create JFR event
  EventHeapDump event;

  AbstractCompressor* compressor = nullptr;

  if (compression > 0) {
    compressor = new (std::nothrow) GZipCompressor(compression);

    if (compressor == nullptr) {
      set_error("Could not allocate gzip compressor");
      return -1;
    }
  }

  DumpWriter writer(path, overwrite, compressor);

  if (writer.error() != nullptr) {
    set_error(writer.error());
    if (out != nullptr) {
      out->print_cr("Unable to create %s: %s", path,
        (error() != nullptr) ? error() : "reason unknown");
    }
    return -1;
  }

  // generate the segmented heap dump into separate files
  VM_HeapDumper dumper(&writer, _gc_before_heap_dump, _oome, num_dump_threads);
  VMThread::execute(&dumper);

  // record any error that the writer may have encountered
  set_error(writer.error());

  // Heap dump process is done in two phases
  //
  // Phase 1: Concurrent threads directly write heap data to multiple heap files.
  //          This is done by VM_HeapDumper, which is performed within safepoint.
  //
  // Phase 2: Merge multiple heap files into one complete heap dump file.
  //          This is done by DumpMerger, which is performed outside safepoint

  DumpMerger merger(path, &writer, dumper.dump_seq());
  // Perform heapdump file merge operation in the current thread prevents us
  // from occupying the VM Thread, which in turn affects the occurrence of
  // GC and other VM operations.
  merger.do_merge();
  if (writer.error() != nullptr) {
    set_error(writer.error());
  }

  // emit JFR event
  if (error() == nullptr) {
    event.set_destination(path);
    event.set_gcBeforeDump(_gc_before_heap_dump);
    event.set_size(writer.bytes_written());
    event.set_onOutOfMemoryError(_oome);
    event.set_overwrite(overwrite);
    event.set_compression(compression);
    event.commit();
  } else {
    log_debug(aot, heap)("Error %s while dumping heap", error());
  }

  // print message in interactive case
  if (out != nullptr) {
    timer()->stop();
    if (error() == nullptr) {
      out->print_cr("Heap dump file created [" JULONG_FORMAT " bytes in %3.3f secs]",
                    writer.bytes_written(), timer()->seconds());
    } else {
      out->print_cr("Dump file is incomplete: %s", writer.error());
    }
  }

  if (compressor != nullptr) {
    delete compressor;
  }
  return (writer.error() == nullptr) ? 0 : -1;
}

// stop timer (if still active), and free any error string we might be holding
HeapDumper::~HeapDumper() {
  if (timer()->is_active()) {
    timer()->stop();
  }
  set_error(nullptr);
}


// returns the error string (resource allocated), or null
char* HeapDumper::error_as_C_string() const {
  if (error() != nullptr) {
    char* str = ResourceArea::strdup(error());
    return str;
  } else {
    return nullptr;
  }
}

// set the error string
void HeapDumper::set_error(char const* error) {
  if (_error != nullptr) {
    os::free(_error);
  }
  if (error == nullptr) {
    _error = nullptr;
  } else {
    _error = os::strdup(error);
    assert(_error != nullptr, "allocation failure");
  }
}

// Called by out-of-memory error reporting by a single Java thread
// outside of a JVM safepoint
void HeapDumper::dump_heap_from_oome() {
  HeapDumper::dump_heap(true);
}

// Called by error reporting by a single Java thread outside of a JVM safepoint,
// or by heap dumping by the VM thread during a (GC) safepoint. Thus, these various
// callers are strictly serialized and guaranteed not to interfere below. For more
// general use, however, this method will need modification to prevent
// inteference when updating the static variables base_path and dump_file_seq below.
void HeapDumper::dump_heap() {
  HeapDumper::dump_heap(false);
}

void HeapDumper::dump_heap(bool oome) {
  static char base_path[JVM_MAXPATHLEN] = {'\0'};
  static uint dump_file_seq = 0;
  char my_path[JVM_MAXPATHLEN];
  const int max_digit_chars = 20;
  const char* dump_file_name = HeapDumpGzipLevel > 0 ? "java_pid%p.hprof.gz" : "java_pid%p.hprof";

  // The dump file defaults to java_pid<pid>.hprof in the current working
  // directory. HeapDumpPath=<file> can be used to specify an alternative
  // dump file name or a directory where dump file is created.
  if (dump_file_seq == 0) { // first time in, we initialize base_path
    // Set base path (name or directory, default or custom, without seq no), doing %p substitution.
    const char *path_src = (HeapDumpPath != nullptr && HeapDumpPath[0] != '\0') ? HeapDumpPath : dump_file_name;
    if (!Arguments::copy_expand_pid(path_src, strlen(path_src), base_path, JVM_MAXPATHLEN - max_digit_chars)) {
      warning("Cannot create heap dump file.  HeapDumpPath is too long.");
      return;
    }
    // Check if the path is an existing directory
    DIR* dir = os::opendir(base_path);
    if (dir != nullptr) {
      os::closedir(dir);
      // Path is a directory.  Append a file separator (if needed).
      size_t fs_len = strlen(os::file_separator());
      if (strlen(base_path) >= fs_len) {
        char* end = base_path;
        end += (strlen(base_path) - fs_len);
        if (strcmp(end, os::file_separator()) != 0) {
          strcat(base_path, os::file_separator());
        }
      }
      // Then add the default name, with %p substitution.  Use my_path temporarily.
      if (!Arguments::copy_expand_pid(dump_file_name, strlen(dump_file_name), my_path, JVM_MAXPATHLEN - max_digit_chars)) {
        warning("Cannot create heap dump file.  HeapDumpPath is too long.");
        return;
      }
      const size_t dlen = strlen(base_path);
      jio_snprintf(&base_path[dlen], sizeof(base_path) - dlen, "%s", my_path);
    }
    strncpy(my_path, base_path, JVM_MAXPATHLEN);
  } else {
    // Append a sequence number id for dumps following the first
    const size_t len = strlen(base_path) + max_digit_chars + 2; // for '.' and \0
    jio_snprintf(my_path, len, "%s.%d", base_path, dump_file_seq);
  }
  dump_file_seq++;   // increment seq number for next time we dump

  HeapDumper dumper(false /* no GC before heap dump */,
                    oome  /* pass along out-of-memory-error flag */);
  dumper.dump(my_path, tty, HeapDumpGzipLevel);
}