allocators.cpp

Go to the documentation of this file.

/*
 *
 * Copyright (c) 1996,1997
 * Silicon Graphics Computer Systems, Inc.
 *
 * Copyright (c) 1997
 * Moscow Center for SPARC Technology
 *
 * Copyright (c) 1999
 * Boris Fomitchev
 *
 * This material is provided "as is", with absolutely no warranty expressed
 * or implied. Any use is at your own risk.
 *
 * Permission to use or copy this software for any purpose is hereby granted
 * without fee, provided the above notices are retained on all copies.
 * Permission to modify the code and to distribute modified code is granted,
 * provided the above notices are retained, and a notice that the code was
 * modified is included with the above copyright notice.
 *
 */

#include "stlport_prefix.h"

#include <memory>

#if defined (__GNUC__) && (defined (__CYGWIN__) || defined (__MINGW32__))
#  include <malloc.h>
#endif

#if defined (_STLP_PTHREADS) && !defined (_STLP_NO_THREADS)
#  include <pthread_alloc>
#  include <cerrno>
#endif

#include <stl/_threads.h>

#include "lock_free_slist.h"

#if defined (__WATCOMC__)
#  pragma warning 13 9
#  pragma warning 367 9
#  pragma warning 368 9
#endif

#if defined (_STLP_SGI_THREADS)
  // We test whether threads are in use before locking.
  // Perhaps this should be moved into stl_threads.h, but that
  // probably makes it harder to avoid the procedure call when
  // it isn't needed.
extern "C" {
  extern int __us_rsthread_malloc;
}
#endif

// Specialised debug form of new operator which does not provide "false"
// memory leaks when run with debug CRT libraries.
#if defined (_STLP_MSVC) && (_STLP_MSVC >= 1020 && defined (_STLP_DEBUG_ALLOC)) && !defined (_STLP_WCE)
#  include <crtdbg.h>
inline char* __stlp_new_chunk(size_t __bytes) {
  void *__chunk = _STLP_CHECK_NULL_ALLOC(::operator new(__bytes, __FILE__, __LINE__));
  return __STATIC_CAST(char*, __chunk);
}
inline void __stlp_delete_chunck(void* __p) { ::operator delete(__p, __FILE__, __LINE__); }
#else
#  ifdef _STLP_NODE_ALLOC_USE_MALLOC
#    include <cstdlib>
inline char* __stlp_new_chunk(size_t __bytes) {
  // do not use _STLP_CHECK_NULL_ALLOC, this macro is dedicated to new operator.
  void *__chunk = _STLP_VENDOR_CSTD::malloc(__bytes);
  if (__chunk == 0) {
    _STLP_THROW_BAD_ALLOC;
  }
  return __STATIC_CAST(char*, __chunk);
}
inline void __stlp_delete_chunck(void* __p) { _STLP_VENDOR_CSTD::free(__p); }
#  else
inline char* __stlp_new_chunk(size_t __bytes)
{ return __STATIC_CAST(char*, _STLP_STD::__stl_new(__bytes)); }
inline void __stlp_delete_chunck(void* __p) { _STLP_STD::__stl_delete(__p); }
#  endif
#endif

/* This is an additional atomic operations to the ones already defined in
 * stl/_threads.h, platform should try to support it to improve performance.
 * __add_atomic_t _STLP_ATOMIC_ADD(volatile __add_atomic_t* __target, __add_atomic_t __val) :
 * does *__target = *__target + __val and returns the old *__target value */
typedef long __add_atomic_t;
typedef unsigned long __uadd_atomic_t;

#if defined (__GNUC__) && defined (__i386__)
inline long _STLP_atomic_add_gcc_x86(long volatile* p, long addend) {
  long result;
  __asm__ __volatile__
    ("lock; xaddl %1, %0;"
    :"=m" (*p), "=r" (result)
    :"m"  (*p), "1"  (addend)
    :"cc");
 return result + addend;
}
#  define _STLP_ATOMIC_ADD(__dst, __val)  _STLP_atomic_add_gcc_x86(__dst, __val)
#elif defined (_STLP_WIN32THREADS)
// The Win32 API function InterlockedExchangeAdd is not available on Windows 95.
#  if !defined (_STLP_WIN95_LIKE)
#    if defined (_STLP_NEW_PLATFORM_SDK)
#      define _STLP_ATOMIC_ADD(__dst, __val) InterlockedExchangeAdd(__dst, __val)
#    else
#      define _STLP_ATOMIC_ADD(__dst, __val) InterlockedExchangeAdd(__CONST_CAST(__add_atomic_t*, __dst), __val)
#    endif
#  endif
#endif

#if defined (__OS400__)
// dums 02/05/2007: is it really necessary ?
enum { _ALIGN = 16, _ALIGN_SHIFT = 4 };
#else
enum { _ALIGN = 2 * sizeof(void*), _ALIGN_SHIFT = 2 + sizeof(void*) / 4 };
#endif

#define _S_FREELIST_INDEX(__bytes) ((__bytes - size_t(1)) >> (int)_ALIGN_SHIFT)

_STLP_BEGIN_NAMESPACE

// malloc_alloc out-of-memory handling
static __oom_handler_type __oom_handler = __STATIC_CAST(__oom_handler_type, 0);

#ifdef _STLP_THREADS
_STLP_mutex __oom_handler_lock;
#endif

void* _STLP_CALL __malloc_alloc::allocate(size_t __n)
{
  void *__result = malloc(__n);
  if ( 0 == __result ) {
    __oom_handler_type __my_malloc_handler;

    for (;;) {
      {
#ifdef _STLP_THREADS
        _STLP_auto_lock _l( __oom_handler_lock );
#endif
        __my_malloc_handler = __oom_handler;
      }
      if ( 0 == __my_malloc_handler) {
        _STLP_THROW_BAD_ALLOC;
      }
      (*__my_malloc_handler)();
      __result = malloc(__n);
      if ( __result )
        return __result;
    }
  }
  return __result;
}

__oom_handler_type _STLP_CALL __malloc_alloc::set_malloc_handler(__oom_handler_type __f)
{
#ifdef _STLP_THREADS
  _STLP_auto_lock _l( __oom_handler_lock );
#endif
  __oom_handler_type __old = __oom_handler;
  __oom_handler = __f;
  return __old;
}

// *******************************************************
// Default node allocator.
// With a reasonable compiler, this should be roughly as fast as the
// original STL class-specific allocators, but with less fragmentation.
//
// Important implementation properties:
// 1. If the client request an object of size > _MAX_BYTES, the resulting
//    object will be obtained directly from malloc.
// 2. In all other cases, we allocate an object of size exactly
//    _S_round_up(requested_size).  Thus the client has enough size
//    information that we can return the object to the proper free list
//    without permanently losing part of the object.
//

#define _STLP_NFREELISTS 16

#if defined (_STLP_LEAKS_PEDANTIC) && defined (_STLP_USE_DYNAMIC_LIB)
/*
 * We can only do cleanup of the node allocator memory pool if we are
 * sure that the STLport library is used as a shared one as it guaranties
 * the unicity of the node allocator instance. Without that guaranty node
 * allocator instances might exchange memory blocks making the implementation
 * of a cleaning process much more complicated.
 */
#  define _STLP_DO_CLEAN_NODE_ALLOC
#endif

/* When STLport is used without multi threaded safety we use the node allocator
 * implementation with locks as locks becomes no-op. The lock free implementation
 * always use system specific atomic operations which are slower than 'normal'
 * ones.
 */
#if defined (_STLP_THREADS) && \
    defined (_STLP_HAS_ATOMIC_FREELIST) && defined (_STLP_ATOMIC_ADD)
/*
 * We have an implementation of the atomic freelist (_STLP_atomic_freelist)
 * for this architecture and compiler.  That means we can use the non-blocking
 * implementation of the node-allocation engine.*/
#  define _STLP_USE_LOCK_FREE_IMPLEMENTATION
#endif

#if !defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
#  if defined (_STLP_THREADS)

class _Node_Alloc_Lock {
  static _STLP_STATIC_MUTEX& _S_Mutex() {
    static _STLP_STATIC_MUTEX mutex _STLP_MUTEX_INITIALIZER;
    return mutex;
  }
public:
  _Node_Alloc_Lock() {
#    if defined (_STLP_SGI_THREADS)
    if (__us_rsthread_malloc)
#    endif
      _S_Mutex()._M_acquire_lock();
  }

  ~_Node_Alloc_Lock() {
#    if defined (_STLP_SGI_THREADS)
    if (__us_rsthread_malloc)
#    endif
      _S_Mutex()._M_release_lock();
  }
};

#  else

class _Node_Alloc_Lock {
public:
  _Node_Alloc_Lock() { }
  ~_Node_Alloc_Lock() { }
};

#  endif

struct _Node_alloc_obj {
  _Node_alloc_obj * _M_next;
};
#endif

class __node_alloc_impl {
  static inline size_t _STLP_CALL _S_round_up(size_t __bytes)
  { return (((__bytes) + (size_t)_ALIGN-1) & ~((size_t)_ALIGN - 1)); }

#if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
  typedef _STLP_atomic_freelist::item   _Obj;
  typedef _STLP_atomic_freelist         _Freelist;
  typedef _STLP_atomic_freelist         _ChunkList;

  // Header of blocks of memory that have been allocated as part of
  // a larger chunk but have not yet been chopped up into nodes.
  struct _FreeBlockHeader : public _STLP_atomic_freelist::item {
    char* _M_end;     // pointer to end of free memory
  };
#else
  typedef _Node_alloc_obj       _Obj;
  typedef _Obj* _STLP_VOLATILE  _Freelist;
  typedef _Obj*                 _ChunkList;
#endif

private:
  // Returns an object of size __n, and optionally adds to size __n free list.
  static _Obj* _S_refill(size_t __n);
  // Allocates a chunk for nobjs of size __p_size.  nobjs may be reduced
  // if it is inconvenient to allocate the requested number.
  static char* _S_chunk_alloc(size_t __p_size, int& __nobjs);
  // Chunk allocation state.
  static _Freelist _S_free_list[_STLP_NFREELISTS];
  // Amount of total allocated memory
#if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
  static _STLP_VOLATILE __add_atomic_t _S_heap_size;
#else
  static size_t _S_heap_size;
#endif

#if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
  // List of blocks of free memory
  static _STLP_atomic_freelist  _S_free_mem_blocks;
#else
  // Start of the current free memory buffer
  static char* _S_start_free;
  // End of the current free memory buffer
  static char* _S_end_free;
#endif

#if defined (_STLP_DO_CLEAN_NODE_ALLOC)
public:
  // Methods to report alloc/dealloc calls to the counter system.
#  if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
  typedef _STLP_VOLATILE __stl_atomic_t _AllocCounter;
#  else
  typedef __stl_atomic_t _AllocCounter;
#  endif
  static _AllocCounter& _STLP_CALL _S_alloc_counter();
  static void _S_alloc_call();
  static void _S_dealloc_call();

private:
  // Free all the allocated chuncks of memory
  static void _S_chunk_dealloc();
  // Beginning of the linked list of allocated chunks of memory
  static _ChunkList _S_chunks;
#endif /* _STLP_DO_CLEAN_NODE_ALLOC */

public:
  /* __n must be > 0      */
  static void* _M_allocate(size_t& __n);
  /* __p may not be 0 */
  static void _M_deallocate(void *__p, size_t __n);
};

#if !defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
void* __node_alloc_impl::_M_allocate(size_t& __n) {
  __n = _S_round_up(__n);
  _Obj * _STLP_VOLATILE * __my_free_list = _S_free_list + _S_FREELIST_INDEX(__n);
  _Obj *__r;

  // Acquire the lock here with a constructor call.
  // This ensures that it is released in exit or during stack
  // unwinding.
  _Node_Alloc_Lock __lock_instance;

  if ( (__r  = *__my_free_list) != 0 ) {
    *__my_free_list = __r->_M_next;
  } else {
    __r = _S_refill(__n);
  }
#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
  _S_alloc_call();
#  endif
  // lock is released here
  return __r;
}

void __node_alloc_impl::_M_deallocate(void *__p, size_t __n) {
  _Obj * _STLP_VOLATILE * __my_free_list = _S_free_list + _S_FREELIST_INDEX(__n);
  _Obj * __pobj = __STATIC_CAST(_Obj*, __p);

  // acquire lock
  _Node_Alloc_Lock __lock_instance;
  __pobj->_M_next = *__my_free_list;
  *__my_free_list = __pobj;

#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
  _S_dealloc_call();
#  endif
  // lock is released here
}

#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
#    define _STLP_OFFSET sizeof(_Obj)
#  else
#    define _STLP_OFFSET 0
#  endif

/* We allocate memory in large chunks in order to avoid fragmenting     */
/* the malloc heap too much.                                            */
/* We assume that size is properly aligned.                             */
/* We hold the allocation lock.                                         */
char* __node_alloc_impl::_S_chunk_alloc(size_t _p_size, int& __nobjs) {
  char* __result;
  size_t __total_bytes = _p_size * __nobjs;
  size_t __bytes_left = _S_end_free - _S_start_free;

  if (__bytes_left > 0) {
    if (__bytes_left >= __total_bytes) {
      __result = _S_start_free;
      _S_start_free += __total_bytes;
      return __result;
    }

    if (__bytes_left >= _p_size) {
      __nobjs = (int)(__bytes_left / _p_size);
      __total_bytes = _p_size * __nobjs;
      __result = _S_start_free;
      _S_start_free += __total_bytes;
      return __result;
    }

    // Try to make use of the left-over piece.
    _Obj* _STLP_VOLATILE* __my_free_list = _S_free_list + _S_FREELIST_INDEX(__bytes_left);
    __REINTERPRET_CAST(_Obj*, _S_start_free)->_M_next = *__my_free_list;
    *__my_free_list = __REINTERPRET_CAST(_Obj*, _S_start_free);
    _S_start_free = _S_end_free = 0;
  }

  size_t __bytes_to_get = 2 * __total_bytes + _S_round_up(_S_heap_size) + _STLP_OFFSET;

  _STLP_TRY {
    _S_start_free = __stlp_new_chunk(__bytes_to_get);
  }
#if defined (_STLP_USE_EXCEPTIONS)
  catch (const _STLP_STD::bad_alloc&) {
    _Obj* _STLP_VOLATILE* __my_free_list;
    _Obj* __p;
    // Try to do with what we have.  That can't hurt.
    // We do not try smaller requests, since that tends
    // to result in disaster on multi-process machines.
    for (size_t __i = _p_size; __i <= (size_t)_MAX_BYTES; __i += (size_t)_ALIGN) {
      __my_free_list = _S_free_list + _S_FREELIST_INDEX(__i);
      __p = *__my_free_list;
      if (0 != __p) {
        *__my_free_list = __p -> _M_next;
        _S_start_free = __REINTERPRET_CAST(char*, __p);
        _S_end_free = _S_start_free + __i;
        return _S_chunk_alloc(_p_size, __nobjs);
        // Any leftover piece will eventually make it to the
        // right free list.
      }
    }
    __bytes_to_get = __total_bytes + _STLP_OFFSET;
    _S_start_free = __stlp_new_chunk(__bytes_to_get);
  }
#endif

  _S_heap_size += __bytes_to_get >> 4;
#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
  __REINTERPRET_CAST(_Obj*, _S_start_free)->_M_next = _S_chunks;
  _S_chunks = __REINTERPRET_CAST(_Obj*, _S_start_free);
#  endif
  _S_end_free = _S_start_free + __bytes_to_get;
  _S_start_free += _STLP_OFFSET;
  return _S_chunk_alloc(_p_size, __nobjs);
}

/* Returns an object of size __n, and optionally adds to size __n free list.*/
/* We assume that __n is properly aligned.                                  */
/* We hold the allocation lock.                                             */
_Node_alloc_obj* __node_alloc_impl::_S_refill(size_t __n) {
  int __nobjs = 20;
  char* __chunk = _S_chunk_alloc(__n, __nobjs);

  if (1 == __nobjs) return __REINTERPRET_CAST(_Obj*, __chunk);

  _Obj* _STLP_VOLATILE* __my_free_list = _S_free_list + _S_FREELIST_INDEX(__n);
  _Obj* __result;
  _Obj* __current_obj;
  _Obj* __next_obj;

  /* Build free list in chunk */
  __result = __REINTERPRET_CAST(_Obj*, __chunk);
  *__my_free_list = __next_obj = __REINTERPRET_CAST(_Obj*, __chunk + __n);
  for (--__nobjs; --__nobjs; ) {
    __current_obj = __next_obj;
    __next_obj = __REINTERPRET_CAST(_Obj*, __REINTERPRET_CAST(char*, __next_obj) + __n);
    __current_obj->_M_next = __next_obj;
  }
  __next_obj->_M_next = 0;
  return __result;
}

#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
void __node_alloc_impl::_S_alloc_call()
{ ++_S_alloc_counter(); }

void __node_alloc_impl::_S_dealloc_call() {
  __stl_atomic_t &counter = _S_alloc_counter();
  if (--counter == 0)
  { _S_chunk_dealloc(); }
}

/* We deallocate all the memory chunks      */
void __node_alloc_impl::_S_chunk_dealloc() {
  _Obj *__pcur = _S_chunks, *__pnext;
  while (__pcur != 0) {
    __pnext = __pcur->_M_next;
    __stlp_delete_chunck(__pcur);
    __pcur = __pnext;
  }
  _S_chunks = 0;
  _S_start_free = _S_end_free = 0;
  _S_heap_size = 0;
  memset(__REINTERPRET_CAST(char*, __CONST_CAST(_Obj**, &_S_free_list[0])), 0, _STLP_NFREELISTS * sizeof(_Obj*));
}
#  endif

#else

void* __node_alloc_impl::_M_allocate(size_t& __n) {
  __n = _S_round_up(__n);
  _Obj* __r = _S_free_list[_S_FREELIST_INDEX(__n)].pop();
  if (__r  == 0)
  { __r = _S_refill(__n); }

#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
  _S_alloc_call();
#  endif
  return __r;
}

void __node_alloc_impl::_M_deallocate(void *__p, size_t __n) {
  _S_free_list[_S_FREELIST_INDEX(__n)].push(__STATIC_CAST(_Obj*, __p));

#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
  _S_dealloc_call();
#  endif
}

/* Returns an object of size __n, and optionally adds additional ones to    */
/* freelist of objects of size __n.                                         */
/* We assume that __n is properly aligned.                                  */
__node_alloc_impl::_Obj* __node_alloc_impl::_S_refill(size_t __n) {
  int __nobjs = 20;
  char* __chunk = _S_chunk_alloc(__n, __nobjs);

  if (__nobjs <= 1)
    return __REINTERPRET_CAST(_Obj*, __chunk);

  // Push all new nodes (minus first one) onto freelist
  _Obj* __result   = __REINTERPRET_CAST(_Obj*, __chunk);
  _Obj* __cur_item = __result;
  _Freelist* __my_freelist = _S_free_list + _S_FREELIST_INDEX(__n);
  for (--__nobjs; __nobjs != 0; --__nobjs) {
    __cur_item  = __REINTERPRET_CAST(_Obj*, __REINTERPRET_CAST(char*, __cur_item) + __n);
    __my_freelist->push(__cur_item);
  }
  return __result;
}

#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
#    define _STLP_OFFSET _ALIGN
#  else
#    define _STLP_OFFSET 0
#  endif

/* We allocate memory in large chunks in order to avoid fragmenting     */
/* the malloc heap too much.                                            */
/* We assume that size is properly aligned.                             */
char* __node_alloc_impl::_S_chunk_alloc(size_t _p_size, int& __nobjs) {
#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
  //We are going to add a small memory block to keep all the allocated blocks
  //address, we need to do so respecting the memory alignment. The following
  //static assert checks that the reserved block is big enough to store a pointer.
  _STLP_STATIC_ASSERT(sizeof(_Obj) <= _ALIGN)
#  endif
  char*  __result       = 0;
  __add_atomic_t __total_bytes  = __STATIC_CAST(__add_atomic_t, _p_size) * __nobjs;

  _FreeBlockHeader* __block = __STATIC_CAST(_FreeBlockHeader*, _S_free_mem_blocks.pop());
  if (__block != 0) {
    // We checked a block out and can now mess with it with impugnity.
    // We'll put the remainder back into the list if we're done with it below.
    char*  __buf_start  = __REINTERPRET_CAST(char*, __block);
    __add_atomic_t __bytes_left = __block->_M_end - __buf_start;

    if ((__bytes_left < __total_bytes) && (__bytes_left >= __STATIC_CAST(__add_atomic_t, _p_size))) {
      // There's enough left for at least one object, but not as much as we wanted
      __result      = __buf_start;
      __nobjs       = (int)(__bytes_left/_p_size);
      __total_bytes = __STATIC_CAST(__add_atomic_t, _p_size) * __nobjs;
      __bytes_left -= __total_bytes;
      __buf_start  += __total_bytes;
    }
    else if (__bytes_left >= __total_bytes) {
      // The block has enough left to satisfy all that was asked for
      __result      = __buf_start;
      __bytes_left -= __total_bytes;
      __buf_start  += __total_bytes;
    }

    if (__bytes_left != 0) {
      // There is still some memory left over in block after we satisfied our request.
      if ((__result != 0) && (__bytes_left >= (__add_atomic_t)sizeof(_FreeBlockHeader))) {
        // We were able to allocate at least one object and there is still enough
        // left to put remainder back into list.
        _FreeBlockHeader* __newblock = __REINTERPRET_CAST(_FreeBlockHeader*, __buf_start);
        __newblock->_M_end  = __block->_M_end;
        _S_free_mem_blocks.push(__newblock);
      }
      else {
        // We were not able to allocate enough for at least one object.
        // Shove into freelist of nearest (rounded-down!) size.
        size_t __rounded_down = _S_round_up(__bytes_left + 1) - (size_t)_ALIGN;
        if (__rounded_down > 0)
          _S_free_list[_S_FREELIST_INDEX(__rounded_down)].push((_Obj*)__buf_start);
      }
    }
    if (__result != 0)
      return __result;
  }

  // We couldn't satisfy it from the list of free blocks, get new memory.
  __add_atomic_t __bytes_to_get = 2 * __total_bytes +
                                  __STATIC_CAST(__add_atomic_t,
                                                _S_round_up(__STATIC_CAST(__uadd_atomic_t, _STLP_ATOMIC_ADD(&_S_heap_size, 0)))) +
                                  _STLP_OFFSET;
  _STLP_TRY {
    __result = __stlp_new_chunk(__bytes_to_get);
  }
#if defined (_STLP_USE_EXCEPTIONS)
  catch (const bad_alloc&) {
    // Allocation failed; try to canibalize from freelist of a larger object size.
    for (size_t __i = _p_size; __i <= (size_t)_MAX_BYTES; __i += (size_t)_ALIGN) {
      _Obj* __p  = _S_free_list[_S_FREELIST_INDEX(__i)].pop();
      if (0 != __p) {
        if (__i < sizeof(_FreeBlockHeader)) {
          // Not enough to put into list of free blocks, divvy it up here.
          // Use as much as possible for this request and shove remainder into freelist.
          __nobjs = (int)(__i/_p_size);
          __total_bytes = __nobjs * __STATIC_CAST(__add_atomic_t, _p_size);
          size_t __bytes_left = __i - __total_bytes;
          size_t __rounded_down = _S_round_up(__bytes_left+1) - (size_t)_ALIGN;
          if (__rounded_down > 0) {
            _S_free_list[_S_FREELIST_INDEX(__rounded_down)].push(__REINTERPRET_CAST(_Obj*, __REINTERPRET_CAST(char*, __p) + __total_bytes));
          }
          return __REINTERPRET_CAST(char*, __p);
        }
        else {
          // Add node to list of available blocks and recursively allocate from it.
          _FreeBlockHeader* __newblock = (_FreeBlockHeader*)__p;
          __newblock->_M_end  = __REINTERPRET_CAST(char*, __p) + __i;
          _S_free_mem_blocks.push(__newblock);
          return _S_chunk_alloc(_p_size, __nobjs);
        }
      }
    }

    // We were not able to find something in a freelist, try to allocate a smaller amount.
    __bytes_to_get  = __total_bytes + _STLP_OFFSET;
    __result = __stlp_new_chunk(__bytes_to_get);

    // This should either throw an exception or remedy the situation.
    // Thus we assume it succeeded.
  }
#endif
  // Alignment check
  _STLP_VERBOSE_ASSERT(((__REINTERPRET_CAST(size_t, __result) & __STATIC_CAST(size_t, _ALIGN - 1)) == 0),
                       _StlMsg_DBA_DELETED_TWICE)
  _STLP_ATOMIC_ADD(&_S_heap_size, __bytes_to_get >> 4);

#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
  // We have to track the allocated memory chunks for release on exit.
  _S_chunks.push(__REINTERPRET_CAST(_Obj*, __result));
  __result       += _ALIGN;
  __bytes_to_get -= _ALIGN;
#  endif

  if (__bytes_to_get > __total_bytes) {
    // Push excess memory allocated in this chunk into list of free memory blocks
    _FreeBlockHeader* __freeblock = __REINTERPRET_CAST(_FreeBlockHeader*, __result + __total_bytes);
    __freeblock->_M_end  = __result + __bytes_to_get;
    _S_free_mem_blocks.push(__freeblock);
  }
  return __result;
}

#  if defined (_STLP_DO_CLEAN_NODE_ALLOC)
void __node_alloc_impl::_S_alloc_call()
{ _STLP_ATOMIC_INCREMENT(&_S_alloc_counter()); }

void __node_alloc_impl::_S_dealloc_call() {
  _STLP_VOLATILE __stl_atomic_t *pcounter = &_S_alloc_counter();
  if (_STLP_ATOMIC_DECREMENT(pcounter) == 0)
    _S_chunk_dealloc();
}

/* We deallocate all the memory chunks      */
void __node_alloc_impl::_S_chunk_dealloc() {
  // Note: The _Node_alloc_helper class ensures that this function
  // will only be called when the (shared) library is unloaded or the
  // process is shutdown.  It's thus not possible that another thread
  // is currently trying to allocate a node (we're not thread-safe here).
  //

  // Clear the free blocks and all freelistst.  This makes sure that if
  // for some reason more memory is allocated again during shutdown
  // (it'd also be really nasty to leave references to deallocated memory).
  _S_free_mem_blocks.clear();
  _S_heap_size      = 0;

  for (size_t __i = 0; __i < _STLP_NFREELISTS; ++__i) {
    _S_free_list[__i].clear();
  }

  // Detach list of chunks and free them all
  _Obj* __chunk = _S_chunks.clear();
  while (__chunk != 0) {
    _Obj* __next = __chunk->_M_next;
    __stlp_delete_chunck(__chunk);
    __chunk  = __next;
  }
}
#  endif

#endif

#if defined (_STLP_DO_CLEAN_NODE_ALLOC)
struct __node_alloc_cleaner {
  ~__node_alloc_cleaner()
  { __node_alloc_impl::_S_dealloc_call(); }
};

#  if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
_STLP_VOLATILE __stl_atomic_t& _STLP_CALL
#  else
__stl_atomic_t& _STLP_CALL
#  endif
__node_alloc_impl::_S_alloc_counter() {
  static _AllocCounter _S_counter = 1;
  static __node_alloc_cleaner _S_node_alloc_cleaner;
  return _S_counter;
}
#endif

#if !defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
_Node_alloc_obj * _STLP_VOLATILE
__node_alloc_impl::_S_free_list[_STLP_NFREELISTS]
= {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
// The 16 zeros are necessary to make version 4.1 of the SunPro
// compiler happy.  Otherwise it appears to allocate too little
// space for the array.
#else
_STLP_atomic_freelist __node_alloc_impl::_S_free_list[_STLP_NFREELISTS];
_STLP_atomic_freelist __node_alloc_impl::_S_free_mem_blocks;
#endif

#if !defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
char *__node_alloc_impl::_S_start_free = 0;
char *__node_alloc_impl::_S_end_free = 0;
#endif

#if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
_STLP_VOLATILE __add_atomic_t
#else
size_t
#endif
__node_alloc_impl::_S_heap_size = 0;

#if defined (_STLP_DO_CLEAN_NODE_ALLOC)
#  if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
_STLP_atomic_freelist __node_alloc_impl::_S_chunks;
#  else
_Node_alloc_obj* __node_alloc_impl::_S_chunks  = 0;
#  endif
#endif

void * _STLP_CALL __node_alloc::_M_allocate(size_t& __n)
{ return __node_alloc_impl::_M_allocate(__n); }

void _STLP_CALL __node_alloc::_M_deallocate(void *__p, size_t __n)
{ __node_alloc_impl::_M_deallocate(__p, __n); }

#if defined (_STLP_PTHREADS) && !defined (_STLP_NO_THREADS)

#  define _STLP_DATA_ALIGNMENT 8

_STLP_MOVE_TO_PRIV_NAMESPACE

// *******************************************************
// __perthread_alloc implementation
union _Pthread_alloc_obj {
  union _Pthread_alloc_obj * __free_list_link;
  char __client_data[_STLP_DATA_ALIGNMENT];    /* The client sees this.    */
};

// Pthread allocators don't appear to the client to have meaningful
// instances.  We do in fact need to associate some state with each
// thread.  That state is represented by _Pthread_alloc_per_thread_state.

struct _Pthread_alloc_per_thread_state {
  typedef _Pthread_alloc_obj __obj;
  enum { _S_NFREELISTS = _MAX_BYTES / _STLP_DATA_ALIGNMENT };

  // Free list link for list of available per thread structures.
  // When one of these becomes available for reuse due to thread
  // termination, any objects in its free list remain associated
  // with it.  The whole structure may then be used by a newly
  // created thread.
  _Pthread_alloc_per_thread_state() : __next(0)
  { memset((void *)__CONST_CAST(_Pthread_alloc_obj**, __free_list), 0, (size_t)_S_NFREELISTS * sizeof(__obj *)); }
  // Returns an object of size __n, and possibly adds to size n free list.
  void *_M_refill(size_t __n);

  _Pthread_alloc_obj* volatile __free_list[_S_NFREELISTS];
  _Pthread_alloc_per_thread_state *__next;
  // this data member is only to be used by per_thread_allocator, which returns memory to the originating thread.
  _STLP_mutex _M_lock;
};

// Pthread-specific allocator.
class _Pthread_alloc_impl {
public: // but only for internal use:
  typedef _Pthread_alloc_per_thread_state __state_type;
  typedef char value_type;

  // Allocates a chunk for nobjs of size size.  nobjs may be reduced
  // if it is inconvenient to allocate the requested number.
  static char *_S_chunk_alloc(size_t __size, size_t &__nobjs, __state_type*);

  enum {_S_ALIGN = _STLP_DATA_ALIGNMENT};

  static size_t _S_round_up(size_t __bytes)
  { return (((__bytes) + (int)_S_ALIGN - 1) & ~((int)_S_ALIGN - 1)); }
  static size_t _S_freelist_index(size_t __bytes)
  { return (((__bytes) + (int)_S_ALIGN - 1) / (int)_S_ALIGN - 1); }

private:
  // Chunk allocation state. And other shared state.
  // Protected by _S_chunk_allocator_lock.
  static _STLP_STATIC_MUTEX _S_chunk_allocator_lock;
  static char *_S_start_free;
  static char *_S_end_free;
  static size_t _S_heap_size;
  static __state_type *_S_free_per_thread_states;
  static pthread_key_t _S_key;
  static bool _S_key_initialized;
  // Pthread key under which per thread state is stored.
  // Allocator instances that are currently unclaimed by any thread.
  static void _S_destructor(void *instance);
  // Function to be called on thread exit to reclaim per thread
  // state.
  static __state_type *_S_new_per_thread_state();
public:
  // Return a recycled or new per thread state.
  static __state_type *_S_get_per_thread_state();
private:
        // ensure that the current thread has an associated
        // per thread state.
  class _M_lock;
  friend class _M_lock;
  class _M_lock {
  public:
    _M_lock () { _S_chunk_allocator_lock._M_acquire_lock(); }
    ~_M_lock () { _S_chunk_allocator_lock._M_release_lock(); }
  };

public:

  /* n must be > 0      */
  static void * allocate(size_t& __n);

  /* p may not be 0 */
  static void deallocate(void *__p, size_t __n);

  // boris : versions for per_thread_allocator
  /* n must be > 0      */
  static void * allocate(size_t& __n, __state_type* __a);

  /* p may not be 0 */
  static void deallocate(void *__p, size_t __n, __state_type* __a);

  static void * reallocate(void *__p, size_t __old_sz, size_t& __new_sz);
};

/* Returns an object of size n, and optionally adds to size n free list.*/
/* We assume that n is properly aligned.                                */
/* We hold the allocation lock.                                         */
void *_Pthread_alloc_per_thread_state::_M_refill(size_t __n) {
  typedef _Pthread_alloc_obj __obj;
  size_t __nobjs = 128;
  char * __chunk = _Pthread_alloc_impl::_S_chunk_alloc(__n, __nobjs, this);
  __obj * volatile * __my_free_list;
  __obj * __result;
  __obj * __current_obj, * __next_obj;
  size_t __i;

  if (1 == __nobjs)  {
    return __chunk;
  }

  __my_free_list = __free_list + _Pthread_alloc_impl::_S_freelist_index(__n);

  /* Build free list in chunk */
  __result = (__obj *)__chunk;
  *__my_free_list = __next_obj = (__obj *)(__chunk + __n);
  for (__i = 1; ; ++__i) {
    __current_obj = __next_obj;
    __next_obj = (__obj *)((char *)__next_obj + __n);
    if (__nobjs - 1 == __i) {
      __current_obj -> __free_list_link = 0;
      break;
    } else {
      __current_obj -> __free_list_link = __next_obj;
    }
  }
  return __result;
}

void _Pthread_alloc_impl::_S_destructor(void *__instance) {
  _M_lock __lock_instance;  // Need to acquire lock here.
  _Pthread_alloc_per_thread_state* __s = (_Pthread_alloc_per_thread_state*)__instance;
  __s -> __next = _S_free_per_thread_states;
  _S_free_per_thread_states = __s;
}

_Pthread_alloc_per_thread_state* _Pthread_alloc_impl::_S_new_per_thread_state() {
  /* lock already held here.  */
  if (0 != _S_free_per_thread_states) {
    _Pthread_alloc_per_thread_state *__result = _S_free_per_thread_states;
    _S_free_per_thread_states = _S_free_per_thread_states -> __next;
    return __result;
  }
  else {
    return new _Pthread_alloc_per_thread_state;
  }
}

_Pthread_alloc_per_thread_state* _Pthread_alloc_impl::_S_get_per_thread_state() {
  int __ret_code;
  __state_type* __result;

  if (_S_key_initialized && (__result = (__state_type*) pthread_getspecific(_S_key)))
    return __result;

  /*REFERENCED*/
  _M_lock __lock_instance;  // Need to acquire lock here.
  if (!_S_key_initialized) {
    if (pthread_key_create(&_S_key, _S_destructor)) {
      _STLP_THROW_BAD_ALLOC;  // failed
    }
    _S_key_initialized = true;
  }

  __result = _S_new_per_thread_state();
  __ret_code = pthread_setspecific(_S_key, __result);
  if (__ret_code) {
    if (__ret_code == ENOMEM) {
      _STLP_THROW_BAD_ALLOC;
    } else {
  // EINVAL
      _STLP_ABORT();
    }
  }
  return __result;
}

/* We allocate memory in large chunks in order to avoid fragmenting     */
/* the malloc heap too much.                                            */
/* We assume that size is properly aligned.                             */
char *_Pthread_alloc_impl::_S_chunk_alloc(size_t __p_size, size_t &__nobjs, _Pthread_alloc_per_thread_state *__a) {
  typedef _Pthread_alloc_obj __obj;
  {
    char * __result;
    size_t __total_bytes;
    size_t __bytes_left;
    /*REFERENCED*/
    _M_lock __lock_instance;         // Acquire lock for this routine

    __total_bytes = __p_size * __nobjs;
    __bytes_left = _S_end_free - _S_start_free;
    if (__bytes_left >= __total_bytes) {
      __result = _S_start_free;
      _S_start_free += __total_bytes;
      return __result;
    } else if (__bytes_left >= __p_size) {
      __nobjs = __bytes_left/__p_size;
      __total_bytes = __p_size * __nobjs;
      __result = _S_start_free;
      _S_start_free += __total_bytes;
      return __result;
    } else {
      size_t __bytes_to_get = 2 * __total_bytes + _S_round_up(_S_heap_size);
      // Try to make use of the left-over piece.
      if (__bytes_left > 0) {
        __obj * volatile * __my_free_list = __a->__free_list + _S_freelist_index(__bytes_left);
        ((__obj *)_S_start_free) -> __free_list_link = *__my_free_list;
        *__my_free_list = (__obj *)_S_start_free;
      }
#  ifdef _SGI_SOURCE
      // Try to get memory that's aligned on something like a
      // cache line boundary, so as to avoid parceling out
      // parts of the same line to different threads and thus
      // possibly different processors.
      {
        const int __cache_line_size = 128;  // probable upper bound
        __bytes_to_get &= ~(__cache_line_size-1);
        _S_start_free = (char *)memalign(__cache_line_size, __bytes_to_get);
        if (0 == _S_start_free) {
          _S_start_free = (char *)__malloc_alloc::allocate(__bytes_to_get);
        }
      }
#  else  /* !SGI_SOURCE */
      _S_start_free = (char *)__malloc_alloc::allocate(__bytes_to_get);
#  endif
      _S_heap_size += __bytes_to_get >> 4;
      _S_end_free = _S_start_free + __bytes_to_get;
    }
  }
  // lock is released here
  return _S_chunk_alloc(__p_size, __nobjs, __a);
}


/* n must be > 0      */
void *_Pthread_alloc_impl::allocate(size_t& __n) {
  typedef _Pthread_alloc_obj __obj;
  __obj * volatile * __my_free_list;
  __obj * __result;
  __state_type* __a;

  if (__n > _MAX_BYTES) {
    return __malloc_alloc::allocate(__n);
  }

  __n = _S_round_up(__n);
  __a = _S_get_per_thread_state();

  __my_free_list = __a->__free_list + _S_freelist_index(__n);
  __result = *__my_free_list;
  if (__result == 0) {
    void *__r = __a->_M_refill(__n);
    return __r;
  }
  *__my_free_list = __result->__free_list_link;
  return __result;
};

/* p may not be 0 */
void _Pthread_alloc_impl::deallocate(void *__p, size_t __n) {
  typedef _Pthread_alloc_obj __obj;
  __obj *__q = (__obj *)__p;
  __obj * volatile * __my_free_list;
  __state_type* __a;

  if (__n > _MAX_BYTES) {
      __malloc_alloc::deallocate(__p, __n);
      return;
  }

  __a = _S_get_per_thread_state();

  __my_free_list = __a->__free_list + _S_freelist_index(__n);
  __q -> __free_list_link = *__my_free_list;
  *__my_free_list = __q;
}

// boris : versions for per_thread_allocator
/* n must be > 0      */
void *_Pthread_alloc_impl::allocate(size_t& __n, __state_type* __a) {
  typedef _Pthread_alloc_obj __obj;
  __obj * volatile * __my_free_list;
  __obj * __result;

  if (__n > _MAX_BYTES) {
    return __malloc_alloc::allocate(__n);
  }
  __n = _S_round_up(__n);

  // boris : here, we have to lock per thread state, as we may be getting memory from
  // different thread pool.
  _STLP_auto_lock __lock(__a->_M_lock);

  __my_free_list = __a->__free_list + _S_freelist_index(__n);
  __result = *__my_free_list;
  if (__result == 0) {
    void *__r = __a->_M_refill(__n);
    return __r;
  }
  *__my_free_list = __result->__free_list_link;
  return __result;
};

/* p may not be 0 */
void _Pthread_alloc_impl::deallocate(void *__p, size_t __n, __state_type* __a) {
  typedef _Pthread_alloc_obj __obj;
  __obj *__q = (__obj *)__p;
  __obj * volatile * __my_free_list;

  if (__n > _MAX_BYTES) {
    __malloc_alloc::deallocate(__p, __n);
    return;
  }

  // boris : here, we have to lock per thread state, as we may be returning memory from
  // different thread.
  _STLP_auto_lock __lock(__a->_M_lock);

  __my_free_list = __a->__free_list + _S_freelist_index(__n);
  __q -> __free_list_link = *__my_free_list;
  *__my_free_list = __q;
}

void *_Pthread_alloc_impl::reallocate(void *__p, size_t __old_sz, size_t& __new_sz) {
  void * __result;
  size_t __copy_sz;

  if (__old_sz > _MAX_BYTES && __new_sz > _MAX_BYTES) {
    return realloc(__p, __new_sz);
  }

  if (_S_round_up(__old_sz) == _S_round_up(__new_sz)) return __p;
  __result = allocate(__new_sz);
  __copy_sz = __new_sz > __old_sz? __old_sz : __new_sz;
  memcpy(__result, __p, __copy_sz);
  deallocate(__p, __old_sz);
  return __result;
}

_Pthread_alloc_per_thread_state* _Pthread_alloc_impl::_S_free_per_thread_states = 0;
pthread_key_t _Pthread_alloc_impl::_S_key = 0;
_STLP_STATIC_MUTEX _Pthread_alloc_impl::_S_chunk_allocator_lock _STLP_MUTEX_INITIALIZER;
bool _Pthread_alloc_impl::_S_key_initialized = false;
char *_Pthread_alloc_impl::_S_start_free = 0;
char *_Pthread_alloc_impl::_S_end_free = 0;
size_t _Pthread_alloc_impl::_S_heap_size = 0;

void * _STLP_CALL _Pthread_alloc::allocate(size_t& __n)
{ return _Pthread_alloc_impl::allocate(__n); }
void _STLP_CALL _Pthread_alloc::deallocate(void *__p, size_t __n)
{ _Pthread_alloc_impl::deallocate(__p, __n); }
void * _STLP_CALL _Pthread_alloc::allocate(size_t& __n, __state_type* __a)
{ return _Pthread_alloc_impl::allocate(__n, __a); }
void _STLP_CALL _Pthread_alloc::deallocate(void *__p, size_t __n, __state_type* __a)
{ _Pthread_alloc_impl::deallocate(__p, __n, __a); }
void * _STLP_CALL _Pthread_alloc::reallocate(void *__p, size_t __old_sz, size_t& __new_sz)
{ return _Pthread_alloc_impl::reallocate(__p, __old_sz, __new_sz); }
_Pthread_alloc_per_thread_state* _STLP_CALL _Pthread_alloc::_S_get_per_thread_state()
{ return _Pthread_alloc_impl::_S_get_per_thread_state(); }

_STLP_MOVE_TO_STD_NAMESPACE

#endif

_STLP_END_NAMESPACE

#undef _S_FREELIST_INDEX