1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/build/stlport/src/allocators.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,1121 @@
1.4 +/*
1.5 + *
1.6 + * Copyright (c) 1996,1997
1.7 + * Silicon Graphics Computer Systems, Inc.
1.8 + *
1.9 + * Copyright (c) 1997
1.10 + * Moscow Center for SPARC Technology
1.11 + *
1.12 + * Copyright (c) 1999
1.13 + * Boris Fomitchev
1.14 + *
1.15 + * This material is provided "as is", with absolutely no warranty expressed
1.16 + * or implied. Any use is at your own risk.
1.17 + *
1.18 + * Permission to use or copy this software for any purpose is hereby granted
1.19 + * without fee, provided the above notices are retained on all copies.
1.20 + * Permission to modify the code and to distribute modified code is granted,
1.21 + * provided the above notices are retained, and a notice that the code was
1.22 + * modified is included with the above copyright notice.
1.23 + *
1.24 + */
1.25 +
1.26 +#include "stlport_prefix.h"
1.27 +
1.28 +#include <memory>
1.29 +
1.30 +#if defined (__GNUC__) && (defined (__CYGWIN__) || defined (__MINGW32__))
1.31 +# include <malloc.h>
1.32 +#endif
1.33 +
1.34 +#if defined (_STLP_PTHREADS) && !defined (_STLP_NO_THREADS)
1.35 +# include <pthread_alloc>
1.36 +# include <cerrno>
1.37 +#endif
1.38 +
1.39 +#include <stl/_threads.h>
1.40 +
1.41 +#include "lock_free_slist.h"
1.42 +
1.43 +#if defined (__WATCOMC__)
1.44 +# pragma warning 13 9
1.45 +# pragma warning 367 9
1.46 +# pragma warning 368 9
1.47 +#endif
1.48 +
1.49 +#if defined (_STLP_SGI_THREADS)
1.50 + // We test whether threads are in use before locking.
1.51 + // Perhaps this should be moved into stl_threads.h, but that
1.52 + // probably makes it harder to avoid the procedure call when
1.53 + // it isn't needed.
1.54 +extern "C" {
1.55 + extern int __us_rsthread_malloc;
1.56 +}
1.57 +#endif
1.58 +
1.59 +// Specialised debug form of new operator which does not provide "false"
1.60 +// memory leaks when run with debug CRT libraries.
1.61 +#if defined (_STLP_MSVC) && (_STLP_MSVC >= 1020 && defined (_STLP_DEBUG_ALLOC)) && !defined (_STLP_WCE)
1.62 +# include <crtdbg.h>
1.63 +inline char* __stlp_new_chunk(size_t __bytes) {
1.64 + void *__chunk = _STLP_CHECK_NULL_ALLOC(::operator new(__bytes, __FILE__, __LINE__));
1.65 + return __STATIC_CAST(char*, __chunk);
1.66 +}
1.67 +inline void __stlp_delete_chunck(void* __p) { ::operator delete(__p, __FILE__, __LINE__); }
1.68 +#else
1.69 +# ifdef _STLP_NODE_ALLOC_USE_MALLOC
1.70 +# include <cstdlib>
1.71 +inline char* __stlp_new_chunk(size_t __bytes) {
1.72 + // do not use _STLP_CHECK_NULL_ALLOC, this macro is dedicated to new operator.
1.73 + void *__chunk = _STLP_VENDOR_CSTD::malloc(__bytes);
1.74 + if (__chunk == 0) {
1.75 + _STLP_THROW_BAD_ALLOC;
1.76 + }
1.77 + return __STATIC_CAST(char*, __chunk);
1.78 +}
1.79 +inline void __stlp_delete_chunck(void* __p) { _STLP_VENDOR_CSTD::free(__p); }
1.80 +# else
1.81 +inline char* __stlp_new_chunk(size_t __bytes)
1.82 +{ return __STATIC_CAST(char*, _STLP_STD::__stl_new(__bytes)); }
1.83 +inline void __stlp_delete_chunck(void* __p) { _STLP_STD::__stl_delete(__p); }
1.84 +# endif
1.85 +#endif
1.86 +
1.87 +/* This is an additional atomic operations to the ones already defined in
1.88 + * stl/_threads.h, platform should try to support it to improve performance.
1.89 + * __add_atomic_t _STLP_ATOMIC_ADD(volatile __add_atomic_t* __target, __add_atomic_t __val) :
1.90 + * does *__target = *__target + __val and returns the old *__target value */
1.91 +typedef long __add_atomic_t;
1.92 +typedef unsigned long __uadd_atomic_t;
1.93 +
1.94 +#if defined (__GNUC__) && defined (__i386__)
1.95 +inline long _STLP_atomic_add_gcc_x86(long volatile* p, long addend) {
1.96 + long result;
1.97 + __asm__ __volatile__
1.98 + ("lock; xaddl %1, %0;"
1.99 + :"=m" (*p), "=r" (result)
1.100 + :"m" (*p), "1" (addend)
1.101 + :"cc");
1.102 + return result + addend;
1.103 +}
1.104 +# define _STLP_ATOMIC_ADD(__dst, __val) _STLP_atomic_add_gcc_x86(__dst, __val)
1.105 +#elif defined (_STLP_WIN32THREADS)
1.106 +// The Win32 API function InterlockedExchangeAdd is not available on Windows 95.
1.107 +# if !defined (_STLP_WIN95_LIKE)
1.108 +# if defined (_STLP_NEW_PLATFORM_SDK)
1.109 +# define _STLP_ATOMIC_ADD(__dst, __val) InterlockedExchangeAdd(__dst, __val)
1.110 +# else
1.111 +# define _STLP_ATOMIC_ADD(__dst, __val) InterlockedExchangeAdd(__CONST_CAST(__add_atomic_t*, __dst), __val)
1.112 +# endif
1.113 +# endif
1.114 +#endif
1.115 +
1.116 +#if defined (__OS400__)
1.117 +// dums 02/05/2007: is it really necessary ?
1.118 +enum { _ALIGN = 16, _ALIGN_SHIFT = 4 };
1.119 +#else
1.120 +enum { _ALIGN = 2 * sizeof(void*), _ALIGN_SHIFT = 2 + sizeof(void*) / 4 };
1.121 +#endif
1.122 +
1.123 +#define _S_FREELIST_INDEX(__bytes) ((__bytes - size_t(1)) >> (int)_ALIGN_SHIFT)
1.124 +
1.125 +_STLP_BEGIN_NAMESPACE
1.126 +
1.127 +// malloc_alloc out-of-memory handling
1.128 +static __oom_handler_type __oom_handler = __STATIC_CAST(__oom_handler_type, 0);
1.129 +
1.130 +#ifdef _STLP_THREADS
1.131 +_STLP_mutex __oom_handler_lock;
1.132 +#endif
1.133 +
1.134 +void* _STLP_CALL __malloc_alloc::allocate(size_t __n)
1.135 +{
1.136 + void *__result = malloc(__n);
1.137 + if ( 0 == __result ) {
1.138 + __oom_handler_type __my_malloc_handler;
1.139 +
1.140 + for (;;) {
1.141 + {
1.142 +#ifdef _STLP_THREADS
1.143 + _STLP_auto_lock _l( __oom_handler_lock );
1.144 +#endif
1.145 + __my_malloc_handler = __oom_handler;
1.146 + }
1.147 + if ( 0 == __my_malloc_handler) {
1.148 + _STLP_THROW_BAD_ALLOC;
1.149 + }
1.150 + (*__my_malloc_handler)();
1.151 + __result = malloc(__n);
1.152 + if ( __result )
1.153 + return __result;
1.154 + }
1.155 + }
1.156 + return __result;
1.157 +}
1.158 +
1.159 +__oom_handler_type _STLP_CALL __malloc_alloc::set_malloc_handler(__oom_handler_type __f)
1.160 +{
1.161 +#ifdef _STLP_THREADS
1.162 + _STLP_auto_lock _l( __oom_handler_lock );
1.163 +#endif
1.164 + __oom_handler_type __old = __oom_handler;
1.165 + __oom_handler = __f;
1.166 + return __old;
1.167 +}
1.168 +
1.169 +// *******************************************************
1.170 +// Default node allocator.
1.171 +// With a reasonable compiler, this should be roughly as fast as the
1.172 +// original STL class-specific allocators, but with less fragmentation.
1.173 +//
1.174 +// Important implementation properties:
1.175 +// 1. If the client request an object of size > _MAX_BYTES, the resulting
1.176 +// object will be obtained directly from malloc.
1.177 +// 2. In all other cases, we allocate an object of size exactly
1.178 +// _S_round_up(requested_size). Thus the client has enough size
1.179 +// information that we can return the object to the proper free list
1.180 +// without permanently losing part of the object.
1.181 +//
1.182 +
1.183 +#define _STLP_NFREELISTS 16
1.184 +
1.185 +#if defined (_STLP_LEAKS_PEDANTIC) && defined (_STLP_USE_DYNAMIC_LIB)
1.186 +/*
1.187 + * We can only do cleanup of the node allocator memory pool if we are
1.188 + * sure that the STLport library is used as a shared one as it guaranties
1.189 + * the unicity of the node allocator instance. Without that guaranty node
1.190 + * allocator instances might exchange memory blocks making the implementation
1.191 + * of a cleaning process much more complicated.
1.192 + */
1.193 +# define _STLP_DO_CLEAN_NODE_ALLOC
1.194 +#endif
1.195 +
1.196 +/* When STLport is used without multi threaded safety we use the node allocator
1.197 + * implementation with locks as locks becomes no-op. The lock free implementation
1.198 + * always use system specific atomic operations which are slower than 'normal'
1.199 + * ones.
1.200 + */
1.201 +#if defined (_STLP_THREADS) && \
1.202 + defined (_STLP_HAS_ATOMIC_FREELIST) && defined (_STLP_ATOMIC_ADD)
1.203 +/*
1.204 + * We have an implementation of the atomic freelist (_STLP_atomic_freelist)
1.205 + * for this architecture and compiler. That means we can use the non-blocking
1.206 + * implementation of the node-allocation engine.*/
1.207 +# define _STLP_USE_LOCK_FREE_IMPLEMENTATION
1.208 +#endif
1.209 +
1.210 +#if !defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.211 +# if defined (_STLP_THREADS)
1.212 +
1.213 +class _Node_Alloc_Lock {
1.214 + static _STLP_STATIC_MUTEX& _S_Mutex() {
1.215 + static _STLP_STATIC_MUTEX mutex _STLP_MUTEX_INITIALIZER;
1.216 + return mutex;
1.217 + }
1.218 +public:
1.219 + _Node_Alloc_Lock() {
1.220 +# if defined (_STLP_SGI_THREADS)
1.221 + if (__us_rsthread_malloc)
1.222 +# endif
1.223 + _S_Mutex()._M_acquire_lock();
1.224 + }
1.225 +
1.226 + ~_Node_Alloc_Lock() {
1.227 +# if defined (_STLP_SGI_THREADS)
1.228 + if (__us_rsthread_malloc)
1.229 +# endif
1.230 + _S_Mutex()._M_release_lock();
1.231 + }
1.232 +};
1.233 +
1.234 +# else
1.235 +
1.236 +class _Node_Alloc_Lock {
1.237 +public:
1.238 + _Node_Alloc_Lock() { }
1.239 + ~_Node_Alloc_Lock() { }
1.240 +};
1.241 +
1.242 +# endif
1.243 +
1.244 +struct _Node_alloc_obj {
1.245 + _Node_alloc_obj * _M_next;
1.246 +};
1.247 +#endif
1.248 +
1.249 +class __node_alloc_impl {
1.250 + static inline size_t _STLP_CALL _S_round_up(size_t __bytes)
1.251 + { return (((__bytes) + (size_t)_ALIGN-1) & ~((size_t)_ALIGN - 1)); }
1.252 +
1.253 +#if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.254 + typedef _STLP_atomic_freelist::item _Obj;
1.255 + typedef _STLP_atomic_freelist _Freelist;
1.256 + typedef _STLP_atomic_freelist _ChunkList;
1.257 +
1.258 + // Header of blocks of memory that have been allocated as part of
1.259 + // a larger chunk but have not yet been chopped up into nodes.
1.260 + struct _FreeBlockHeader : public _STLP_atomic_freelist::item {
1.261 + char* _M_end; // pointer to end of free memory
1.262 + };
1.263 +#else
1.264 + typedef _Node_alloc_obj _Obj;
1.265 + typedef _Obj* _STLP_VOLATILE _Freelist;
1.266 + typedef _Obj* _ChunkList;
1.267 +#endif
1.268 +
1.269 +private:
1.270 + // Returns an object of size __n, and optionally adds to size __n free list.
1.271 + static _Obj* _S_refill(size_t __n);
1.272 + // Allocates a chunk for nobjs of size __p_size. nobjs may be reduced
1.273 + // if it is inconvenient to allocate the requested number.
1.274 + static char* _S_chunk_alloc(size_t __p_size, int& __nobjs);
1.275 + // Chunk allocation state.
1.276 + static _Freelist _S_free_list[_STLP_NFREELISTS];
1.277 + // Amount of total allocated memory
1.278 +#if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.279 + static _STLP_VOLATILE __add_atomic_t _S_heap_size;
1.280 +#else
1.281 + static size_t _S_heap_size;
1.282 +#endif
1.283 +
1.284 +#if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.285 + // List of blocks of free memory
1.286 + static _STLP_atomic_freelist _S_free_mem_blocks;
1.287 +#else
1.288 + // Start of the current free memory buffer
1.289 + static char* _S_start_free;
1.290 + // End of the current free memory buffer
1.291 + static char* _S_end_free;
1.292 +#endif
1.293 +
1.294 +#if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.295 +public:
1.296 + // Methods to report alloc/dealloc calls to the counter system.
1.297 +# if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.298 + typedef _STLP_VOLATILE __stl_atomic_t _AllocCounter;
1.299 +# else
1.300 + typedef __stl_atomic_t _AllocCounter;
1.301 +# endif
1.302 + static _AllocCounter& _STLP_CALL _S_alloc_counter();
1.303 + static void _S_alloc_call();
1.304 + static void _S_dealloc_call();
1.305 +
1.306 +private:
1.307 + // Free all the allocated chuncks of memory
1.308 + static void _S_chunk_dealloc();
1.309 + // Beginning of the linked list of allocated chunks of memory
1.310 + static _ChunkList _S_chunks;
1.311 +#endif /* _STLP_DO_CLEAN_NODE_ALLOC */
1.312 +
1.313 +public:
1.314 + /* __n must be > 0 */
1.315 + static void* _M_allocate(size_t& __n);
1.316 + /* __p may not be 0 */
1.317 + static void _M_deallocate(void *__p, size_t __n);
1.318 +};
1.319 +
1.320 +#if !defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.321 +void* __node_alloc_impl::_M_allocate(size_t& __n) {
1.322 + __n = _S_round_up(__n);
1.323 + _Obj * _STLP_VOLATILE * __my_free_list = _S_free_list + _S_FREELIST_INDEX(__n);
1.324 + _Obj *__r;
1.325 +
1.326 + // Acquire the lock here with a constructor call.
1.327 + // This ensures that it is released in exit or during stack
1.328 + // unwinding.
1.329 + _Node_Alloc_Lock __lock_instance;
1.330 +
1.331 + if ( (__r = *__my_free_list) != 0 ) {
1.332 + *__my_free_list = __r->_M_next;
1.333 + } else {
1.334 + __r = _S_refill(__n);
1.335 + }
1.336 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.337 + _S_alloc_call();
1.338 +# endif
1.339 + // lock is released here
1.340 + return __r;
1.341 +}
1.342 +
1.343 +void __node_alloc_impl::_M_deallocate(void *__p, size_t __n) {
1.344 + _Obj * _STLP_VOLATILE * __my_free_list = _S_free_list + _S_FREELIST_INDEX(__n);
1.345 + _Obj * __pobj = __STATIC_CAST(_Obj*, __p);
1.346 +
1.347 + // acquire lock
1.348 + _Node_Alloc_Lock __lock_instance;
1.349 + __pobj->_M_next = *__my_free_list;
1.350 + *__my_free_list = __pobj;
1.351 +
1.352 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.353 + _S_dealloc_call();
1.354 +# endif
1.355 + // lock is released here
1.356 +}
1.357 +
1.358 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.359 +# define _STLP_OFFSET sizeof(_Obj)
1.360 +# else
1.361 +# define _STLP_OFFSET 0
1.362 +# endif
1.363 +
1.364 +/* We allocate memory in large chunks in order to avoid fragmenting */
1.365 +/* the malloc heap too much. */
1.366 +/* We assume that size is properly aligned. */
1.367 +/* We hold the allocation lock. */
1.368 +char* __node_alloc_impl::_S_chunk_alloc(size_t _p_size, int& __nobjs) {
1.369 + char* __result;
1.370 + size_t __total_bytes = _p_size * __nobjs;
1.371 + size_t __bytes_left = _S_end_free - _S_start_free;
1.372 +
1.373 + if (__bytes_left > 0) {
1.374 + if (__bytes_left >= __total_bytes) {
1.375 + __result = _S_start_free;
1.376 + _S_start_free += __total_bytes;
1.377 + return __result;
1.378 + }
1.379 +
1.380 + if (__bytes_left >= _p_size) {
1.381 + __nobjs = (int)(__bytes_left / _p_size);
1.382 + __total_bytes = _p_size * __nobjs;
1.383 + __result = _S_start_free;
1.384 + _S_start_free += __total_bytes;
1.385 + return __result;
1.386 + }
1.387 +
1.388 + // Try to make use of the left-over piece.
1.389 + _Obj* _STLP_VOLATILE* __my_free_list = _S_free_list + _S_FREELIST_INDEX(__bytes_left);
1.390 + __REINTERPRET_CAST(_Obj*, _S_start_free)->_M_next = *__my_free_list;
1.391 + *__my_free_list = __REINTERPRET_CAST(_Obj*, _S_start_free);
1.392 + _S_start_free = _S_end_free = 0;
1.393 + }
1.394 +
1.395 + size_t __bytes_to_get = 2 * __total_bytes + _S_round_up(_S_heap_size) + _STLP_OFFSET;
1.396 +
1.397 + _STLP_TRY {
1.398 + _S_start_free = __stlp_new_chunk(__bytes_to_get);
1.399 + }
1.400 +#if defined (_STLP_USE_EXCEPTIONS)
1.401 + catch (const _STLP_STD::bad_alloc&) {
1.402 + _Obj* _STLP_VOLATILE* __my_free_list;
1.403 + _Obj* __p;
1.404 + // Try to do with what we have. That can't hurt.
1.405 + // We do not try smaller requests, since that tends
1.406 + // to result in disaster on multi-process machines.
1.407 + for (size_t __i = _p_size; __i <= (size_t)_MAX_BYTES; __i += (size_t)_ALIGN) {
1.408 + __my_free_list = _S_free_list + _S_FREELIST_INDEX(__i);
1.409 + __p = *__my_free_list;
1.410 + if (0 != __p) {
1.411 + *__my_free_list = __p -> _M_next;
1.412 + _S_start_free = __REINTERPRET_CAST(char*, __p);
1.413 + _S_end_free = _S_start_free + __i;
1.414 + return _S_chunk_alloc(_p_size, __nobjs);
1.415 + // Any leftover piece will eventually make it to the
1.416 + // right free list.
1.417 + }
1.418 + }
1.419 + __bytes_to_get = __total_bytes + _STLP_OFFSET;
1.420 + _S_start_free = __stlp_new_chunk(__bytes_to_get);
1.421 + }
1.422 +#endif
1.423 +
1.424 + _S_heap_size += __bytes_to_get >> 4;
1.425 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.426 + __REINTERPRET_CAST(_Obj*, _S_start_free)->_M_next = _S_chunks;
1.427 + _S_chunks = __REINTERPRET_CAST(_Obj*, _S_start_free);
1.428 +# endif
1.429 + _S_end_free = _S_start_free + __bytes_to_get;
1.430 + _S_start_free += _STLP_OFFSET;
1.431 + return _S_chunk_alloc(_p_size, __nobjs);
1.432 +}
1.433 +
1.434 +/* Returns an object of size __n, and optionally adds to size __n free list.*/
1.435 +/* We assume that __n is properly aligned. */
1.436 +/* We hold the allocation lock. */
1.437 +_Node_alloc_obj* __node_alloc_impl::_S_refill(size_t __n) {
1.438 + int __nobjs = 20;
1.439 + char* __chunk = _S_chunk_alloc(__n, __nobjs);
1.440 +
1.441 + if (1 == __nobjs) return __REINTERPRET_CAST(_Obj*, __chunk);
1.442 +
1.443 + _Obj* _STLP_VOLATILE* __my_free_list = _S_free_list + _S_FREELIST_INDEX(__n);
1.444 + _Obj* __result;
1.445 + _Obj* __current_obj;
1.446 + _Obj* __next_obj;
1.447 +
1.448 + /* Build free list in chunk */
1.449 + __result = __REINTERPRET_CAST(_Obj*, __chunk);
1.450 + *__my_free_list = __next_obj = __REINTERPRET_CAST(_Obj*, __chunk + __n);
1.451 + for (--__nobjs; --__nobjs; ) {
1.452 + __current_obj = __next_obj;
1.453 + __next_obj = __REINTERPRET_CAST(_Obj*, __REINTERPRET_CAST(char*, __next_obj) + __n);
1.454 + __current_obj->_M_next = __next_obj;
1.455 + }
1.456 + __next_obj->_M_next = 0;
1.457 + return __result;
1.458 +}
1.459 +
1.460 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.461 +void __node_alloc_impl::_S_alloc_call()
1.462 +{ ++_S_alloc_counter(); }
1.463 +
1.464 +void __node_alloc_impl::_S_dealloc_call() {
1.465 + __stl_atomic_t &counter = _S_alloc_counter();
1.466 + if (--counter == 0)
1.467 + { _S_chunk_dealloc(); }
1.468 +}
1.469 +
1.470 +/* We deallocate all the memory chunks */
1.471 +void __node_alloc_impl::_S_chunk_dealloc() {
1.472 + _Obj *__pcur = _S_chunks, *__pnext;
1.473 + while (__pcur != 0) {
1.474 + __pnext = __pcur->_M_next;
1.475 + __stlp_delete_chunck(__pcur);
1.476 + __pcur = __pnext;
1.477 + }
1.478 + _S_chunks = 0;
1.479 + _S_start_free = _S_end_free = 0;
1.480 + _S_heap_size = 0;
1.481 + memset(__REINTERPRET_CAST(char*, __CONST_CAST(_Obj**, &_S_free_list[0])), 0, _STLP_NFREELISTS * sizeof(_Obj*));
1.482 +}
1.483 +# endif
1.484 +
1.485 +#else
1.486 +
1.487 +void* __node_alloc_impl::_M_allocate(size_t& __n) {
1.488 + __n = _S_round_up(__n);
1.489 + _Obj* __r = _S_free_list[_S_FREELIST_INDEX(__n)].pop();
1.490 + if (__r == 0)
1.491 + { __r = _S_refill(__n); }
1.492 +
1.493 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.494 + _S_alloc_call();
1.495 +# endif
1.496 + return __r;
1.497 +}
1.498 +
1.499 +void __node_alloc_impl::_M_deallocate(void *__p, size_t __n) {
1.500 + _S_free_list[_S_FREELIST_INDEX(__n)].push(__STATIC_CAST(_Obj*, __p));
1.501 +
1.502 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.503 + _S_dealloc_call();
1.504 +# endif
1.505 +}
1.506 +
1.507 +/* Returns an object of size __n, and optionally adds additional ones to */
1.508 +/* freelist of objects of size __n. */
1.509 +/* We assume that __n is properly aligned. */
1.510 +__node_alloc_impl::_Obj* __node_alloc_impl::_S_refill(size_t __n) {
1.511 + int __nobjs = 20;
1.512 + char* __chunk = _S_chunk_alloc(__n, __nobjs);
1.513 +
1.514 + if (__nobjs <= 1)
1.515 + return __REINTERPRET_CAST(_Obj*, __chunk);
1.516 +
1.517 + // Push all new nodes (minus first one) onto freelist
1.518 + _Obj* __result = __REINTERPRET_CAST(_Obj*, __chunk);
1.519 + _Obj* __cur_item = __result;
1.520 + _Freelist* __my_freelist = _S_free_list + _S_FREELIST_INDEX(__n);
1.521 + for (--__nobjs; __nobjs != 0; --__nobjs) {
1.522 + __cur_item = __REINTERPRET_CAST(_Obj*, __REINTERPRET_CAST(char*, __cur_item) + __n);
1.523 + __my_freelist->push(__cur_item);
1.524 + }
1.525 + return __result;
1.526 +}
1.527 +
1.528 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.529 +# define _STLP_OFFSET _ALIGN
1.530 +# else
1.531 +# define _STLP_OFFSET 0
1.532 +# endif
1.533 +
1.534 +/* We allocate memory in large chunks in order to avoid fragmenting */
1.535 +/* the malloc heap too much. */
1.536 +/* We assume that size is properly aligned. */
1.537 +char* __node_alloc_impl::_S_chunk_alloc(size_t _p_size, int& __nobjs) {
1.538 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.539 + //We are going to add a small memory block to keep all the allocated blocks
1.540 + //address, we need to do so respecting the memory alignment. The following
1.541 + //static assert checks that the reserved block is big enough to store a pointer.
1.542 + _STLP_STATIC_ASSERT(sizeof(_Obj) <= _ALIGN)
1.543 +# endif
1.544 + char* __result = 0;
1.545 + __add_atomic_t __total_bytes = __STATIC_CAST(__add_atomic_t, _p_size) * __nobjs;
1.546 +
1.547 + _FreeBlockHeader* __block = __STATIC_CAST(_FreeBlockHeader*, _S_free_mem_blocks.pop());
1.548 + if (__block != 0) {
1.549 + // We checked a block out and can now mess with it with impugnity.
1.550 + // We'll put the remainder back into the list if we're done with it below.
1.551 + char* __buf_start = __REINTERPRET_CAST(char*, __block);
1.552 + __add_atomic_t __bytes_left = __block->_M_end - __buf_start;
1.553 +
1.554 + if ((__bytes_left < __total_bytes) && (__bytes_left >= __STATIC_CAST(__add_atomic_t, _p_size))) {
1.555 + // There's enough left for at least one object, but not as much as we wanted
1.556 + __result = __buf_start;
1.557 + __nobjs = (int)(__bytes_left/_p_size);
1.558 + __total_bytes = __STATIC_CAST(__add_atomic_t, _p_size) * __nobjs;
1.559 + __bytes_left -= __total_bytes;
1.560 + __buf_start += __total_bytes;
1.561 + }
1.562 + else if (__bytes_left >= __total_bytes) {
1.563 + // The block has enough left to satisfy all that was asked for
1.564 + __result = __buf_start;
1.565 + __bytes_left -= __total_bytes;
1.566 + __buf_start += __total_bytes;
1.567 + }
1.568 +
1.569 + if (__bytes_left != 0) {
1.570 + // There is still some memory left over in block after we satisfied our request.
1.571 + if ((__result != 0) && (__bytes_left >= (__add_atomic_t)sizeof(_FreeBlockHeader))) {
1.572 + // We were able to allocate at least one object and there is still enough
1.573 + // left to put remainder back into list.
1.574 + _FreeBlockHeader* __newblock = __REINTERPRET_CAST(_FreeBlockHeader*, __buf_start);
1.575 + __newblock->_M_end = __block->_M_end;
1.576 + _S_free_mem_blocks.push(__newblock);
1.577 + }
1.578 + else {
1.579 + // We were not able to allocate enough for at least one object.
1.580 + // Shove into freelist of nearest (rounded-down!) size.
1.581 + size_t __rounded_down = _S_round_up(__bytes_left + 1) - (size_t)_ALIGN;
1.582 + if (__rounded_down > 0)
1.583 + _S_free_list[_S_FREELIST_INDEX(__rounded_down)].push((_Obj*)__buf_start);
1.584 + }
1.585 + }
1.586 + if (__result != 0)
1.587 + return __result;
1.588 + }
1.589 +
1.590 + // We couldn't satisfy it from the list of free blocks, get new memory.
1.591 + __add_atomic_t __bytes_to_get = 2 * __total_bytes +
1.592 + __STATIC_CAST(__add_atomic_t,
1.593 + _S_round_up(__STATIC_CAST(__uadd_atomic_t, _STLP_ATOMIC_ADD(&_S_heap_size, 0)))) +
1.594 + _STLP_OFFSET;
1.595 + _STLP_TRY {
1.596 + __result = __stlp_new_chunk(__bytes_to_get);
1.597 + }
1.598 +#if defined (_STLP_USE_EXCEPTIONS)
1.599 + catch (const bad_alloc&) {
1.600 + // Allocation failed; try to canibalize from freelist of a larger object size.
1.601 + for (size_t __i = _p_size; __i <= (size_t)_MAX_BYTES; __i += (size_t)_ALIGN) {
1.602 + _Obj* __p = _S_free_list[_S_FREELIST_INDEX(__i)].pop();
1.603 + if (0 != __p) {
1.604 + if (__i < sizeof(_FreeBlockHeader)) {
1.605 + // Not enough to put into list of free blocks, divvy it up here.
1.606 + // Use as much as possible for this request and shove remainder into freelist.
1.607 + __nobjs = (int)(__i/_p_size);
1.608 + __total_bytes = __nobjs * __STATIC_CAST(__add_atomic_t, _p_size);
1.609 + size_t __bytes_left = __i - __total_bytes;
1.610 + size_t __rounded_down = _S_round_up(__bytes_left+1) - (size_t)_ALIGN;
1.611 + if (__rounded_down > 0) {
1.612 + _S_free_list[_S_FREELIST_INDEX(__rounded_down)].push(__REINTERPRET_CAST(_Obj*, __REINTERPRET_CAST(char*, __p) + __total_bytes));
1.613 + }
1.614 + return __REINTERPRET_CAST(char*, __p);
1.615 + }
1.616 + else {
1.617 + // Add node to list of available blocks and recursively allocate from it.
1.618 + _FreeBlockHeader* __newblock = (_FreeBlockHeader*)__p;
1.619 + __newblock->_M_end = __REINTERPRET_CAST(char*, __p) + __i;
1.620 + _S_free_mem_blocks.push(__newblock);
1.621 + return _S_chunk_alloc(_p_size, __nobjs);
1.622 + }
1.623 + }
1.624 + }
1.625 +
1.626 + // We were not able to find something in a freelist, try to allocate a smaller amount.
1.627 + __bytes_to_get = __total_bytes + _STLP_OFFSET;
1.628 + __result = __stlp_new_chunk(__bytes_to_get);
1.629 +
1.630 + // This should either throw an exception or remedy the situation.
1.631 + // Thus we assume it succeeded.
1.632 + }
1.633 +#endif
1.634 + // Alignment check
1.635 + _STLP_VERBOSE_ASSERT(((__REINTERPRET_CAST(size_t, __result) & __STATIC_CAST(size_t, _ALIGN - 1)) == 0),
1.636 + _StlMsg_DBA_DELETED_TWICE)
1.637 + _STLP_ATOMIC_ADD(&_S_heap_size, __bytes_to_get >> 4);
1.638 +
1.639 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.640 + // We have to track the allocated memory chunks for release on exit.
1.641 + _S_chunks.push(__REINTERPRET_CAST(_Obj*, __result));
1.642 + __result += _ALIGN;
1.643 + __bytes_to_get -= _ALIGN;
1.644 +# endif
1.645 +
1.646 + if (__bytes_to_get > __total_bytes) {
1.647 + // Push excess memory allocated in this chunk into list of free memory blocks
1.648 + _FreeBlockHeader* __freeblock = __REINTERPRET_CAST(_FreeBlockHeader*, __result + __total_bytes);
1.649 + __freeblock->_M_end = __result + __bytes_to_get;
1.650 + _S_free_mem_blocks.push(__freeblock);
1.651 + }
1.652 + return __result;
1.653 +}
1.654 +
1.655 +# if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.656 +void __node_alloc_impl::_S_alloc_call()
1.657 +{ _STLP_ATOMIC_INCREMENT(&_S_alloc_counter()); }
1.658 +
1.659 +void __node_alloc_impl::_S_dealloc_call() {
1.660 + _STLP_VOLATILE __stl_atomic_t *pcounter = &_S_alloc_counter();
1.661 + if (_STLP_ATOMIC_DECREMENT(pcounter) == 0)
1.662 + _S_chunk_dealloc();
1.663 +}
1.664 +
1.665 +/* We deallocate all the memory chunks */
1.666 +void __node_alloc_impl::_S_chunk_dealloc() {
1.667 + // Note: The _Node_alloc_helper class ensures that this function
1.668 + // will only be called when the (shared) library is unloaded or the
1.669 + // process is shutdown. It's thus not possible that another thread
1.670 + // is currently trying to allocate a node (we're not thread-safe here).
1.671 + //
1.672 +
1.673 + // Clear the free blocks and all freelistst. This makes sure that if
1.674 + // for some reason more memory is allocated again during shutdown
1.675 + // (it'd also be really nasty to leave references to deallocated memory).
1.676 + _S_free_mem_blocks.clear();
1.677 + _S_heap_size = 0;
1.678 +
1.679 + for (size_t __i = 0; __i < _STLP_NFREELISTS; ++__i) {
1.680 + _S_free_list[__i].clear();
1.681 + }
1.682 +
1.683 + // Detach list of chunks and free them all
1.684 + _Obj* __chunk = _S_chunks.clear();
1.685 + while (__chunk != 0) {
1.686 + _Obj* __next = __chunk->_M_next;
1.687 + __stlp_delete_chunck(__chunk);
1.688 + __chunk = __next;
1.689 + }
1.690 +}
1.691 +# endif
1.692 +
1.693 +#endif
1.694 +
1.695 +#if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.696 +struct __node_alloc_cleaner {
1.697 + ~__node_alloc_cleaner()
1.698 + { __node_alloc_impl::_S_dealloc_call(); }
1.699 +};
1.700 +
1.701 +# if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.702 +_STLP_VOLATILE __stl_atomic_t& _STLP_CALL
1.703 +# else
1.704 +__stl_atomic_t& _STLP_CALL
1.705 +# endif
1.706 +__node_alloc_impl::_S_alloc_counter() {
1.707 + static _AllocCounter _S_counter = 1;
1.708 + static __node_alloc_cleaner _S_node_alloc_cleaner;
1.709 + return _S_counter;
1.710 +}
1.711 +#endif
1.712 +
1.713 +#if !defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.714 +_Node_alloc_obj * _STLP_VOLATILE
1.715 +__node_alloc_impl::_S_free_list[_STLP_NFREELISTS]
1.716 += {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
1.717 +// The 16 zeros are necessary to make version 4.1 of the SunPro
1.718 +// compiler happy. Otherwise it appears to allocate too little
1.719 +// space for the array.
1.720 +#else
1.721 +_STLP_atomic_freelist __node_alloc_impl::_S_free_list[_STLP_NFREELISTS];
1.722 +_STLP_atomic_freelist __node_alloc_impl::_S_free_mem_blocks;
1.723 +#endif
1.724 +
1.725 +#if !defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.726 +char *__node_alloc_impl::_S_start_free = 0;
1.727 +char *__node_alloc_impl::_S_end_free = 0;
1.728 +#endif
1.729 +
1.730 +#if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.731 +_STLP_VOLATILE __add_atomic_t
1.732 +#else
1.733 +size_t
1.734 +#endif
1.735 +__node_alloc_impl::_S_heap_size = 0;
1.736 +
1.737 +#if defined (_STLP_DO_CLEAN_NODE_ALLOC)
1.738 +# if defined (_STLP_USE_LOCK_FREE_IMPLEMENTATION)
1.739 +_STLP_atomic_freelist __node_alloc_impl::_S_chunks;
1.740 +# else
1.741 +_Node_alloc_obj* __node_alloc_impl::_S_chunks = 0;
1.742 +# endif
1.743 +#endif
1.744 +
1.745 +void * _STLP_CALL __node_alloc::_M_allocate(size_t& __n)
1.746 +{ return __node_alloc_impl::_M_allocate(__n); }
1.747 +
1.748 +void _STLP_CALL __node_alloc::_M_deallocate(void *__p, size_t __n)
1.749 +{ __node_alloc_impl::_M_deallocate(__p, __n); }
1.750 +
1.751 +#if defined (_STLP_PTHREADS) && !defined (_STLP_NO_THREADS)
1.752 +
1.753 +# define _STLP_DATA_ALIGNMENT 8
1.754 +
1.755 +_STLP_MOVE_TO_PRIV_NAMESPACE
1.756 +
1.757 +// *******************************************************
1.758 +// __perthread_alloc implementation
1.759 +union _Pthread_alloc_obj {
1.760 + union _Pthread_alloc_obj * __free_list_link;
1.761 + char __client_data[_STLP_DATA_ALIGNMENT]; /* The client sees this. */
1.762 +};
1.763 +
1.764 +// Pthread allocators don't appear to the client to have meaningful
1.765 +// instances. We do in fact need to associate some state with each
1.766 +// thread. That state is represented by _Pthread_alloc_per_thread_state.
1.767 +
1.768 +struct _Pthread_alloc_per_thread_state {
1.769 + typedef _Pthread_alloc_obj __obj;
1.770 + enum { _S_NFREELISTS = _MAX_BYTES / _STLP_DATA_ALIGNMENT };
1.771 +
1.772 + // Free list link for list of available per thread structures.
1.773 + // When one of these becomes available for reuse due to thread
1.774 + // termination, any objects in its free list remain associated
1.775 + // with it. The whole structure may then be used by a newly
1.776 + // created thread.
1.777 + _Pthread_alloc_per_thread_state() : __next(0)
1.778 + { memset((void *)__CONST_CAST(_Pthread_alloc_obj**, __free_list), 0, (size_t)_S_NFREELISTS * sizeof(__obj *)); }
1.779 + // Returns an object of size __n, and possibly adds to size n free list.
1.780 + void *_M_refill(size_t __n);
1.781 +
1.782 + _Pthread_alloc_obj* volatile __free_list[_S_NFREELISTS];
1.783 + _Pthread_alloc_per_thread_state *__next;
1.784 + // this data member is only to be used by per_thread_allocator, which returns memory to the originating thread.
1.785 + _STLP_mutex _M_lock;
1.786 +};
1.787 +
1.788 +// Pthread-specific allocator.
1.789 +class _Pthread_alloc_impl {
1.790 +public: // but only for internal use:
1.791 + typedef _Pthread_alloc_per_thread_state __state_type;
1.792 + typedef char value_type;
1.793 +
1.794 + // Allocates a chunk for nobjs of size size. nobjs may be reduced
1.795 + // if it is inconvenient to allocate the requested number.
1.796 + static char *_S_chunk_alloc(size_t __size, size_t &__nobjs, __state_type*);
1.797 +
1.798 + enum {_S_ALIGN = _STLP_DATA_ALIGNMENT};
1.799 +
1.800 + static size_t _S_round_up(size_t __bytes)
1.801 + { return (((__bytes) + (int)_S_ALIGN - 1) & ~((int)_S_ALIGN - 1)); }
1.802 + static size_t _S_freelist_index(size_t __bytes)
1.803 + { return (((__bytes) + (int)_S_ALIGN - 1) / (int)_S_ALIGN - 1); }
1.804 +
1.805 +private:
1.806 + // Chunk allocation state. And other shared state.
1.807 + // Protected by _S_chunk_allocator_lock.
1.808 + static _STLP_STATIC_MUTEX _S_chunk_allocator_lock;
1.809 + static char *_S_start_free;
1.810 + static char *_S_end_free;
1.811 + static size_t _S_heap_size;
1.812 + static __state_type *_S_free_per_thread_states;
1.813 + static pthread_key_t _S_key;
1.814 + static bool _S_key_initialized;
1.815 + // Pthread key under which per thread state is stored.
1.816 + // Allocator instances that are currently unclaimed by any thread.
1.817 + static void _S_destructor(void *instance);
1.818 + // Function to be called on thread exit to reclaim per thread
1.819 + // state.
1.820 + static __state_type *_S_new_per_thread_state();
1.821 +public:
1.822 + // Return a recycled or new per thread state.
1.823 + static __state_type *_S_get_per_thread_state();
1.824 +private:
1.825 + // ensure that the current thread has an associated
1.826 + // per thread state.
1.827 + class _M_lock;
1.828 + friend class _M_lock;
1.829 + class _M_lock {
1.830 + public:
1.831 + _M_lock () { _S_chunk_allocator_lock._M_acquire_lock(); }
1.832 + ~_M_lock () { _S_chunk_allocator_lock._M_release_lock(); }
1.833 + };
1.834 +
1.835 +public:
1.836 +
1.837 + /* n must be > 0 */
1.838 + static void * allocate(size_t& __n);
1.839 +
1.840 + /* p may not be 0 */
1.841 + static void deallocate(void *__p, size_t __n);
1.842 +
1.843 + // boris : versions for per_thread_allocator
1.844 + /* n must be > 0 */
1.845 + static void * allocate(size_t& __n, __state_type* __a);
1.846 +
1.847 + /* p may not be 0 */
1.848 + static void deallocate(void *__p, size_t __n, __state_type* __a);
1.849 +
1.850 + static void * reallocate(void *__p, size_t __old_sz, size_t& __new_sz);
1.851 +};
1.852 +
1.853 +/* Returns an object of size n, and optionally adds to size n free list.*/
1.854 +/* We assume that n is properly aligned. */
1.855 +/* We hold the allocation lock. */
1.856 +void *_Pthread_alloc_per_thread_state::_M_refill(size_t __n) {
1.857 + typedef _Pthread_alloc_obj __obj;
1.858 + size_t __nobjs = 128;
1.859 + char * __chunk = _Pthread_alloc_impl::_S_chunk_alloc(__n, __nobjs, this);
1.860 + __obj * volatile * __my_free_list;
1.861 + __obj * __result;
1.862 + __obj * __current_obj, * __next_obj;
1.863 + size_t __i;
1.864 +
1.865 + if (1 == __nobjs) {
1.866 + return __chunk;
1.867 + }
1.868 +
1.869 + __my_free_list = __free_list + _Pthread_alloc_impl::_S_freelist_index(__n);
1.870 +
1.871 + /* Build free list in chunk */
1.872 + __result = (__obj *)__chunk;
1.873 + *__my_free_list = __next_obj = (__obj *)(__chunk + __n);
1.874 + for (__i = 1; ; ++__i) {
1.875 + __current_obj = __next_obj;
1.876 + __next_obj = (__obj *)((char *)__next_obj + __n);
1.877 + if (__nobjs - 1 == __i) {
1.878 + __current_obj -> __free_list_link = 0;
1.879 + break;
1.880 + } else {
1.881 + __current_obj -> __free_list_link = __next_obj;
1.882 + }
1.883 + }
1.884 + return __result;
1.885 +}
1.886 +
1.887 +void _Pthread_alloc_impl::_S_destructor(void *__instance) {
1.888 + _M_lock __lock_instance; // Need to acquire lock here.
1.889 + _Pthread_alloc_per_thread_state* __s = (_Pthread_alloc_per_thread_state*)__instance;
1.890 + __s -> __next = _S_free_per_thread_states;
1.891 + _S_free_per_thread_states = __s;
1.892 +}
1.893 +
1.894 +_Pthread_alloc_per_thread_state* _Pthread_alloc_impl::_S_new_per_thread_state() {
1.895 + /* lock already held here. */
1.896 + if (0 != _S_free_per_thread_states) {
1.897 + _Pthread_alloc_per_thread_state *__result = _S_free_per_thread_states;
1.898 + _S_free_per_thread_states = _S_free_per_thread_states -> __next;
1.899 + return __result;
1.900 + }
1.901 + else {
1.902 + return new _Pthread_alloc_per_thread_state;
1.903 + }
1.904 +}
1.905 +
1.906 +_Pthread_alloc_per_thread_state* _Pthread_alloc_impl::_S_get_per_thread_state() {
1.907 + int __ret_code;
1.908 + __state_type* __result;
1.909 +
1.910 + if (_S_key_initialized && (__result = (__state_type*) pthread_getspecific(_S_key)))
1.911 + return __result;
1.912 +
1.913 + /*REFERENCED*/
1.914 + _M_lock __lock_instance; // Need to acquire lock here.
1.915 + if (!_S_key_initialized) {
1.916 + if (pthread_key_create(&_S_key, _S_destructor)) {
1.917 + _STLP_THROW_BAD_ALLOC; // failed
1.918 + }
1.919 + _S_key_initialized = true;
1.920 + }
1.921 +
1.922 + __result = _S_new_per_thread_state();
1.923 + __ret_code = pthread_setspecific(_S_key, __result);
1.924 + if (__ret_code) {
1.925 + if (__ret_code == ENOMEM) {
1.926 + _STLP_THROW_BAD_ALLOC;
1.927 + } else {
1.928 + // EINVAL
1.929 + _STLP_ABORT();
1.930 + }
1.931 + }
1.932 + return __result;
1.933 +}
1.934 +
1.935 +/* We allocate memory in large chunks in order to avoid fragmenting */
1.936 +/* the malloc heap too much. */
1.937 +/* We assume that size is properly aligned. */
1.938 +char *_Pthread_alloc_impl::_S_chunk_alloc(size_t __p_size, size_t &__nobjs, _Pthread_alloc_per_thread_state *__a) {
1.939 + typedef _Pthread_alloc_obj __obj;
1.940 + {
1.941 + char * __result;
1.942 + size_t __total_bytes;
1.943 + size_t __bytes_left;
1.944 + /*REFERENCED*/
1.945 + _M_lock __lock_instance; // Acquire lock for this routine
1.946 +
1.947 + __total_bytes = __p_size * __nobjs;
1.948 + __bytes_left = _S_end_free - _S_start_free;
1.949 + if (__bytes_left >= __total_bytes) {
1.950 + __result = _S_start_free;
1.951 + _S_start_free += __total_bytes;
1.952 + return __result;
1.953 + } else if (__bytes_left >= __p_size) {
1.954 + __nobjs = __bytes_left/__p_size;
1.955 + __total_bytes = __p_size * __nobjs;
1.956 + __result = _S_start_free;
1.957 + _S_start_free += __total_bytes;
1.958 + return __result;
1.959 + } else {
1.960 + size_t __bytes_to_get = 2 * __total_bytes + _S_round_up(_S_heap_size);
1.961 + // Try to make use of the left-over piece.
1.962 + if (__bytes_left > 0) {
1.963 + __obj * volatile * __my_free_list = __a->__free_list + _S_freelist_index(__bytes_left);
1.964 + ((__obj *)_S_start_free) -> __free_list_link = *__my_free_list;
1.965 + *__my_free_list = (__obj *)_S_start_free;
1.966 + }
1.967 +# ifdef _SGI_SOURCE
1.968 + // Try to get memory that's aligned on something like a
1.969 + // cache line boundary, so as to avoid parceling out
1.970 + // parts of the same line to different threads and thus
1.971 + // possibly different processors.
1.972 + {
1.973 + const int __cache_line_size = 128; // probable upper bound
1.974 + __bytes_to_get &= ~(__cache_line_size-1);
1.975 + _S_start_free = (char *)memalign(__cache_line_size, __bytes_to_get);
1.976 + if (0 == _S_start_free) {
1.977 + _S_start_free = (char *)__malloc_alloc::allocate(__bytes_to_get);
1.978 + }
1.979 + }
1.980 +# else /* !SGI_SOURCE */
1.981 + _S_start_free = (char *)__malloc_alloc::allocate(__bytes_to_get);
1.982 +# endif
1.983 + _S_heap_size += __bytes_to_get >> 4;
1.984 + _S_end_free = _S_start_free + __bytes_to_get;
1.985 + }
1.986 + }
1.987 + // lock is released here
1.988 + return _S_chunk_alloc(__p_size, __nobjs, __a);
1.989 +}
1.990 +
1.991 +
1.992 +/* n must be > 0 */
1.993 +void *_Pthread_alloc_impl::allocate(size_t& __n) {
1.994 + typedef _Pthread_alloc_obj __obj;
1.995 + __obj * volatile * __my_free_list;
1.996 + __obj * __result;
1.997 + __state_type* __a;
1.998 +
1.999 + if (__n > _MAX_BYTES) {
1.1000 + return __malloc_alloc::allocate(__n);
1.1001 + }
1.1002 +
1.1003 + __n = _S_round_up(__n);
1.1004 + __a = _S_get_per_thread_state();
1.1005 +
1.1006 + __my_free_list = __a->__free_list + _S_freelist_index(__n);
1.1007 + __result = *__my_free_list;
1.1008 + if (__result == 0) {
1.1009 + void *__r = __a->_M_refill(__n);
1.1010 + return __r;
1.1011 + }
1.1012 + *__my_free_list = __result->__free_list_link;
1.1013 + return __result;
1.1014 +};
1.1015 +
1.1016 +/* p may not be 0 */
1.1017 +void _Pthread_alloc_impl::deallocate(void *__p, size_t __n) {
1.1018 + typedef _Pthread_alloc_obj __obj;
1.1019 + __obj *__q = (__obj *)__p;
1.1020 + __obj * volatile * __my_free_list;
1.1021 + __state_type* __a;
1.1022 +
1.1023 + if (__n > _MAX_BYTES) {
1.1024 + __malloc_alloc::deallocate(__p, __n);
1.1025 + return;
1.1026 + }
1.1027 +
1.1028 + __a = _S_get_per_thread_state();
1.1029 +
1.1030 + __my_free_list = __a->__free_list + _S_freelist_index(__n);
1.1031 + __q -> __free_list_link = *__my_free_list;
1.1032 + *__my_free_list = __q;
1.1033 +}
1.1034 +
1.1035 +// boris : versions for per_thread_allocator
1.1036 +/* n must be > 0 */
1.1037 +void *_Pthread_alloc_impl::allocate(size_t& __n, __state_type* __a) {
1.1038 + typedef _Pthread_alloc_obj __obj;
1.1039 + __obj * volatile * __my_free_list;
1.1040 + __obj * __result;
1.1041 +
1.1042 + if (__n > _MAX_BYTES) {
1.1043 + return __malloc_alloc::allocate(__n);
1.1044 + }
1.1045 + __n = _S_round_up(__n);
1.1046 +
1.1047 + // boris : here, we have to lock per thread state, as we may be getting memory from
1.1048 + // different thread pool.
1.1049 + _STLP_auto_lock __lock(__a->_M_lock);
1.1050 +
1.1051 + __my_free_list = __a->__free_list + _S_freelist_index(__n);
1.1052 + __result = *__my_free_list;
1.1053 + if (__result == 0) {
1.1054 + void *__r = __a->_M_refill(__n);
1.1055 + return __r;
1.1056 + }
1.1057 + *__my_free_list = __result->__free_list_link;
1.1058 + return __result;
1.1059 +};
1.1060 +
1.1061 +/* p may not be 0 */
1.1062 +void _Pthread_alloc_impl::deallocate(void *__p, size_t __n, __state_type* __a) {
1.1063 + typedef _Pthread_alloc_obj __obj;
1.1064 + __obj *__q = (__obj *)__p;
1.1065 + __obj * volatile * __my_free_list;
1.1066 +
1.1067 + if (__n > _MAX_BYTES) {
1.1068 + __malloc_alloc::deallocate(__p, __n);
1.1069 + return;
1.1070 + }
1.1071 +
1.1072 + // boris : here, we have to lock per thread state, as we may be returning memory from
1.1073 + // different thread.
1.1074 + _STLP_auto_lock __lock(__a->_M_lock);
1.1075 +
1.1076 + __my_free_list = __a->__free_list + _S_freelist_index(__n);
1.1077 + __q -> __free_list_link = *__my_free_list;
1.1078 + *__my_free_list = __q;
1.1079 +}
1.1080 +
1.1081 +void *_Pthread_alloc_impl::reallocate(void *__p, size_t __old_sz, size_t& __new_sz) {
1.1082 + void * __result;
1.1083 + size_t __copy_sz;
1.1084 +
1.1085 + if (__old_sz > _MAX_BYTES && __new_sz > _MAX_BYTES) {
1.1086 + return realloc(__p, __new_sz);
1.1087 + }
1.1088 +
1.1089 + if (_S_round_up(__old_sz) == _S_round_up(__new_sz)) return __p;
1.1090 + __result = allocate(__new_sz);
1.1091 + __copy_sz = __new_sz > __old_sz? __old_sz : __new_sz;
1.1092 + memcpy(__result, __p, __copy_sz);
1.1093 + deallocate(__p, __old_sz);
1.1094 + return __result;
1.1095 +}
1.1096 +
1.1097 +_Pthread_alloc_per_thread_state* _Pthread_alloc_impl::_S_free_per_thread_states = 0;
1.1098 +pthread_key_t _Pthread_alloc_impl::_S_key = 0;
1.1099 +_STLP_STATIC_MUTEX _Pthread_alloc_impl::_S_chunk_allocator_lock _STLP_MUTEX_INITIALIZER;
1.1100 +bool _Pthread_alloc_impl::_S_key_initialized = false;
1.1101 +char *_Pthread_alloc_impl::_S_start_free = 0;
1.1102 +char *_Pthread_alloc_impl::_S_end_free = 0;
1.1103 +size_t _Pthread_alloc_impl::_S_heap_size = 0;
1.1104 +
1.1105 +void * _STLP_CALL _Pthread_alloc::allocate(size_t& __n)
1.1106 +{ return _Pthread_alloc_impl::allocate(__n); }
1.1107 +void _STLP_CALL _Pthread_alloc::deallocate(void *__p, size_t __n)
1.1108 +{ _Pthread_alloc_impl::deallocate(__p, __n); }
1.1109 +void * _STLP_CALL _Pthread_alloc::allocate(size_t& __n, __state_type* __a)
1.1110 +{ return _Pthread_alloc_impl::allocate(__n, __a); }
1.1111 +void _STLP_CALL _Pthread_alloc::deallocate(void *__p, size_t __n, __state_type* __a)
1.1112 +{ _Pthread_alloc_impl::deallocate(__p, __n, __a); }
1.1113 +void * _STLP_CALL _Pthread_alloc::reallocate(void *__p, size_t __old_sz, size_t& __new_sz)
1.1114 +{ return _Pthread_alloc_impl::reallocate(__p, __old_sz, __new_sz); }
1.1115 +_Pthread_alloc_per_thread_state* _STLP_CALL _Pthread_alloc::_S_get_per_thread_state()
1.1116 +{ return _Pthread_alloc_impl::_S_get_per_thread_state(); }
1.1117 +
1.1118 +_STLP_MOVE_TO_STD_NAMESPACE
1.1119 +
1.1120 +#endif
1.1121 +
1.1122 +_STLP_END_NAMESPACE
1.1123 +
1.1124 +#undef _S_FREELIST_INDEX
