1 /* ===---------- emutls.c - Implements __emutls_get_address ---------------===
3 * The LLVM Compiler Infrastructure
5 * This file is dual licensed under the MIT and the University of Illinois Open
6 * Source Licenses. See LICENSE.TXT for details.
8 * ===----------------------------------------------------------------------===
18 /* There are 4 pthread key cleanup rounds on Bionic. Delay emutls deallocation
19 to round 2. We need to delay deallocation because:
20 - Android versions older than M lack __cxa_thread_atexit_impl, so apps
21 use a pthread key destructor to call C++ destructors.
22 - Apps might use __thread/thread_local variables in pthread destructors.
23 We can't wait until the final two rounds, because jemalloc needs two rounds
24 after the final malloc/free call to free its thread-specific data (see
25 https://reviews.llvm.org/D46978#1107507). */
26 #define EMUTLS_SKIP_DESTRUCTOR_ROUNDS 1
28 #define EMUTLS_SKIP_DESTRUCTOR_ROUNDS 0
31 typedef struct emutls_address_array {
32 uintptr_t skip_destructor_rounds;
33 uintptr_t size; /* number of elements in the 'data' array */
35 } emutls_address_array;
37 static void emutls_shutdown(emutls_address_array *array);
43 static pthread_mutex_t emutls_mutex = PTHREAD_MUTEX_INITIALIZER;
44 static pthread_key_t emutls_pthread_key;
46 typedef unsigned int gcc_word __attribute__((mode(word)));
47 typedef unsigned int gcc_pointer __attribute__((mode(pointer)));
49 /* Default is not to use posix_memalign, so systems like Android
50 * can use thread local data without heavier POSIX memory allocators.
52 #ifndef EMUTLS_USE_POSIX_MEMALIGN
53 #define EMUTLS_USE_POSIX_MEMALIGN 0
56 static __inline void *emutls_memalign_alloc(size_t align, size_t size) {
58 #if EMUTLS_USE_POSIX_MEMALIGN
59 if (posix_memalign(&base, align, size) != 0)
62 #define EXTRA_ALIGN_PTR_BYTES (align - 1 + sizeof(void*))
64 if ((object = (char*)malloc(EXTRA_ALIGN_PTR_BYTES + size)) == NULL)
66 base = (void*)(((uintptr_t)(object + EXTRA_ALIGN_PTR_BYTES))
67 & ~(uintptr_t)(align - 1));
69 ((void**)base)[-1] = object;
74 static __inline void emutls_memalign_free(void *base) {
75 #if EMUTLS_USE_POSIX_MEMALIGN
78 /* The mallocated address is in ((void**)base)[-1] */
79 free(((void**)base)[-1]);
83 static __inline void emutls_setspecific(emutls_address_array *value) {
84 pthread_setspecific(emutls_pthread_key, (void*) value);
87 static __inline emutls_address_array* emutls_getspecific() {
88 return (emutls_address_array*) pthread_getspecific(emutls_pthread_key);
91 static void emutls_key_destructor(void* ptr) {
92 emutls_address_array *array = (emutls_address_array*)ptr;
93 if (array->skip_destructor_rounds > 0) {
94 /* emutls is deallocated using a pthread key destructor. These
95 * destructors are called in several rounds to accommodate destructor
96 * functions that (re)initialize key values with pthread_setspecific.
97 * Delay the emutls deallocation to accommodate other end-of-thread
98 * cleanup tasks like calling thread_local destructors (e.g. the
99 * __cxa_thread_atexit fallback in libc++abi).
101 array->skip_destructor_rounds--;
102 emutls_setspecific(array);
104 emutls_shutdown(array);
109 static __inline void emutls_init(void) {
110 if (pthread_key_create(&emutls_pthread_key, emutls_key_destructor) != 0)
114 static __inline void emutls_init_once(void) {
115 static pthread_once_t once = PTHREAD_ONCE_INIT;
116 pthread_once(&once, emutls_init);
119 static __inline void emutls_lock() {
120 pthread_mutex_lock(&emutls_mutex);
123 static __inline void emutls_unlock() {
124 pthread_mutex_unlock(&emutls_mutex);
134 static LPCRITICAL_SECTION emutls_mutex;
135 static DWORD emutls_tls_index = TLS_OUT_OF_INDEXES;
137 typedef uintptr_t gcc_word;
138 typedef void * gcc_pointer;
140 static void win_error(DWORD last_err, const char *hint) {
142 if (FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER |
143 FORMAT_MESSAGE_FROM_SYSTEM |
144 FORMAT_MESSAGE_MAX_WIDTH_MASK,
145 NULL, last_err, 0, (LPSTR)&buffer, 1, NULL)) {
146 fprintf(stderr, "Windows error: %s\n", buffer);
148 fprintf(stderr, "Unkown Windows error: %s\n", hint);
153 static __inline void win_abort(DWORD last_err, const char *hint) {
154 win_error(last_err, hint);
158 static __inline void *emutls_memalign_alloc(size_t align, size_t size) {
159 void *base = _aligned_malloc(size, align);
161 win_abort(GetLastError(), "_aligned_malloc");
165 static __inline void emutls_memalign_free(void *base) {
169 static void emutls_exit(void) {
171 DeleteCriticalSection(emutls_mutex);
172 _aligned_free(emutls_mutex);
175 if (emutls_tls_index != TLS_OUT_OF_INDEXES) {
176 emutls_shutdown((emutls_address_array*)TlsGetValue(emutls_tls_index));
177 TlsFree(emutls_tls_index);
178 emutls_tls_index = TLS_OUT_OF_INDEXES;
182 #pragma warning (push)
183 #pragma warning (disable : 4100)
184 static BOOL CALLBACK emutls_init(PINIT_ONCE p0, PVOID p1, PVOID *p2) {
185 emutls_mutex = (LPCRITICAL_SECTION)_aligned_malloc(sizeof(CRITICAL_SECTION), 16);
187 win_error(GetLastError(), "_aligned_malloc");
190 InitializeCriticalSection(emutls_mutex);
192 emutls_tls_index = TlsAlloc();
193 if (emutls_tls_index == TLS_OUT_OF_INDEXES) {
195 win_error(GetLastError(), "TlsAlloc");
198 atexit(&emutls_exit);
202 static __inline void emutls_init_once(void) {
203 static INIT_ONCE once;
204 InitOnceExecuteOnce(&once, emutls_init, NULL, NULL);
207 static __inline void emutls_lock() {
208 EnterCriticalSection(emutls_mutex);
211 static __inline void emutls_unlock() {
212 LeaveCriticalSection(emutls_mutex);
215 static __inline void emutls_setspecific(emutls_address_array *value) {
216 if (TlsSetValue(emutls_tls_index, (LPVOID) value) == 0)
217 win_abort(GetLastError(), "TlsSetValue");
220 static __inline emutls_address_array* emutls_getspecific() {
221 LPVOID value = TlsGetValue(emutls_tls_index);
223 const DWORD err = GetLastError();
224 if (err != ERROR_SUCCESS)
225 win_abort(err, "TlsGetValue");
227 return (emutls_address_array*) value;
230 /* Provide atomic load/store functions for emutls_get_index if built with MSVC.
232 #if !defined(__ATOMIC_RELEASE)
235 enum { __ATOMIC_ACQUIRE = 2, __ATOMIC_RELEASE = 3 };
237 static __inline uintptr_t __atomic_load_n(void *ptr, unsigned type) {
238 assert(type == __ATOMIC_ACQUIRE);
239 // These return the previous value - but since we do an OR with 0,
240 // it's equivalent to a plain load.
242 return InterlockedOr64(ptr, 0);
244 return InterlockedOr(ptr, 0);
248 static __inline void __atomic_store_n(void *ptr, uintptr_t val, unsigned type) {
249 assert(type == __ATOMIC_RELEASE);
250 InterlockedExchangePointer((void *volatile *)ptr, (void *)val);
253 #endif /* __ATOMIC_RELEASE */
255 #pragma warning (pop)
259 static size_t emutls_num_object = 0; /* number of allocated TLS objects */
261 /* Free the allocated TLS data
263 static void emutls_shutdown(emutls_address_array *array) {
266 for (i = 0; i < array->size; ++i) {
268 emutls_memalign_free(array->data[i]);
273 /* For every TLS variable xyz,
274 * there is one __emutls_control variable named __emutls_v.xyz.
275 * If xyz has non-zero initial value, __emutls_v.xyz's "value"
276 * will point to __emutls_t.xyz, which has the initial value.
278 typedef struct __emutls_control {
279 /* Must use gcc_word here, instead of size_t, to match GCC. When
280 gcc_word is larger than size_t, the upper extra bits are all
281 zeros. We can use variables of size_t to operate on size and
283 gcc_word size; /* size of the object in bytes */
284 gcc_word align; /* alignment of the object in bytes */
286 uintptr_t index; /* data[index-1] is the object address */
287 void* address; /* object address, when in single thread env */
289 void* value; /* null or non-zero initial value for the object */
292 /* Emulated TLS objects are always allocated at run-time. */
293 static __inline void *emutls_allocate_object(__emutls_control *control) {
294 /* Use standard C types, check with gcc's emutls.o. */
295 COMPILE_TIME_ASSERT(sizeof(uintptr_t) == sizeof(gcc_pointer));
296 COMPILE_TIME_ASSERT(sizeof(uintptr_t) == sizeof(void*));
298 size_t size = control->size;
299 size_t align = control->align;
301 if (align < sizeof(void*))
302 align = sizeof(void*);
303 /* Make sure that align is power of 2. */
304 if ((align & (align - 1)) != 0)
307 base = emutls_memalign_alloc(align, size);
309 memcpy(base, control->value, size);
311 memset(base, 0, size);
316 /* Returns control->object.index; set index if not allocated yet. */
317 static __inline uintptr_t emutls_get_index(__emutls_control *control) {
318 uintptr_t index = __atomic_load_n(&control->object.index, __ATOMIC_ACQUIRE);
322 index = control->object.index;
324 index = ++emutls_num_object;
325 __atomic_store_n(&control->object.index, index, __ATOMIC_RELEASE);
332 /* Updates newly allocated thread local emutls_address_array. */
333 static __inline void emutls_check_array_set_size(emutls_address_array *array,
338 emutls_setspecific(array);
341 /* Returns the new 'data' array size, number of elements,
342 * which must be no smaller than the given index.
344 static __inline uintptr_t emutls_new_data_array_size(uintptr_t index) {
345 /* Need to allocate emutls_address_array with extra slots
346 * to store the header.
347 * Round up the emutls_address_array size to multiple of 16.
349 uintptr_t header_words = sizeof(emutls_address_array) / sizeof(void *);
350 return ((index + header_words + 15) & ~((uintptr_t)15)) - header_words;
353 /* Returns the size in bytes required for an emutls_address_array with
354 * N number of elements for data field.
356 static __inline uintptr_t emutls_asize(uintptr_t N) {
357 return N * sizeof(void *) + sizeof(emutls_address_array);
360 /* Returns the thread local emutls_address_array.
361 * Extends its size if necessary to hold address at index.
363 static __inline emutls_address_array *
364 emutls_get_address_array(uintptr_t index) {
365 emutls_address_array* array = emutls_getspecific();
367 uintptr_t new_size = emutls_new_data_array_size(index);
368 array = (emutls_address_array*) malloc(emutls_asize(new_size));
370 memset(array->data, 0, new_size * sizeof(void*));
371 array->skip_destructor_rounds = EMUTLS_SKIP_DESTRUCTOR_ROUNDS;
373 emutls_check_array_set_size(array, new_size);
374 } else if (index > array->size) {
375 uintptr_t orig_size = array->size;
376 uintptr_t new_size = emutls_new_data_array_size(index);
377 array = (emutls_address_array*) realloc(array, emutls_asize(new_size));
379 memset(array->data + orig_size, 0,
380 (new_size - orig_size) * sizeof(void*));
381 emutls_check_array_set_size(array, new_size);
386 void* __emutls_get_address(__emutls_control* control) {
387 uintptr_t index = emutls_get_index(control);
388 emutls_address_array* array = emutls_get_address_array(index--);
389 if (array->data[index] == NULL)
390 array->data[index] = emutls_allocate_object(control);
391 return array->data[index];