/* * kmp_threadprivate.cpp -- OpenMP threadprivate support library */ //===----------------------------------------------------------------------===// // // The LLVM Compiler Infrastructure // // This file is dual licensed under the MIT and the University of Illinois Open // Source Licenses. See LICENSE.txt for details. // //===----------------------------------------------------------------------===// #include "kmp.h" #include "kmp_i18n.h" #include "kmp_itt.h" #define USE_CHECKS_COMMON #define KMP_INLINE_SUBR 1 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr, void *data_addr, size_t pc_size); struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr, void *data_addr, size_t pc_size); struct shared_table __kmp_threadprivate_d_table; static #ifdef KMP_INLINE_SUBR __forceinline #endif struct private_common * __kmp_threadprivate_find_task_common(struct common_table *tbl, int gtid, void *pc_addr) { struct private_common *tn; #ifdef KMP_TASK_COMMON_DEBUG KC_TRACE(10, ("__kmp_threadprivate_find_task_common: thread#%d, called with " "address %p\n", gtid, pc_addr)); dump_list(); #endif for (tn = tbl->data[KMP_HASH(pc_addr)]; tn; tn = tn->next) { if (tn->gbl_addr == pc_addr) { #ifdef KMP_TASK_COMMON_DEBUG KC_TRACE(10, ("__kmp_threadprivate_find_task_common: thread#%d, found " "node %p on list\n", gtid, pc_addr)); #endif return tn; } } return 0; } static #ifdef KMP_INLINE_SUBR __forceinline #endif struct shared_common * __kmp_find_shared_task_common(struct shared_table *tbl, int gtid, void *pc_addr) { struct shared_common *tn; for (tn = tbl->data[KMP_HASH(pc_addr)]; tn; tn = tn->next) { if (tn->gbl_addr == pc_addr) { #ifdef KMP_TASK_COMMON_DEBUG KC_TRACE( 10, ("__kmp_find_shared_task_common: thread#%d, found node %p on list\n", gtid, pc_addr)); #endif return tn; } } return 0; } // Create a template for the data initialized storage. Either the template is // NULL indicating zero fill, or the template is a copy of the original data. static struct private_data *__kmp_init_common_data(void *pc_addr, size_t pc_size) { struct private_data *d; size_t i; char *p; d = (struct private_data *)__kmp_allocate(sizeof(struct private_data)); /* d->data = 0; // AC: commented out because __kmp_allocate zeroes the memory d->next = 0; */ d->size = pc_size; d->more = 1; p = (char *)pc_addr; for (i = pc_size; i > 0; --i) { if (*p++ != '\0') { d->data = __kmp_allocate(pc_size); KMP_MEMCPY(d->data, pc_addr, pc_size); break; } } return d; } // Initialize the data area from the template. static void __kmp_copy_common_data(void *pc_addr, struct private_data *d) { char *addr = (char *)pc_addr; int i, offset; for (offset = 0; d != 0; d = d->next) { for (i = d->more; i > 0; --i) { if (d->data == 0) memset(&addr[offset], '\0', d->size); else KMP_MEMCPY(&addr[offset], d->data, d->size); offset += d->size; } } } /* we are called from __kmp_serial_initialize() with __kmp_initz_lock held. */ void __kmp_common_initialize(void) { if (!TCR_4(__kmp_init_common)) { int q; #ifdef KMP_DEBUG int gtid; #endif __kmp_threadpriv_cache_list = NULL; #ifdef KMP_DEBUG /* verify the uber masters were initialized */ for (gtid = 0; gtid < __kmp_threads_capacity; gtid++) if (__kmp_root[gtid]) { KMP_DEBUG_ASSERT(__kmp_root[gtid]->r.r_uber_thread); for (q = 0; q < KMP_HASH_TABLE_SIZE; ++q) KMP_DEBUG_ASSERT( !__kmp_root[gtid]->r.r_uber_thread->th.th_pri_common->data[q]); /* __kmp_root[ gitd ]-> r.r_uber_thread -> * th.th_pri_common -> data[ q ] = 0;*/ } #endif /* KMP_DEBUG */ for (q = 0; q < KMP_HASH_TABLE_SIZE; ++q) __kmp_threadprivate_d_table.data[q] = 0; TCW_4(__kmp_init_common, TRUE); } } /* Call all destructors for threadprivate data belonging to all threads. Currently unused! */ void __kmp_common_destroy(void) { if (TCR_4(__kmp_init_common)) { int q; TCW_4(__kmp_init_common, FALSE); for (q = 0; q < KMP_HASH_TABLE_SIZE; ++q) { int gtid; struct private_common *tn; struct shared_common *d_tn; /* C++ destructors need to be called once per thread before exiting. Don't call destructors for master thread though unless we used copy constructor */ for (d_tn = __kmp_threadprivate_d_table.data[q]; d_tn; d_tn = d_tn->next) { if (d_tn->is_vec) { if (d_tn->dt.dtorv != 0) { for (gtid = 0; gtid < __kmp_all_nth; ++gtid) { if (__kmp_threads[gtid]) { if ((__kmp_foreign_tp) ? (!KMP_INITIAL_GTID(gtid)) : (!KMP_UBER_GTID(gtid))) { tn = __kmp_threadprivate_find_task_common( __kmp_threads[gtid]->th.th_pri_common, gtid, d_tn->gbl_addr); if (tn) { (*d_tn->dt.dtorv)(tn->par_addr, d_tn->vec_len); } } } } if (d_tn->obj_init != 0) { (*d_tn->dt.dtorv)(d_tn->obj_init, d_tn->vec_len); } } } else { if (d_tn->dt.dtor != 0) { for (gtid = 0; gtid < __kmp_all_nth; ++gtid) { if (__kmp_threads[gtid]) { if ((__kmp_foreign_tp) ? (!KMP_INITIAL_GTID(gtid)) : (!KMP_UBER_GTID(gtid))) { tn = __kmp_threadprivate_find_task_common( __kmp_threads[gtid]->th.th_pri_common, gtid, d_tn->gbl_addr); if (tn) { (*d_tn->dt.dtor)(tn->par_addr); } } } } if (d_tn->obj_init != 0) { (*d_tn->dt.dtor)(d_tn->obj_init); } } } } __kmp_threadprivate_d_table.data[q] = 0; } } } /* Call all destructors for threadprivate data belonging to this thread */ void __kmp_common_destroy_gtid(int gtid) { struct private_common *tn; struct shared_common *d_tn; KC_TRACE(10, ("__kmp_common_destroy_gtid: T#%d called\n", gtid)); if ((__kmp_foreign_tp) ? (!KMP_INITIAL_GTID(gtid)) : (!KMP_UBER_GTID(gtid))) { if (TCR_4(__kmp_init_common)) { /* Cannot do this here since not all threads have destroyed their data */ /* TCW_4(__kmp_init_common, FALSE); */ for (tn = __kmp_threads[gtid]->th.th_pri_head; tn; tn = tn->link) { d_tn = __kmp_find_shared_task_common(&__kmp_threadprivate_d_table, gtid, tn->gbl_addr); KMP_DEBUG_ASSERT(d_tn); if (d_tn->is_vec) { if (d_tn->dt.dtorv != 0) { (void)(*d_tn->dt.dtorv)(tn->par_addr, d_tn->vec_len); } if (d_tn->obj_init != 0) { (void)(*d_tn->dt.dtorv)(d_tn->obj_init, d_tn->vec_len); } } else { if (d_tn->dt.dtor != 0) { (void)(*d_tn->dt.dtor)(tn->par_addr); } if (d_tn->obj_init != 0) { (void)(*d_tn->dt.dtor)(d_tn->obj_init); } } } KC_TRACE(30, ("__kmp_common_destroy_gtid: T#%d threadprivate destructors " "complete\n", gtid)); } } } #ifdef KMP_TASK_COMMON_DEBUG static void dump_list(void) { int p, q; for (p = 0; p < __kmp_all_nth; ++p) { if (!__kmp_threads[p]) continue; for (q = 0; q < KMP_HASH_TABLE_SIZE; ++q) { if (__kmp_threads[p]->th.th_pri_common->data[q]) { struct private_common *tn; KC_TRACE(10, ("\tdump_list: gtid:%d addresses\n", p)); for (tn = __kmp_threads[p]->th.th_pri_common->data[q]; tn; tn = tn->next) { KC_TRACE(10, ("\tdump_list: THREADPRIVATE: Serial %p -> Parallel %p\n", tn->gbl_addr, tn->par_addr)); } } } } } #endif /* KMP_TASK_COMMON_DEBUG */ // NOTE: this routine is to be called only from the serial part of the program. void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr, void *data_addr, size_t pc_size) { struct shared_common **lnk_tn, *d_tn; KMP_DEBUG_ASSERT(__kmp_threads[gtid] && __kmp_threads[gtid]->th.th_root->r.r_active == 0); d_tn = __kmp_find_shared_task_common(&__kmp_threadprivate_d_table, gtid, pc_addr); if (d_tn == 0) { d_tn = (struct shared_common *)__kmp_allocate(sizeof(struct shared_common)); d_tn->gbl_addr = pc_addr; d_tn->pod_init = __kmp_init_common_data(data_addr, pc_size); /* d_tn->obj_init = 0; // AC: commented out because __kmp_allocate zeroes the memory d_tn->ct.ctor = 0; d_tn->cct.cctor = 0;; d_tn->dt.dtor = 0; d_tn->is_vec = FALSE; d_tn->vec_len = 0L; */ d_tn->cmn_size = pc_size; __kmp_acquire_lock(&__kmp_global_lock, gtid); lnk_tn = &(__kmp_threadprivate_d_table.data[KMP_HASH(pc_addr)]); d_tn->next = *lnk_tn; *lnk_tn = d_tn; __kmp_release_lock(&__kmp_global_lock, gtid); } } struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr, void *data_addr, size_t pc_size) { struct private_common *tn, **tt; struct shared_common *d_tn; /* +++++++++ START OF CRITICAL SECTION +++++++++ */ __kmp_acquire_lock(&__kmp_global_lock, gtid); tn = (struct private_common *)__kmp_allocate(sizeof(struct private_common)); tn->gbl_addr = pc_addr; d_tn = __kmp_find_shared_task_common( &__kmp_threadprivate_d_table, gtid, pc_addr); /* Only the MASTER data table exists. */ if (d_tn != 0) { /* This threadprivate variable has already been seen. */ if (d_tn->pod_init == 0 && d_tn->obj_init == 0) { d_tn->cmn_size = pc_size; if (d_tn->is_vec) { if (d_tn->ct.ctorv != 0) { /* Construct from scratch so no prototype exists */ d_tn->obj_init = 0; } else if (d_tn->cct.cctorv != 0) { /* Now data initialize the prototype since it was previously * registered */ d_tn->obj_init = (void *)__kmp_allocate(d_tn->cmn_size); (void)(*d_tn->cct.cctorv)(d_tn->obj_init, pc_addr, d_tn->vec_len); } else { d_tn->pod_init = __kmp_init_common_data(data_addr, d_tn->cmn_size); } } else { if (d_tn->ct.ctor != 0) { /* Construct from scratch so no prototype exists */ d_tn->obj_init = 0; } else if (d_tn->cct.cctor != 0) { /* Now data initialize the prototype since it was previously registered */ d_tn->obj_init = (void *)__kmp_allocate(d_tn->cmn_size); (void)(*d_tn->cct.cctor)(d_tn->obj_init, pc_addr); } else { d_tn->pod_init = __kmp_init_common_data(data_addr, d_tn->cmn_size); } } } } else { struct shared_common **lnk_tn; d_tn = (struct shared_common *)__kmp_allocate(sizeof(struct shared_common)); d_tn->gbl_addr = pc_addr; d_tn->cmn_size = pc_size; d_tn->pod_init = __kmp_init_common_data(data_addr, pc_size); /* d_tn->obj_init = 0; // AC: commented out because __kmp_allocate zeroes the memory d_tn->ct.ctor = 0; d_tn->cct.cctor = 0; d_tn->dt.dtor = 0; d_tn->is_vec = FALSE; d_tn->vec_len = 0L; */ lnk_tn = &(__kmp_threadprivate_d_table.data[KMP_HASH(pc_addr)]); d_tn->next = *lnk_tn; *lnk_tn = d_tn; } tn->cmn_size = d_tn->cmn_size; if ((__kmp_foreign_tp) ? (KMP_INITIAL_GTID(gtid)) : (KMP_UBER_GTID(gtid))) { tn->par_addr = (void *)pc_addr; } else { tn->par_addr = (void *)__kmp_allocate(tn->cmn_size); } __kmp_release_lock(&__kmp_global_lock, gtid); /* +++++++++ END OF CRITICAL SECTION +++++++++ */ #ifdef USE_CHECKS_COMMON if (pc_size > d_tn->cmn_size) { KC_TRACE( 10, ("__kmp_threadprivate_insert: THREADPRIVATE: %p (%" KMP_UINTPTR_SPEC " ,%" KMP_UINTPTR_SPEC ")\n", pc_addr, pc_size, d_tn->cmn_size)); KMP_FATAL(TPCommonBlocksInconsist); } #endif /* USE_CHECKS_COMMON */ tt = &(__kmp_threads[gtid]->th.th_pri_common->data[KMP_HASH(pc_addr)]); #ifdef KMP_TASK_COMMON_DEBUG if (*tt != 0) { KC_TRACE( 10, ("__kmp_threadprivate_insert: WARNING! thread#%d: collision on %p\n", gtid, pc_addr)); } #endif tn->next = *tt; *tt = tn; #ifdef KMP_TASK_COMMON_DEBUG KC_TRACE(10, ("__kmp_threadprivate_insert: thread#%d, inserted node %p on list\n", gtid, pc_addr)); dump_list(); #endif /* Link the node into a simple list */ tn->link = __kmp_threads[gtid]->th.th_pri_head; __kmp_threads[gtid]->th.th_pri_head = tn; #ifdef BUILD_TV __kmp_tv_threadprivate_store(__kmp_threads[gtid], tn->gbl_addr, tn->par_addr); #endif if ((__kmp_foreign_tp) ? (KMP_INITIAL_GTID(gtid)) : (KMP_UBER_GTID(gtid))) return tn; /* if C++ object with copy constructor, use it; * else if C++ object with constructor, use it for the non-master copies only; * else use pod_init and memcpy * * C++ constructors need to be called once for each non-master thread on * allocate * C++ copy constructors need to be called once for each thread on allocate */ /* C++ object with constructors/destructors; don't call constructors for master thread though */ if (d_tn->is_vec) { if (d_tn->ct.ctorv != 0) { (void)(*d_tn->ct.ctorv)(tn->par_addr, d_tn->vec_len); } else if (d_tn->cct.cctorv != 0) { (void)(*d_tn->cct.cctorv)(tn->par_addr, d_tn->obj_init, d_tn->vec_len); } else if (tn->par_addr != tn->gbl_addr) { __kmp_copy_common_data(tn->par_addr, d_tn->pod_init); } } else { if (d_tn->ct.ctor != 0) { (void)(*d_tn->ct.ctor)(tn->par_addr); } else if (d_tn->cct.cctor != 0) { (void)(*d_tn->cct.cctor)(tn->par_addr, d_tn->obj_init); } else if (tn->par_addr != tn->gbl_addr) { __kmp_copy_common_data(tn->par_addr, d_tn->pod_init); } } /* !BUILD_OPENMP_C if (tn->par_addr != tn->gbl_addr) __kmp_copy_common_data( tn->par_addr, d_tn->pod_init ); */ return tn; } /* ------------------------------------------------------------------------ */ /* We are currently parallel, and we know the thread id. */ /* ------------------------------------------------------------------------ */ /*! @ingroup THREADPRIVATE @param loc source location information @param data pointer to data being privatized @param ctor pointer to constructor function for data @param cctor pointer to copy constructor function for data @param dtor pointer to destructor function for data Register constructors and destructors for thread private data. This function is called when executing in parallel, when we know the thread id. */ void __kmpc_threadprivate_register(ident_t *loc, void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor) { struct shared_common *d_tn, **lnk_tn; KC_TRACE(10, ("__kmpc_threadprivate_register: called\n")); #ifdef USE_CHECKS_COMMON /* copy constructor must be zero for current code gen (Nov 2002 - jph) */ KMP_ASSERT(cctor == 0); #endif /* USE_CHECKS_COMMON */ /* Only the global data table exists. */ d_tn = __kmp_find_shared_task_common(&__kmp_threadprivate_d_table, -1, data); if (d_tn == 0) { d_tn = (struct shared_common *)__kmp_allocate(sizeof(struct shared_common)); d_tn->gbl_addr = data; d_tn->ct.ctor = ctor; d_tn->cct.cctor = cctor; d_tn->dt.dtor = dtor; /* d_tn->is_vec = FALSE; // AC: commented out because __kmp_allocate zeroes the memory d_tn->vec_len = 0L; d_tn->obj_init = 0; d_tn->pod_init = 0; */ lnk_tn = &(__kmp_threadprivate_d_table.data[KMP_HASH(data)]); d_tn->next = *lnk_tn; *lnk_tn = d_tn; } } void *__kmpc_threadprivate(ident_t *loc, kmp_int32 global_tid, void *data, size_t size) { void *ret; struct private_common *tn; KC_TRACE(10, ("__kmpc_threadprivate: T#%d called\n", global_tid)); #ifdef USE_CHECKS_COMMON if (!__kmp_init_serial) KMP_FATAL(RTLNotInitialized); #endif /* USE_CHECKS_COMMON */ if (!__kmp_threads[global_tid]->th.th_root->r.r_active && !__kmp_foreign_tp) { /* The parallel address will NEVER overlap with the data_address */ /* dkp: 3rd arg to kmp_threadprivate_insert_private_data() is the * data_address; use data_address = data */ KC_TRACE(20, ("__kmpc_threadprivate: T#%d inserting private data\n", global_tid)); kmp_threadprivate_insert_private_data(global_tid, data, data, size); ret = data; } else { KC_TRACE( 50, ("__kmpc_threadprivate: T#%d try to find private data at address %p\n", global_tid, data)); tn = __kmp_threadprivate_find_task_common( __kmp_threads[global_tid]->th.th_pri_common, global_tid, data); if (tn) { KC_TRACE(20, ("__kmpc_threadprivate: T#%d found data\n", global_tid)); #ifdef USE_CHECKS_COMMON if ((size_t)size > tn->cmn_size) { KC_TRACE(10, ("THREADPRIVATE: %p (%" KMP_UINTPTR_SPEC " ,%" KMP_UINTPTR_SPEC ")\n", data, size, tn->cmn_size)); KMP_FATAL(TPCommonBlocksInconsist); } #endif /* USE_CHECKS_COMMON */ } else { /* The parallel address will NEVER overlap with the data_address */ /* dkp: 3rd arg to kmp_threadprivate_insert() is the data_address; use * data_address = data */ KC_TRACE(20, ("__kmpc_threadprivate: T#%d inserting data\n", global_tid)); tn = kmp_threadprivate_insert(global_tid, data, data, size); } ret = tn->par_addr; } KC_TRACE(10, ("__kmpc_threadprivate: T#%d exiting; return value = %p\n", global_tid, ret)); return ret; } /*! @ingroup THREADPRIVATE @param loc source location information @param global_tid global thread number @param data pointer to data to privatize @param size size of data to privatize @param cache pointer to cache @return pointer to private storage Allocate private storage for threadprivate data. */ void * __kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, // gtid. void *data, // Pointer to original global variable. size_t size, // Size of original global variable. void ***cache) { KC_TRACE(10, ("__kmpc_threadprivate_cached: T#%d called with cache: %p, " "address: %p, size: %" KMP_SIZE_T_SPEC "\n", global_tid, *cache, data, size)); if (TCR_PTR(*cache) == 0) { __kmp_acquire_lock(&__kmp_global_lock, global_tid); if (TCR_PTR(*cache) == 0) { __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock); __kmp_tp_cached = 1; __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock); void **my_cache; KMP_ITT_IGNORE( my_cache = (void **)__kmp_allocate( sizeof(void *) * __kmp_tp_capacity + sizeof(kmp_cached_addr_t));); // No need to zero the allocated memory; __kmp_allocate does that. KC_TRACE( 50, ("__kmpc_threadprivate_cached: T#%d allocated cache at address %p\n", global_tid, my_cache)); /* TODO: free all this memory in __kmp_common_destroy using * __kmp_threadpriv_cache_list */ /* Add address of mycache to linked list for cleanup later */ kmp_cached_addr_t *tp_cache_addr; tp_cache_addr = (kmp_cached_addr_t *)&my_cache[__kmp_tp_capacity]; tp_cache_addr->addr = my_cache; tp_cache_addr->next = __kmp_threadpriv_cache_list; __kmp_threadpriv_cache_list = tp_cache_addr; KMP_MB(); TCW_PTR(*cache, my_cache); KMP_MB(); } __kmp_release_lock(&__kmp_global_lock, global_tid); } void *ret; if ((ret = TCR_PTR((*cache)[global_tid])) == 0) { ret = __kmpc_threadprivate(loc, global_tid, data, (size_t)size); TCW_PTR((*cache)[global_tid], ret); } KC_TRACE(10, ("__kmpc_threadprivate_cached: T#%d exiting; return value = %p\n", global_tid, ret)); return ret; } /*! @ingroup THREADPRIVATE @param loc source location information @param data pointer to data being privatized @param ctor pointer to constructor function for data @param cctor pointer to copy constructor function for data @param dtor pointer to destructor function for data @param vector_length length of the vector (bytes or elements?) Register vector constructors and destructors for thread private data. */ void __kmpc_threadprivate_register_vec(ident_t *loc, void *data, kmpc_ctor_vec ctor, kmpc_cctor_vec cctor, kmpc_dtor_vec dtor, size_t vector_length) { struct shared_common *d_tn, **lnk_tn; KC_TRACE(10, ("__kmpc_threadprivate_register_vec: called\n")); #ifdef USE_CHECKS_COMMON /* copy constructor must be zero for current code gen (Nov 2002 - jph) */ KMP_ASSERT(cctor == 0); #endif /* USE_CHECKS_COMMON */ d_tn = __kmp_find_shared_task_common( &__kmp_threadprivate_d_table, -1, data); /* Only the global data table exists. */ if (d_tn == 0) { d_tn = (struct shared_common *)__kmp_allocate(sizeof(struct shared_common)); d_tn->gbl_addr = data; d_tn->ct.ctorv = ctor; d_tn->cct.cctorv = cctor; d_tn->dt.dtorv = dtor; d_tn->is_vec = TRUE; d_tn->vec_len = (size_t)vector_length; /* d_tn->obj_init = 0; // AC: commented out because __kmp_allocate zeroes the memory d_tn->pod_init = 0; */ lnk_tn = &(__kmp_threadprivate_d_table.data[KMP_HASH(data)]); d_tn->next = *lnk_tn; *lnk_tn = d_tn; } }