Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Parmap: add missing atomic load/store, and convert legacy builtins.
[simgrid.git] / src / include / xbt / parmap.hpp
index dd5037f..77d13e0 100644 (file)
@@ -90,7 +90,6 @@ private:
      */
     virtual void worker_wait(unsigned) = 0;
 
-  protected:
     Parmap<T>& parmap;
   };
 
@@ -225,7 +224,7 @@ template <typename T> void Parmap<T>::apply(void (*fun)(T), const std::vector<T>
  */
 template <typename T> boost::optional<T> Parmap<T>::next()
 {
-  unsigned index = this->index++;
+  unsigned index = this->index.fetch_add(1, std::memory_order_relaxed);
   if (index < this->data->size())
     return (*this->data)[index];
   else
@@ -237,11 +236,11 @@ template <typename T> boost::optional<T> Parmap<T>::next()
  */
 template <typename T> void Parmap<T>::work()
 {
-  unsigned index = this->index++;
   unsigned length = this->data->size();
+  unsigned index  = this->index.fetch_add(1, std::memory_order_relaxed);
   while (index < length) {
     this->fun((*this->data)[index]);
-    index = this->index++;
+    index = this->index.fetch_add(1, std::memory_order_relaxed);
   }
 }
 
@@ -382,25 +381,25 @@ template <typename T> inline void Parmap<T>::FutexSynchro::futex_wake(unsigned*
 
 template <typename T> void Parmap<T>::FutexSynchro::master_signal()
 {
-  this->parmap.thread_counter = 1;
-  __sync_add_and_fetch(&this->parmap.work_round, 1);
+  __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
+  __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
   /* wake all workers */
   futex_wake(&this->parmap.work_round, std::numeric_limits<int>::max());
 }
 
 template <typename T> void Parmap<T>::FutexSynchro::master_wait()
 {
-  unsigned count = this->parmap.thread_counter;
+  unsigned count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
   while (count < this->parmap.num_workers) {
     /* wait for all workers to be ready */
     futex_wait(&this->parmap.thread_counter, count);
-    count = this->parmap.thread_counter;
+    count = __atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST);
   }
 }
 
 template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
 {
-  unsigned count = __sync_add_and_fetch(&this->parmap.thread_counter, 1);
+  unsigned count = __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
   if (count == this->parmap.num_workers) {
     /* all workers have finished, wake the controller */
     futex_wake(&this->parmap.thread_counter, std::numeric_limits<int>::max());
@@ -409,37 +408,37 @@ template <typename T> void Parmap<T>::FutexSynchro::worker_signal()
 
 template <typename T> void Parmap<T>::FutexSynchro::worker_wait(unsigned round)
 {
-  unsigned work_round = this->parmap.work_round;
+  unsigned work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
   /* wait for more work */
   while (work_round != round) {
     futex_wait(&this->parmap.work_round, work_round);
-    work_round = this->parmap.work_round;
+    work_round = __atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST);
   }
 }
 #endif
 
 template <typename T> void Parmap<T>::BusyWaitSynchro::master_signal()
 {
-  this->parmap.thread_counter = 1;
-  __sync_add_and_fetch(&this->parmap.work_round, 1);
+  __atomic_store_n(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
+  __atomic_add_fetch(&this->parmap.work_round, 1, __ATOMIC_SEQ_CST);
 }
 
 template <typename T> void Parmap<T>::BusyWaitSynchro::master_wait()
 {
-  while (this->parmap.thread_counter < this->parmap.num_workers) {
+  while (__atomic_load_n(&this->parmap.thread_counter, __ATOMIC_SEQ_CST) < this->parmap.num_workers) {
     xbt_os_thread_yield();
   }
 }
 
 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_signal()
 {
-  __sync_add_and_fetch(&this->parmap.thread_counter, 1);
+  __atomic_add_fetch(&this->parmap.thread_counter, 1, __ATOMIC_SEQ_CST);
 }
 
 template <typename T> void Parmap<T>::BusyWaitSynchro::worker_wait(unsigned round)
 {
   /* wait for more work */
-  while (this->parmap.work_round != round) {
+  while (__atomic_load_n(&this->parmap.work_round, __ATOMIC_SEQ_CST) != round) {
     xbt_os_thread_yield();
   }
 }