fix: Protect local memory stores with shmem_barrier_all()

Author: dalcinl

src/libshmem/fallback.h

@@ -311,7 +311,7 @@ static inline long *_py_shmem_pSync()
     _py_shmem_pSync_array = (long *) shmem_malloc(SHMEM_SYNC_SIZE * sizeof(long));
     for (int i = 0; i < SHMEM_SYNC_SIZE; i++)
       _py_shmem_pSync_array[i] = SHMEM_SYNC_VALUE;
-    shmem_sync_all();
+    shmem_barrier_all();
   }
   return _py_shmem_pSync_array;
 }
@@ -328,7 +328,7 @@ static inline void *_py_shmem_pWrk(size_t nreduce, size_t eltsize)
     shmem_free(_py_shmem_pWrk_array);
     _py_shmem_pWrk_size  = wrk_size;
     _py_shmem_pWrk_array = shmem_malloc(wrk_size);
-    shmem_sync_all();
+    shmem_barrier_all();
   }
   return _py_shmem_pWrk_array;
 }

src/libshmem/memalloc.h

@@ -13,7 +13,7 @@ static void *shmem_py_alloc(size_t size, size_t align, long hints, int clear)
   }
   if (clear) {
     memset(ptr, 0, size);
-    shmem_sync_all();
+    shmem_barrier_all();
   }
   return ptr;
 }

src/shmem4py/shmem.py

@@ -889,7 +889,7 @@ def array(
     if tmp.ndim > 1:
         a.strides = tmp.strides
     np.copyto(a, tmp, casting='no')
-    lib.shmem_sync_all()
+    lib.shmem_barrier_all()
     return a
 
 
@@ -969,7 +969,7 @@ def ones(
     """
     a = new_array(shape, dtype, order, align=align, hints=hints, clear=False)
     np.copyto(a, 1, casting='unsafe')
-    lib.shmem_sync_all()
+    lib.shmem_barrier_all()
     return a
 
 
@@ -1001,7 +1001,7 @@ def full(
         dtype = np.array(fill_value).dtype
     a = new_array(shape, dtype, order, align=align, hints=hints, clear=False)
     np.copyto(a, fill_value, casting='unsafe')
-    lib.shmem_sync_all()
+    lib.shmem_barrier_all()
     return a