Merge pull request #405 from jdinan/pr/release-updates

Updates for v1.3.3rc1 release

NEWS

@@ -2,6 +2,18 @@
 
 Sandia OpenSHMEM NEWS -- history of user-visible changes.
 
+v1.3.3rc1
+---------
+- Fixed dissemination barrier race that could lead to incorrect synchronization.
+- Improved collect and all-to-all collectives performance.
+- Improved error reporting and debugging output.
+- Added support for proposed bitwise atomics (see shmemx.h).
+- Added shmemx_pcontrol function to profiling interface extension (see shmemx.h).
+- Added symbol hiding support in SHMEM library.
+- Fixed symmetric heap corruption bug in GUPS benchmark (see test/apps/gups.c).
+- Added target-side measurement and asymmetric configuration support to
+  performance suite benchmarks.
+
 v1.3.2
 ------
 - Improved support for proposed multithreading extension (see shmemx.h),

configure.ac

@@ -13,7 +13,7 @@
 
 dnl Init Autoconf/Automake/Libtool
 
-AC_INIT([Sandia OpenSHMEM], [1.3.2], [https://github.com/Sandia-OpenSHMEM/SOS])
+AC_INIT([Sandia OpenSHMEM], [1.3.3rc1], [https://github.com/Sandia-OpenSHMEM/SOS])
 AC_PREREQ([2.60])
 AC_CONFIG_AUX_DIR([config])
 AC_CONFIG_MACRO_DIR([config])

src/symmetric_heap_c.c

@@ -302,9 +302,13 @@ shmem_free(void *ptr)
 
     shmem_internal_barrier_all();
 
-    SHMEM_MUTEX_LOCK(shmem_internal_mutex_alloc);
-    dlfree(ptr);
-    SHMEM_MUTEX_UNLOCK(shmem_internal_mutex_alloc);
+    /* It's fine to call dlfree with NULL, but better to avoid unnecessarily
+     * taking the mutex in the threaded case. */
+    if (ptr != NULL) {
+        SHMEM_MUTEX_LOCK(shmem_internal_mutex_alloc);
+        dlfree(ptr);
+        SHMEM_MUTEX_UNLOCK(shmem_internal_mutex_alloc);
+    }
 }
 
 
@@ -321,7 +325,12 @@ shmem_realloc(void *ptr, size_t size)
     shmem_internal_barrier_all();
 
     SHMEM_MUTEX_LOCK(shmem_internal_mutex_alloc);
-    ret = dlrealloc(ptr, size);
+    if (size == 0 && ptr != NULL) {
+        dlfree(ptr);
+        ret = NULL;
+    } else {
+        ret = dlrealloc(ptr, size);
+    }
     SHMEM_MUTEX_UNLOCK(shmem_internal_mutex_alloc);
 
     shmem_internal_barrier_all();
@@ -337,6 +346,24 @@ shmem_align(size_t alignment, size_t size)
 
     SHMEM_ERR_CHECK_INITIALIZED();
 
+    /* Alignment must be at least sizeof(void*) */
+    if (alignment < sizeof(void*))
+        return NULL;
+
+    /* Round alignment up to the nearest power of two if
+     * not already a power of two */
+    if ((alignment & (alignment-1)) != 0) {
+        size_t c, log2_alignment = 0;
+
+        for (c = alignment >> 1; c; c >>= 1)
+            log2_alignment++;
+
+        DEBUG_MSG("Alignment was rounded up from %zu to %zu\n",
+                  alignment, (size_t) 1 << (log2_alignment + 1));
+
+        alignment = 1 << (log2_alignment + 1);
+    }
+
     SHMEM_MUTEX_LOCK(shmem_internal_mutex_alloc);
     ret = dlmemalign(alignment, size);
     SHMEM_MUTEX_UNLOCK(shmem_internal_mutex_alloc);

test/unit/shmemalign.c

@@ -154,7 +154,7 @@ main(int argc, char **argv)
     for(l=0; l < loops; l++)
     {
         /* align 2**2 ... 2**23; 24 exceeds symetric heap max */
-        for(j=0,c=2; j < 23; j++,c<<=1)
+        for(j=0,c=8; j < 21; j++,c<<=1)
         {
             target_sz = nWords * sizeof(DataType);
             if (!(target = (DataType *)shmem_align(c,target_sz))) {