Always SFENCE at end of memcpy_nontemporal

It might be needed for correct ordering with write-combining memory,
such as used by gdrcopy.

doc/changelog.rst

@@ -8,6 +8,8 @@ Changelog
 - Make the ibverbs sender compatible with `PeerDirect`_.
 - Add examples programs showing integration with `gdrcopy`_ and
   `PeerDirect`_.
+- Always use SFENCE at end of :cpp:func:`memcpy_nontemporal` so that it is
+  appropriate for use with `gdrcopy`_.
 - Fix a memory leak when receiving with ibverbs.
 
 .. _gdrcopy: https://github.com/NVIDIA/gdrcopy

src/common_memcpy.cpp

@@ -43,7 +43,14 @@ void *memcpy_nontemporal(void * __restrict__ dest, const void * __restrict__ src
     {
         if (head >= n)
         {
-            return std::memcpy(dest_c, src_c, n);
+            std::memcpy(dest_c, src_c, n);
+            /* Not normally required, but if the destination is
+             * write-combining memory then this will flush the combining
+             * buffers. That may be necessary if the memory is actually on
+             * a GPU or other accelerator.
+             */
+            _mm_sfence();
+            return dest;
         }
         std::memcpy(dest_c, src_c, head);
         dest_c += head;