Force inlining memcpy for short data

This work is based on @derbeyn patch provided on #6678. I reworked it to
be more inclusive (works now with both gcc and icc) and to cover more
standard size lengths (4, 8, 16).

Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
Signed-off-by: Nadia Derbey <Nadia.Derbey@atos.net>

opal/datatype/opal_datatype_memcpy.h

@@ -1,6 +1,6 @@
 /* -*- Mode: C; c-basic-offset:4 ; -*- */
 /*
- * Copyright (c) 2004-2009 The University of Tennessee and The University
+ * Copyright (c) 2004-2019 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2009      Oak Ridge National Labs.  All rights reserved.
@@ -11,10 +11,45 @@
  * $HEADER$
  */
 
+#include "opal/runtime/opal.h"
+
 #ifndef OPAL_DATATYPE_MEMCPY_H_HAS_BEEN_INCLUDED
 #define OPAL_DATATYPE_MEMCPY_H_HAS_BEEN_INCLUDED
 
-#define MEMCPY( DST, SRC, BLENGTH ) \
-    memcpy( (DST), (SRC), (BLENGTH) )
+#define TYPED_MEMCPY( DST, SRC, BLENGTH, ALIGN, TYPE)                   \
+    if( ((BLENGTH) >= sizeof(TYPE)) && (0 == ((ALIGN) & (sizeof(TYPE) - 1))) ) { \
+        TYPE *__dst = (TYPE*)(DST), *__src = (TYPE*)(SRC);              \
+        if( (BLENGTH) == sizeof(TYPE) ) {                               \
+            *__dst = *__src;                                            \
+            break;                                                      \
+        }                                                               \
+        size_t _cnt = ((BLENGTH) / sizeof(TYPE));                       \
+        for( ; _cnt > 0; _cnt--, __dst++, __src++ ) {                   \
+            *__dst = *__src;                                            \
+            (BLENGTH) -= sizeof(TYPE);                                  \
+        }                                                               \
+        if( 0 == (BLENGTH) ) break;                                     \
+        (DST) = __dst;                                                  \
+        (SRC) = __src;                                                  \
+    }
+
+/*
+ * This macro is called whenever we are packing/unpacking a DDT that
+ * that is built with basic datatypes.
+ * Specifying a fixed size for the memcpy() makes the intel compiler
+ * inline it as an assignment operation.
+ */
+#define MEMCPY( DST, SRC, BLENGTH )                                     \
+    do {                                                                \
+        void *_dst = (void*)(DST), *_src = (void*)(SRC);                \
+        size_t _blength = (size_t)(BLENGTH);                            \
+        if( _blength < (size_t)opal_cache_line_size ) {                 \
+            uintptr_t align = ((uintptr_t)_dst) | ((uintptr_t)_src);    \
+            TYPED_MEMCPY( _dst, _src, _blength, align, int64_t );       \
+            TYPED_MEMCPY( _dst, _src, _blength, align, int32_t );       \
+            TYPED_MEMCPY( _dst, _src, _blength, align, int8_t );        \
+        }                                                               \
+        memcpy( _dst, _src, _blength );                                 \
+    } while (0)
 
 #endif  /* OPAL_DATATYPE_MEMCPY_H_HAS_BEEN_INCLUDED */