diff --git a/tests/tcctest.c b/tests/tcctest.c
index 30d78b8..f8a9ca6 100644
--- a/tests/tcctest.c
+++ b/tests/tcctest.c
@@ -2677,6 +2677,7 @@ void asm_test(void)
        asm block gets the outer one.  */
     int base_func = 42;
     void override_func3 (void);
+    unsigned long asmret;
 
     printf("inline asm:\n");
 
@@ -2720,6 +2721,11 @@ void asm_test(void)
        the global one, not the local decl from this function.  */
     asm volatile(".weak override_func3\n.set override_func3, base_func");
     override_func3();
+    /* Check that we can also load structs of appropriate layout
+       into registers.  */
+    asm volatile("" : "=r" (asmret) : "0"(s2));
+    if (asmret != s2.addr)
+      printf("asmstr: failed\n");
     return;
  label1:
     goto label2;
diff --git a/x86_64-gen.c b/x86_64-gen.c
index 8a2d03a..ea06708 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -382,6 +382,22 @@ void load(int r, SValue *sv)
             load(fr, &v1);
         }
         ll = 0;
+	/* Like GCC we can load from small enough properly sized
+	   structs and unions as well.
+	   XXX maybe move to generic operand handling, but should
+	   occur only with asm, so tccasm.c might also be a better place */
+	if ((ft & VT_BTYPE) == VT_STRUCT) {
+	    int align;
+	    switch (type_size(&sv->type, &align)) {
+		case 1: ft = VT_BYTE; break;
+		case 2: ft = VT_SHORT; break;
+		case 4: ft = VT_INT; break;
+		case 8: ft = VT_LLONG; break;
+		default:
+		    tcc_error("invalid aggregate type for register load");
+		    break;
+	    }
+	}
         if ((ft & VT_BTYPE) == VT_FLOAT) {
             b = 0x6e0f66;
             r = REG_VALUE(r); /* movd */