This backport contains 1 patch from gcc main stream tree.
The commit id of these patchs list as following in the order of time.

0001-tree-optimization-94949-fix-load-eliding-in-SM.patch
0424a5ece5307cc22bbc0fe97edf4707d7a798ed

diff -Nurp a/gcc/testsuite/gcc.dg/torture/pr94949.c b/gcc/testsuite/gcc.dg/torture/pr94949.c
--- a/gcc/testsuite/gcc.dg/torture/pr94949.c	1970-01-01 08:00:00.000000000 +0800
+++ b/gcc/testsuite/gcc.dg/torture/pr94949.c	2020-08-24 21:40:32.208000000 +0800
@@ -0,0 +1,17 @@
+/* { dg-do run } */
+/* { dg-additional-options "-fallow-store-data-races" } */
+
+static int x = 1;
+static volatile int y = -1;
+int
+main()
+{
+  for (int i = 0; i < 128; ++i)
+    {
+      if (i == y)
+	x = i;
+    }
+  if (x != 1)
+    __builtin_abort ();
+  return 0;
+}
diff -Nurp a/gcc/tree-ssa-loop-im.c b/gcc/tree-ssa-loop-im.c
--- a/gcc/tree-ssa-loop-im.c	2020-08-24 21:40:14.164000000 +0800
+++ b/gcc/tree-ssa-loop-im.c	2020-08-24 21:40:32.208000000 +0800
@@ -2115,9 +2115,9 @@ execute_sm (struct loop *loop, vec<edge>
   fmt_data.orig_loop = loop;
   for_each_index (&ref->mem.ref, force_move_till, &fmt_data);
 
+  bool always_stored = ref_always_accessed_p (loop, ref, true);
   if (bb_in_transaction (loop_preheader_edge (loop)->src)
-      || (! flag_store_data_races
-	  && ! ref_always_accessed_p (loop, ref, true)))
+      || (! flag_store_data_races && ! always_stored))
     multi_threaded_model_p = true;
 
   if (multi_threaded_model_p)
@@ -2132,8 +2132,10 @@ execute_sm (struct loop *loop, vec<edge>
 
   /* Avoid doing a load if there was no load of the ref in the loop.
      Esp. when the ref is not always stored we cannot optimize it
-     away later.  */
-  if (ref->loaded && bitmap_bit_p (ref->loaded, loop->num))
+     away later.  But when it is not always stored we must use a conditional
+     store then.  */
+  if ((!always_stored && !multi_threaded_model_p)
+      || (ref->loaded && bitmap_bit_p (ref->loaded, loop->num)))
     {
       load = gimple_build_assign (tmp_var, unshare_expr (ref->mem.ref));
       lim_data = init_lim_data (load);