diff --git a/[refs] b/[refs]
index 2795f8e570fa..7a42700d1cbf 100644
--- a/[refs]
+++ b/[refs]
@@ -1,2 +1,2 @@
 ---
-refs/heads/master: b287f69676a34a9fc341de4d79a9c74e1959dec6
+refs/heads/master: efb734d8ed040b053f53fd53589ed5d9c9b5cd04
diff --git a/trunk/arch/tile/lib/Makefile b/trunk/arch/tile/lib/Makefile
index 0c26086ecbef..985f59858234 100644
--- a/trunk/arch/tile/lib/Makefile
+++ b/trunk/arch/tile/lib/Makefile
@@ -7,6 +7,7 @@ lib-y = cacheflush.o checksum.o cpumask.o delay.o uaccess.o \
 	strchr_$(BITS).o strlen_$(BITS).o
 
 ifeq ($(CONFIG_TILEGX),y)
+CFLAGS_REMOVE_memcpy_user_64.o = -fno-omit-frame-pointer
 lib-y += memcpy_user_64.o
 else
 lib-y += atomic_32.o atomic_asm_32.o memcpy_tile64.o
diff --git a/trunk/arch/tile/lib/memcpy_user_64.c b/trunk/arch/tile/lib/memcpy_user_64.c
index 4763b3aff1cc..37440caa7370 100644
--- a/trunk/arch/tile/lib/memcpy_user_64.c
+++ b/trunk/arch/tile/lib/memcpy_user_64.c
@@ -14,7 +14,13 @@
  * Do memcpy(), but trap and return "n" when a load or store faults.
  *
  * Note: this idiom only works when memcpy() compiles to a leaf function.
- * If "sp" is updated during memcpy, the "jrp lr" will be incorrect.
+ * Here leaf function not only means it does not have calls, but also
+ * requires no stack operations (sp, stack frame pointer) and no
+ * use of callee-saved registers, else "jrp lr" will be incorrect since
+ * unwinding stack frame is bypassed. Since memcpy() is not complex so
+ * these conditions are satisfied here, but we need to be careful when
+ * modifying this file. This is not a clean solution but is the best
+ * one so far.
  *
  * Also note that we are capturing "n" from the containing scope here.
  */