diff --git a/ChangeLog.tile b/ChangeLog.tile
index 5cdc293f8e..83136aa6c3 100644
--- a/ChangeLog.tile
+++ b/ChangeLog.tile
@@ -1,3 +1,8 @@
+2012-05-12  Chris Metcalf  <cmetcalf@tilera.com>
+
+	* sysdeps/tile/tilegx/memcpy.c: Allow memcpy(p, p, n)
+	without corrupting memory at "p".
+
 2012-05-12  Chris Metcalf  <cmetcalf@tilera.com>
 
 	* sysdeps/tile/__tls_get_addr.S: Use __WORDSIZE, not _LP64.
diff --git a/sysdeps/tile/tilegx/memcpy.c b/sysdeps/tile/tilegx/memcpy.c
index 9cfb283c82..dd6e30dd60 100644
--- a/sysdeps/tile/tilegx/memcpy.c
+++ b/sysdeps/tile/tilegx/memcpy.c
@@ -1,4 +1,4 @@
-/* Copyright (C) 2011 Free Software Foundation, Inc.
+/* Copyright (C) 2011-2012 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
    Contributed by Chris Metcalf <cmetcalf@tilera.com>, 2011.
 
@@ -107,6 +107,17 @@ __memcpy (void *__restrict dstv, const void *__restrict srcv, size_t n)
                n -= sizeof (word_t))
             *dst8++ = *src8++;
 
+          /* If copying to self, return.  The test is cheap enough
+             that we do it despite the fact that the memcpy() contract
+             doesn't require us to support overlapping dst and src.
+             This is the most common case of overlap, and any close
+             overlap will cause corruption due to the wh64 below.
+             This case is particularly important since the compiler
+             will emit memcpy() calls for aggregate copies even if it
+             can't prove that src != dst.  */
+          if (__builtin_expect (dst8 == src8, 0))
+            return dstv;
+
           for (; n >= CHIP_L2_LINE_SIZE ();)
             {
               __insn_wh64 (dst8);