diff --git a/include/config/SkUserConfig.h b/include/config/SkUserConfig.h
index 81033f3..5f59b91 100644
--- a/include/config/SkUserConfig.h
+++ b/include/config/SkUserConfig.h
@@ -176,5 +176,6 @@
  */
 //#define SK_ATOMICS_PLATFORM_H "SkAtomics_xxx.h"
 //#define SK_MUTEX_PLATFORM_H "SkMutex_xxx.h"
+//#define SK_BARRIERS_PLATFORM_H "SkBarriers_xxx.h"
 
 #endif
diff --git a/include/core/SkOnce.h b/include/core/SkOnce.h
index d5330b9..d39a05b 100644
--- a/include/core/SkOnce.h
+++ b/include/core/SkOnce.h
@@ -71,44 +71,6 @@
     SkSpinlock lock;
 };
 
-// TODO(bungeman, mtklein): move all these *barrier* functions to SkThread when refactoring lands.
-
-#ifdef SK_BUILD_FOR_WIN
-#  include <intrin.h>
-inline static void compiler_barrier() {
-    _ReadWriteBarrier();
-}
-#else
-inline static void compiler_barrier() {
-    asm volatile("" : : : "memory");
-}
-#endif
-
-inline static void full_barrier_on_arm() {
-#if (defined(SK_CPU_ARM) && SK_ARM_ARCH >= 7) || defined(SK_CPU_ARM64)
-    asm volatile("dmb ish" : : : "memory");
-#elif defined(SK_CPU_ARM)
-    asm volatile("mcr p15, 0, %0, c7, c10, 5" : : "r" (0) : "memory");
-#endif
-}
-
-// On every platform, we issue a compiler barrier to prevent it from reordering
-// code.  That's enough for platforms like x86 where release and acquire
-// barriers are no-ops.  On other platforms we may need to be more careful;
-// ARM, in particular, needs real code for both acquire and release.  We use a
-// full barrier, which acts as both, because that the finest precision ARM
-// provides.
-
-inline static void release_barrier() {
-    compiler_barrier();
-    full_barrier_on_arm();
-}
-
-inline static void acquire_barrier() {
-    compiler_barrier();
-    full_barrier_on_arm();
-}
-
 // Works with SkSpinlock or SkMutex.
 template <typename Lock>
 class SkAutoLockAcquire {
@@ -143,8 +105,7 @@
         //
         // We'll use this in the fast path to make sure f(arg)'s effects are
         // observable whenever we observe *done == true.
-        release_barrier();
-        *done = true;
+        sk_release_store(done, true);
     }
 }
 
@@ -156,15 +117,15 @@
     }
     // Also known as a load-load/load-store barrier, this acquire barrier makes
     // sure that anything we read from memory---in particular, memory written by
-    // calling f(arg)---is at least as current as the value we read from once->done.
+    // calling f(arg)---is at least as current as the value we read from done.
     //
     // In version control terms, this is a lot like saying "sync up to the
-    // commit where we wrote once->done = true".
+    // commit where we wrote done = true".
     //
-    // The release barrier in sk_once_slow guaranteed that once->done = true
-    // happens after f(arg), so by syncing to once->done = true here we're
+    // The release barrier in sk_once_slow guaranteed that done = true
+    // happens after f(arg), so by syncing to done = true here we're
     // forcing ourselves to also wait until the effects of f(arg) are readble.
-    acquire_barrier();
+    SkAssertResult(sk_acquire_load(done));
 }
 
 template <typename Func, typename Arg>
diff --git a/include/core/SkPostConfig.h b/include/core/SkPostConfig.h
index 2538522..d4d9857 100644
--- a/include/core/SkPostConfig.h
+++ b/include/core/SkPostConfig.h
@@ -399,6 +399,14 @@
 #  endif
 #endif
 
+#ifndef SK_BARRIERS_PLATFORM_H
+#  if defined(SK_CPU_ARM) || defined(SK_CPU_ARM64)
+#    define SK_BARRIERS_PLATFORM_H "../../src/ports/SkBarriers_arm.h"
+#  else
+#    define SK_BARRIERS_PLATFORM_H "../../src/ports/SkBarriers_x86.h"
+#  endif
+#endif
+
 
 //////////////////////////////////////////////////////////////////////
 
diff --git a/include/core/SkThread.h b/include/core/SkThread.h
index c8cd4e9..3038e2d 100644
--- a/include/core/SkThread.h
+++ b/include/core/SkThread.h
@@ -60,6 +60,25 @@
     return prev;
 }
 
+// SK_BARRIERS_PLATFORM_H must provide implementations for the following declarations:
+
+/** Prevent the compiler from reordering across this barrier. */
+static void sk_compiler_barrier();
+
+/** Read T*, with at least an acquire barrier.
+ *
+ *  Only needs to be implemented for T which can be atomically read.
+ */
+template <typename T> T sk_acquire_load(T*);
+
+/** Write T*, with at least a release barrier.
+ *
+ *  Only needs to be implemented for T which can be atomically written.
+ */
+template <typename T> void sk_release_store(T*, T);
+
+#include SK_BARRIERS_PLATFORM_H
+
 /** SK_MUTEX_PLATFORM_H must provide the following (or equivalent) declarations.
 
 class SkBaseMutex {
diff --git a/src/ports/SkBarriers_arm.h b/src/ports/SkBarriers_arm.h
new file mode 100644
index 0000000..9161cdd
--- /dev/null
+++ b/src/ports/SkBarriers_arm.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright 2014 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkBarriers_arm_DEFINED
+#define SkBarriers_arm_DEFINED
+
+static inline void sk_compiler_barrier() { asm volatile("" : : : "memory"); }
+
+template <typename T>
+T sk_acquire_load(T* ptr) {
+    T val = *ptr;
+    __sync_synchronize();  // Issue a full barrier, which is an overkill acquire barrier.
+    return val;
+}
+
+template <typename T>
+void sk_release_store(T* ptr, T val) {
+    __sync_synchronize();  // Issue a full barrier, which is an overkill release barrier.
+    *ptr = val;
+}
+
+#endif//SkBarriers_x86_DEFINED
diff --git a/src/ports/SkBarriers_x86.h b/src/ports/SkBarriers_x86.h
new file mode 100644
index 0000000..fc57615
--- /dev/null
+++ b/src/ports/SkBarriers_x86.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright 2014 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkBarriers_x86_DEFINED
+#define SkBarriers_x86_DEFINED
+
+#ifdef SK_BUILD_FOR_WIN
+#  include <intrin.h>
+static inline void sk_compiler_barrier() { _ReadWriteBarrier(); }
+#else
+static inline void sk_compiler_barrier() { asm volatile("" : : : "memory"); }
+#endif
+
+template <typename T>
+T sk_acquire_load(T* ptr) {
+    T val = *ptr;
+    // On x86, all loads are acquire loads, so we only need a compiler barrier.
+    sk_compiler_barrier();
+    return val;
+}
+
+template <typename T>
+void sk_release_store(T* ptr, T val) {
+    // On x86, all stores are release stores, so we only need a compiler barrier.
+    sk_compiler_barrier();
+    *ptr = val;
+}
+
+#endif//SkBarriers_x86_DEFINED