diff --git a/lib/Transforms/ObjCARC/PtrState.cpp b/lib/Transforms/ObjCARC/PtrState.cpp
index a5afc8ad977..c1bbc4e96b1 100644
--- a/lib/Transforms/ObjCARC/PtrState.cpp
+++ b/lib/Transforms/ObjCARC/PtrState.cpp
@@ -351,8 +351,10 @@ bool TopDownPtrState::HandlePotentialAlterRefCount(Instruction *Inst,
                                                    const Value *Ptr,
                                                    ProvenanceAnalysis &PA,
                                                    ARCInstKind Class) {
-  // Check for possible releases.
-  if (!CanAlterRefCount(Inst, Ptr, PA, Class))
+  // Check for possible releases. Treat clang.arc.use as a releasing instruction
+  // to prevent sinking a retain past it.
+  if (!CanAlterRefCount(Inst, Ptr, PA, Class) &&
+      Class != ARCInstKind::IntrinsicUser)
     return false;
 
   DEBUG(dbgs() << "            CanAlterRefCount: Seq: " << GetSeq() << "; " << *Ptr
diff --git a/test/Transforms/ObjCARC/clang-arc-use-barrier.ll b/test/Transforms/ObjCARC/clang-arc-use-barrier.ll
new file mode 100644
index 00000000000..98d49ec512e
--- /dev/null
+++ b/test/Transforms/ObjCARC/clang-arc-use-barrier.ll
@@ -0,0 +1,45 @@
+; RUN: opt -objc-arc -S %s | FileCheck %s
+
+%0 = type opaque
+
+; Make sure ARC optimizer doesn't sink @obj_retain past @clang.arc.use.
+
+; CHECK: call i8* @objc_retain(
+; CHECK: call void (...) @clang.arc.use(
+; CHECK: call i8* @objc_retain(
+; CHECK: call void (...) @clang.arc.use(
+
+define void @runTest() local_unnamed_addr {
+  %1 = alloca %0*, align 8
+  %2 = alloca %0*, align 8
+  %3 = tail call %0* @foo0()
+  %4 = bitcast %0* %3 to i8*
+  %5 = tail call i8* @objc_retainAutoreleasedReturnValue(i8* %4)
+  store %0* %3, %0** %1, align 8
+  call void @foo1(%0** nonnull %1)
+  %6 = load %0*, %0** %1, align 8
+  %7 = bitcast %0* %6 to i8*
+  %8 = call i8* @objc_retain(i8* %7)
+  call void (...) @clang.arc.use(%0* %3)
+  call void @objc_release(i8* %4)
+  store %0* %6, %0** %2, align 8
+  call void @foo1(%0** nonnull %2)
+  %9 = load %0*, %0** %2, align 8
+  %10 = bitcast %0* %9 to i8*
+  %11 = call i8* @objc_retain(i8* %10)
+  call void (...) @clang.arc.use(%0* %6)
+  %tmp1 = load %0*, %0** %2, align 8
+  call void @objc_release(i8* %7)
+  call void @foo2(%0* %9)
+  call void @objc_release(i8* %10)
+  ret void
+}
+
+declare %0* @foo0() local_unnamed_addr
+declare void @foo1(%0**) local_unnamed_addr
+declare void @foo2(%0*) local_unnamed_addr
+
+declare i8* @objc_retainAutoreleasedReturnValue(i8*) local_unnamed_addr
+declare i8* @objc_retain(i8*) local_unnamed_addr
+declare void @clang.arc.use(...) local_unnamed_addr
+declare void @objc_release(i8*) local_unnamed_addr
diff --git a/test/Transforms/ObjCARC/intrinsic-use.ll b/test/Transforms/ObjCARC/intrinsic-use.ll
index f75b1872b17..f5956201454 100644
--- a/test/Transforms/ObjCARC/intrinsic-use.ll
+++ b/test/Transforms/ObjCARC/intrinsic-use.ll
@@ -14,23 +14,20 @@ declare void @test0_helper(i8*, i8**)
 ; Ensure that we honor clang.arc.use as a use and don't miscompile
 ; the reduced test case from <rdar://13195034>.
 ;
-; FIXME: the fact that we re-order retains w.r.t. @clang.arc.use could
-; be problematic if we get run twice, e.g. under LTO.
-;
 ; CHECK-LABEL:      define void @test0(
 ; CHECK:        @objc_retain(i8* %x)
 ; CHECK-NEXT:   store i8* %y, i8** %temp0
 ; CHECK-NEXT:   @objc_retain(i8* %y)
 ; CHECK-NEXT:   call void @test0_helper
 ; CHECK-NEXT:   [[VAL1:%.*]] = load i8*, i8** %temp0
-; CHECK-NEXT:   call void (...) @clang.arc.use(i8* %y)
 ; CHECK-NEXT:   @objc_retain(i8* [[VAL1]])
+; CHECK-NEXT:   call void (...) @clang.arc.use(i8* %y)
 ; CHECK-NEXT:   @objc_release(i8* %y)
 ; CHECK-NEXT:   store i8* [[VAL1]], i8** %temp1
 ; CHECK-NEXT:   call void @test0_helper
 ; CHECK-NEXT:   [[VAL2:%.*]] = load i8*, i8** %temp1
-; CHECK-NEXT:   call void (...) @clang.arc.use(i8* [[VAL1]])
 ; CHECK-NEXT:   @objc_retain(i8* [[VAL2]])
+; CHECK-NEXT:   call void (...) @clang.arc.use(i8* [[VAL1]])
 ; CHECK-NEXT:   @objc_release(i8* [[VAL1]])
 ; CHECK-NEXT:   @objc_autorelease(i8* %x)
 ; CHECK-NEXT:   store i8* %x, i8** %out
@@ -71,14 +68,14 @@ entry:
 ; CHECK-NEXT:   @objc_retain(i8* %y)
 ; CHECK-NEXT:   call void @test0_helper
 ; CHECK-NEXT:   [[VAL1:%.*]] = load i8*, i8** %temp0
-; CHECK-NEXT:   call void (...) @clang.arc.use(i8* %y)
 ; CHECK-NEXT:   @objc_retain(i8* [[VAL1]])
+; CHECK-NEXT:   call void (...) @clang.arc.use(i8* %y)
 ; CHECK-NEXT:   @objc_release(i8* %y)
 ; CHECK-NEXT:   store i8* [[VAL1]], i8** %temp1
 ; CHECK-NEXT:   call void @test0_helper
 ; CHECK-NEXT:   [[VAL2:%.*]] = load i8*, i8** %temp1
-; CHECK-NEXT:   call void (...) @clang.arc.use(i8* [[VAL1]])
 ; CHECK-NEXT:   @objc_retain(i8* [[VAL2]])
+; CHECK-NEXT:   call void (...) @clang.arc.use(i8* [[VAL1]])
 ; CHECK-NEXT:   @objc_release(i8* [[VAL1]])
 ; CHECK-NEXT:   @objc_autorelease(i8* %x)
 ; CHECK-NEXT:   @objc_release(i8* [[VAL2]])