diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -8944,8 +8944,9 @@
       // for now.
       if (Args[i].IsReturned && !Op.getValueType().isVector() &&
           CanLowerReturn) {
-        assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
-               "unexpected use of 'returned'");
+        assert((CLI.RetTy == Args[i].Ty ||
+                CLI.RetTy->isPointerTy() && Args[i].Ty->isPointerTy()) &&
+               RetTys.size() == NumValues && "unexpected use of 'returned'");
         // Before passing 'returned' to the target lowering code, ensure that
         // either the register MVT and the actual EVT are the same size or that
         // the return value and argument are extended in the same way; in these
diff --git a/llvm/test/CodeGen/Generic/arg_returned_bitcast.ll b/llvm/test/CodeGen/Generic/arg_returned_bitcast.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/Generic/arg_returned_bitcast.ll
@@ -0,0 +1,13 @@
+; RUN: llc < %s -asm-verbose=false
+
+; Test that the "returned" attribute "works" even if there is a bitcast between
+; the argument and return value.
+
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+
+declare double* @bar(i8* returned)
+
+define double* @foo(i8*) {
+  %r = tail call double* @bar(i8* %0)
+  ret double* %r
+}