diff --git a/lib/LLVMPasses/LLVMMergeFunctions.cpp b/lib/LLVMPasses/LLVMMergeFunctions.cpp
index 25b4d3f61b220..07734c20e2166 100644
--- a/lib/LLVMPasses/LLVMMergeFunctions.cpp
+++ b/lib/LLVMPasses/LLVMMergeFunctions.cpp
@@ -1188,6 +1188,9 @@ void SwiftMergeFunctions::removeEquivalenceClassFromTree(FunctionEntry *FE) {
 // Selects proper bitcast operation,
 // but a bit simpler then CastInst::getCastOpcode.
 static Value *createCast(IRBuilder<> &Builder, Value *V, Type *DestTy) {
+  if (V->getType() == DestTy)
+    return V;
+
   Type *SrcTy = V->getType();
   if (SrcTy->isStructTy()) {
     assert(DestTy->isStructTy());
@@ -1357,7 +1360,8 @@ bool SwiftMergeFunctions::replaceDirectCallers(Function *Old, Function *New,
                                             NewArgAttrs);
     newAttrList = fixUpTypesInByValAndStructRetAttributes(FType, newAttrList);
     NewCI->setAttributes(newAttrList);
-    CI->replaceAllUsesWith(NewCI);
+    Value *retVal = createCast(Builder, NewCI, CI->getType());
+    CI->replaceAllUsesWith(retVal);
     CI->eraseFromParent();
   }
   assert(Old->use_empty() && "should have replaced all uses of old function");
diff --git a/test/LLVMPasses/merge_func_opaque_pointers.ll b/test/LLVMPasses/merge_func_opaque_pointers.ll
new file mode 100644
index 0000000000000..7165fc5b32fad
--- /dev/null
+++ b/test/LLVMPasses/merge_func_opaque_pointers.ll
@@ -0,0 +1,34 @@
+; RUN: %swift-llvm-opt -swift-merge-functions -swiftmergefunc-threshold=4 -opaque-pointers %s | %FileCheck %s
+
+; REQUIRES: PTRSIZE=64
+
+@g1 = external global i32
+
+define internal i64 @return_0(i32 %x, i32 %y) {
+  %sum = add i32 %x, %y
+  %sum2 = add i32 %sum, %y
+  %l = load i32, i32* @g1, align 4
+  %sum3 = add i32 %sum2, %y
+  ret i64 0
+}
+
+define internal ptr @return_null(i32 %x, i32 %y) {
+  %sum = add i32 %x, %y
+  %sum2 = add i32 %sum, %y
+  %l = load i32, i32* @g1, align 4
+  %sum3 = add i32 %sum2, %y
+  ret ptr null
+}
+
+declare void @take_results(i64, ptr)
+
+
+; CHECK-LABEL: define void @call_return_null
+; CHECK:         %1 = call i64 @return_0Tm(i32 0, i32 0)
+; CHECK:         %2 = call ptr @return_0Tm(i32 0, i32 0)
+define void @call_return_null()  {
+  %r1 = call i64 @return_0(i32 0, i32 0)
+  %r2 = call ptr @return_null(i32 0, i32 0)
+  call void @take_results(i64 %r1, ptr %r2)
+  ret void
+}