Index: lib/Transforms/Utils/LowerMemIntrinsics.cpp
===================================================================
--- lib/Transforms/Utils/LowerMemIntrinsics.cpp
+++ lib/Transforms/Utils/LowerMemIntrinsics.cpp
@@ -174,7 +174,7 @@
   Value *RuntimeBytesCopied = PLBuilder.CreateSub(CopyLen, RuntimeResidual);
 
   BasicBlock *LoopBB =
-      BasicBlock::Create(Ctx, "loop-memcpy-expansion", ParentFunc, nullptr);
+      BasicBlock::Create(Ctx, "loop-memcpy-expansion", ParentFunc, PostLoopBB);
   IRBuilder<> LoopBuilder(LoopBB);
 
   PHINode *LoopIndex = LoopBuilder.CreatePHI(CopyLenType, 2, "loop-index");
@@ -193,7 +193,8 @@
   if (LoopOpType != Int8Type) {
     // Loop body for the residual copy.
     BasicBlock *ResLoopBB = BasicBlock::Create(Ctx, "loop-memcpy-residual",
-                                               PreLoopBB->getParent(), nullptr);
+                                               PreLoopBB->getParent(),
+                                               PostLoopBB);
     // Residual loop header.
     BasicBlock *ResHeaderBB = BasicBlock::Create(
         Ctx, "loop-memcpy-residual-header", PreLoopBB->getParent(), nullptr);
Index: test/CodeGen/NVPTX/lower-aggr-copies.ll
===================================================================
--- test/CodeGen/NVPTX/lower-aggr-copies.ll
+++ test/CodeGen/NVPTX/lower-aggr-copies.ll
@@ -50,6 +50,9 @@
 ; WIR:         [[IndexInc]] = add i64 %loop-index, 1
 ; WIR:         [[Cond2:%[0-9]+]] = icmp ult i64 [[IndexInc]], [[LoopCount]]
 ; WIR:         br i1 [[Cond2]], label %loop-memcpy-expansion, label %post-loop-memcpy-expansion
+
+; WIR-LABEL:   post-loop-memcpy-expansion:
+; WIR:         ret i8* %dst
 }
 
 define i8* @memcpy_volatile_caller(i8* %dst, i8* %src, i64 %n) #0 {
@@ -85,6 +88,9 @@
 ; WIR:         [[IndexInc]] = add i64 %loop-index, 1
 ; WIR:         [[Cond2:%[0-9]+]] = icmp ult i64 [[IndexInc]], [[LoopCount]]
 ; WIR:         br i1 [[Cond2]], label %loop-memcpy-expansion, label %post-loop-memcpy-expansion
+
+; WIR-LABEL:   post-loop-memcpy-expansion:
+; WIR:         ret i8* %dst
 }
 
 define i8* @memcpy_casting_caller(i32* %dst, i32* %src, i64 %n) #0 {