diff --git a/SYCL/Regression/unoptimized_stream.cpp b/SYCL/Regression/unoptimized_stream.cpp
new file mode 100644
index 0000000000..e9d800bdc2
--- /dev/null
+++ b/SYCL/Regression/unoptimized_stream.cpp
@@ -0,0 +1,20 @@
+// RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -O0 -o %t.out
+// RUN: %HOST_RUN_PLACEHOLDER %t.out
+// RUN: %CPU_RUN_PLACEHOLDER %t.out
+// RUN: %GPU_RUN_PLACEHOLDER %t.out
+
+// NOTE: The libclc target used by the CUDA backend used to generate atomic load
+//       variants that were unsupported by NVPTX. Even if they were not used
+//       directly, sycl::stream and other operations would keep the invalid
+//       operations in when optimizations were disabled.
+
+#include <sycl/sycl.hpp>
+
+int main() {
+  sycl::queue q;
+  q.submit([&](sycl::handler &cgh) {
+    sycl::stream os(1024, 256, cgh);
+    cgh.single_task([=]() { os << "test"; });
+  });
+  return 0;
+}