diff --git a/lib/CodeGen/CGCUDABuiltin.cpp b/lib/CodeGen/CGCUDABuiltin.cpp
index 0ccba8982a6da6d85a6ebd8070d7f6d81d152c02..ea3b888635c3197cef7f0a8637580b7e7b560f54 100644
--- a/lib/CodeGen/CGCUDABuiltin.cpp
+++ b/lib/CodeGen/CGCUDABuiltin.cpp
@@ -83,6 +83,13 @@ CodeGenFunction::EmitCUDADevicePrintfCallExpr(const CallExpr *E,
                E->arguments(), E->getDirectCallee(),
                /* ParamsToSkip = */ 0);
 
+  // We don't know how to emit non-scalar varargs.
+  if (std::any_of(Args.begin() + 1, Args.end(),
+                  [](const CallArg &A) { return !A.RV.isScalar(); })) {
+    CGM.ErrorUnsupported(E, "non-scalar arg to printf");
+    return RValue::get(llvm::ConstantInt::get(IntTy, 0));
+  }
+
   // Construct and fill the args buffer that we'll pass to vprintf.
   llvm::Value *BufferPtr;
   if (Args.size() <= 1) {
diff --git a/test/CodeGenCUDA/printf-aggregate.cu b/test/CodeGenCUDA/printf-aggregate.cu
new file mode 100644
index 0000000000000000000000000000000000000000..2e703b81d09b7ad653f9d8c094ff47ab7b8b72a4
--- /dev/null
+++ b/test/CodeGenCUDA/printf-aggregate.cu
@@ -0,0 +1,17 @@
+// REQUIRES: x86-registered-target
+// REQUIRES: nvptx-registered-target
+
+// RUN: not %clang_cc1 -triple nvptx64-nvidia-cuda -fcuda-is-device -emit-llvm \
+// RUN:   -o - %s 2>&1 | FileCheck %s
+
+#include "Inputs/cuda.h"
+
+// Check that we don't crash when asked to printf a non-scalar arg.
+struct Struct {
+  int x;
+  int y;
+};
+__device__ void PrintfNonScalar() {
+  // CHECK: cannot compile this non-scalar arg to printf
+  printf("%d", Struct());
+}