diff --git a/clang/lib/Frontend/CompilerInstance.cpp b/clang/lib/Frontend/CompilerInstance.cpp
index b96dc70..6fb880f 100644
--- a/clang/lib/Frontend/CompilerInstance.cpp
+++ b/clang/lib/Frontend/CompilerInstance.cpp
@@ -334,9 +334,16 @@
   InitializePreprocessor(*PP, PPOpts, getPCHContainerReader(),
                          getFrontendOpts());
 
-  // Initialize the header search object.
+  // Initialize the header search object.  In CUDA compilations, we use the aux
+  // triple (the host triple) to initialize our header search, since we need to
+  // find the host headers in order to compile the CUDA code.
+  const llvm::Triple *HeaderSearchTriple = &PP->getTargetInfo().getTriple();
+  if (PP->getTargetInfo().getTriple().getOS() == llvm::Triple::CUDA &&
+      PP->getAuxTargetInfo())
+    HeaderSearchTriple = &PP->getAuxTargetInfo()->getTriple();
+
   ApplyHeaderSearchOptions(PP->getHeaderSearchInfo(), getHeaderSearchOpts(),
-                           PP->getLangOpts(), PP->getTargetInfo().getTriple());
+                           PP->getLangOpts(), *HeaderSearchTriple);
 
   PP->setPreprocessedOutput(getPreprocessorOutputOpts().ShowCPP);