[CUDA] Fix CUDA compilation broken by D50845

[tra]

Summary

This keeps predefined macros for CUDA to work as they were before and lets OpenMP control the set of macros it needs.

[Hahnfeld]

Not needed anymore after the reverts in rC341115 and rC341118, right?

[tra]

Not needed anymore after the reverts in rC341115 and rC341118, right?

Correct.

clang/lib/Frontend/InitPreprocessor.cpp

void clang::InitializePreprocessor(
   // Emit line markers for various builtin sections of the file.  We don't do
   // this in asm preprocessor mode, because "# 4" is not a line marker directive
   // in this mode.
   if (!PP.getLangOpts().AsmPreprocessor)
     Builder.append("# 1 \"<built-in>\" 3");
 
   // Install things like __POWERPC__, __GNUC__, etc into the macro table.
   if (InitOpts.UsePredefines) {
+    // CUDA and OpenMP handle preprocessing differently.  CUDA wants to see
+    // identical preprocessed output (as close to it as possible), so it
+    // provides full set of predefined macros for both sides of compilation.
+    if (LangOpts.CUDA && PP.getAuxTargetInfo())
+      InitializePredefinedMacros(*PP.getAuxTargetInfo(), LangOpts, FEOpts,
+                                 Builder);
     InitializePredefinedMacros(PP.getTargetInfo(), LangOpts, FEOpts, Builder);
-    if ((LangOpts.CUDA || LangOpts.OpenMPIsDevice) && PP.getAuxTargetInfo())
+    // OpenMP relies on selectively picking predefined macros to work around the
+    // bits of host includes it can't compile during device-side compilation.
+    if (LangOpts.OpenMPIsDevice && PP.getAuxTargetInfo())
       InitializePredefinedAuxMacros(*PP.getAuxTargetInfo(), LangOpts, Builder);
 
     // Install definitions to make Objective-C++ ARC work well with various
     // C++ Standard Library implementations.
     if (LangOpts.ObjC1 && LangOpts.CPlusPlus &&
         (LangOpts.ObjCAutoRefCount || LangOpts.ObjCWeak)) {
       switch (InitOpts.ObjCXXARCStandardLibrary) {
       case ARCXX_nolib: