[CUDA] Do a better job at detecting wrong-side calls.

Summary:
Move CheckCUDACall from ActOnCallExpr and BuildDeclRefExpr to
DiagnoseUseOfDecl.  This lets us catch some edge cases we were missing,
specifically around class operators.

This necessitates a few other changes:

 - Avoid emitting duplicate deferred diags in CheckCUDACall.

   Previously we'd carefully placed our call to CheckCUDACall such that
   it would only ever run once for a particular callsite.  But now this
   isn't the case.

 - Emit deferred diagnostics from a template
   specialization/instantiation's primary template, in addition to from
   the specialization/instantiation itself.  DiagnoseUseOfDecl ends up
   putting the deferred diagnostics on the template, rather than the
   specialization, so we need to check both.

Reviewers: rsmith

Subscribers: cfe-commits, tra

Differential Revision: https://reviews.llvm.org/D24573

llvm-svn: 283637

clang/lib/CodeGen/CodeGenModule.cpp

@@ -2923,6 +2923,10 @@ void CodeGenModule::EmitGlobalFunctionDefinition(GlobalDecl GD,
   // non-error diags here, because order can be significant, e.g. with notes
   // that follow errors.)
   auto Diags = D->takeDeferredDiags();
+  if (auto *Templ = D->getPrimaryTemplate()) {
+    auto TemplDiags = Templ->getAsFunction()->takeDeferredDiags();
+    Diags.insert(Diags.end(), TemplDiags.begin(), TemplDiags.end());
+  }
   bool HasError = llvm::any_of(Diags, [this](const PartialDiagnosticAt &PDAt) {
     return getDiags().getDiagnosticLevel(PDAt.second.getDiagID(), PDAt.first) >=
            DiagnosticsEngine::Error;