[CUDA] Bail, rather than crash, on va_arg in device code.

[jlebar]

Rename test file.

[tra]

Small nit. LGTM otherwise.

lib/Sema/SemaExpr.cpp

11732

Currently CUDAIsDevice does not imply that .CUDA is set. You may end up taking
this path if -fcuda-is-device is passed to C++ compilation.

For now add .CUDA check. We'll fix CUDA options processing separately.

11732–11739

Unneeded {}.

[jlebar]

Fixed and submitted.

include/clang/Basic/DiagnosticSemaKinds.td

def warn_host_calls_from_host_device : Warning<
   "calling __host__ function %0 from __host__ __device__ function %1 can lead to runtime errors">,
   InGroup<CudaCompat>;
 def warn_kern_is_method : Extension<
   "kernel function %0 is a member function; this may not be accepted by nvcc">,
   InGroup<CudaCompat>;
 def warn_kern_is_inline : Warning<
   "ignored 'inline' attribute on kernel function %0">,
   InGroup<CudaCompat>;
+def err_va_arg_in_device : Error<
+  "CUDA device code does not support va_arg">;
 
 def warn_non_pod_vararg_with_format_string : Warning<
   "cannot pass %select{non-POD|non-trivial}0 object of type %1 to variadic "
   "%select{function|block|method|constructor}2; expected type from format "
   "string was %3">, InGroup<NonPODVarargs>, DefaultError;
 // The arguments to this diagnostic should match the warning above.
 def err_cannot_pass_objc_interface_to_vararg_format : Error<
   "cannot pass object with interface type %1 by value to variadic "

lib/Sema/SemaExpr.cpp

for (const auto &CI : Result->getBlockDecl()->captures()) {
         break;
       }
     }
   }
 
   return Result;
 }
 
-ExprResult Sema::ActOnVAArg(SourceLocation BuiltinLoc,
-                                        Expr *E, ParsedType Ty,
-                                        SourceLocation RPLoc) {
+ExprResult Sema::ActOnVAArg(SourceLocation BuiltinLoc, Expr *E, ParsedType Ty,
   TypeSourceInfo *TInfo;
   GetTypeFromParser(Ty, &TInfo);
   return BuildVAArgExpr(BuiltinLoc, E, TInfo, RPLoc);
 }
 
 ExprResult Sema::BuildVAArgExpr(SourceLocation BuiltinLoc,
                                 Expr *E, TypeSourceInfo *TInfo,
                                 SourceLocation RPLoc) {
   Expr *OrigExpr = E;
   bool IsMS = false;
 
+  // CUDA device code does not support varargs.
+  if (getLangOpts().CUDAIsDevice) {
+    if (const FunctionDecl *F = dyn_cast<FunctionDecl>(CurContext)) {
+      CUDAFunctionTarget T = IdentifyCUDATarget(F);
+      if (T == CFT_Global || T == CFT_Device || T == CFT_HostDevice) {
+        return ExprError(Diag(E->getLocStart(), diag::err_va_arg_in_device));
+      }
+    }
+  }
+
   // It might be a __builtin_ms_va_list. (But don't ever mark a va_arg()
   // as Microsoft ABI on an actual Microsoft platform, where
   // __builtin_ms_va_list and __builtin_va_list are the same.)
   if (!E->isTypeDependent() && Context.getTargetInfo().hasBuiltinMSVaList() &&
       Context.getTargetInfo().getBuiltinVaListKind() != TargetInfo::CharPtrBuiltinVaList) {
     QualType MSVaListType = Context.getBuiltinMSVaListType();
     if (Context.hasSameType(MSVaListType, E->getType())) {
       if (CheckForModifiableLvalue(E, BuiltinLoc, *this))

test/SemaCUDA/va-arg.cu

+// REQUIRES: x86-registered-target
+// REQUIRES: nvptx-registered-target
+// RUN: %clang_cc1 -triple nvptx64-nvidia-cuda -fcuda-is-device -fsyntax-only \
+// RUN:   -verify -DEXPECT_ERR %s
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fsyntax-only %s
+
+#include <stdarg.h>
+#include "Inputs/cuda.h"
+
+__device__ void foo() {
+  va_list list;
+  va_arg(list, int);
+#ifdef EXPECT_ERR
+  // expected-error@-2 {{CUDA device code does not support va_arg}}
+#endif
+}
+
+void bar() {
+  va_list list;
+  va_arg(list, int);  // OK: host-only
+}
+
+__device__ void baz() {
+#if !defined(__CUDA_ARCH__)
+  va_list list;
+  va_arg(list, int);  // OK: only seen when compiling for host
+#endif
+}