[CUDA] Disallow variable-length arrays in CUDA device code.

clang/include/clang/Basic/DiagnosticSemaKinds.td

 def err_dynamic_var_init : Error<
     "dynamic initialization is not supported for "
     "__device__, __constant__, and __shared__ variables.">;
 def err_shared_var_init : Error<
     "initialization is not supported for __shared__ variables.">;
 def err_device_static_local_var : Error<
     "Within a __device__/__global__ function, "
     "only __shared__ variables may be marked \"static\"">;
+def err_cuda_vla : Error<
+    "cannot use variable-length arrays in "
+    "%select{__device__|__global__|__host__|__host__ __device__}0 functions">;
+
 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 "
   "%select{function|block|method|constructor}2; expected type from format "

clang/include/clang/Sema/Sema.h

   ///
   /// 'try' and 'throw' are never allowed in CUDA __device__ functions, and are
   /// allowed in __host__ __device__ functions only if those functions are never
   /// codegen'ed for the device.
   ///
   /// ExprTy should be the string "try" or "throw", as appropriate.
   bool CheckCUDAExceptionExpr(SourceLocation Loc, StringRef ExprTy);
 
+  bool CheckCUDAVLA(SourceLocation Loc);
+
   /// Finds a function in \p Matches with highest calling priority
   /// from \p Caller context and erases all functions with lower
   /// calling priority.
   void EraseUnwantedCUDAMatches(const FunctionDecl *Caller,
                                 SmallVectorImpl<FunctionDecl *> &Matches);
   void EraseUnwantedCUDAMatches(const FunctionDecl *Caller,
                                 SmallVectorImpl<DeclAccessPair> &Matches);
   void EraseUnwantedCUDAMatches(

clang/lib/Sema/SemaCUDA.cpp

if (Target == CFT_HostDevice && getLangOpts().CUDAIsDevice) {
     PartialDiagnostic ErrPD{PartialDiagnostic::NullDiagnostic()};
     ErrPD.Reset(diag::err_cuda_device_exceptions);
     ErrPD << ExprTy << Target << CurFn;
     CurFn->addDeferredDiag({Loc, std::move(ErrPD)});
     return false;
   }
   return true;
 }
+
+bool Sema::CheckCUDAVLA(SourceLocation Loc) {
+  assert(getLangOpts().CUDA && "Should only be called during CUDA compilation");
+  FunctionDecl *CurFn = dyn_cast<FunctionDecl>(CurContext);
+  if (!CurFn)
+    return true;
+  CUDAFunctionTarget Target = IdentifyCUDATarget(CurFn);
+  if (Target == CFT_Global || Target == CFT_Device) {
+    Diag(Loc, diag::err_cuda_vla) << Target;
+    return false;
+  }
+  if (Target == CFT_HostDevice && getLangOpts().CUDAIsDevice) {
+    PartialDiagnostic ErrPD{PartialDiagnostic::NullDiagnostic()};
+    ErrPD.Reset(diag::err_cuda_vla);
+    ErrPD << Target;
+    CurFn->addDeferredDiag({Loc, std::move(ErrPD)});
+    return false;
+  }
+  return true;
+}

clang/lib/Sema/SemaType.cpp

if (!ArraySize) {
     T = Context.getConstantArrayType(T, ConstVal, ASM, Quals);
   }
 
   // OpenCL v1.2 s6.9.d: variable length arrays are not supported.
   if (getLangOpts().OpenCL && T->isVariableArrayType()) {
     Diag(Loc, diag::err_opencl_vla);
     return QualType();
   }
+  // CUDA device code doesn't support VLAs.
+  if (getLangOpts().CUDA && T->isVariableArrayType() && !CheckCUDAVLA(Loc))
+    return QualType();
+
   // If this is not C99, extwarn about VLA's and C99 array size modifiers.
   if (!getLangOpts().C99) {
     if (T->isVariableArrayType()) {
       // Prohibit the use of VLAs during template argument deduction.
       if (isSFINAEContext()) {
         Diag(Loc, diag::err_vla_in_sfinae);
         return QualType();
       }

clang/test/SemaCUDA/vla-host-device.cu

+// RUN: %clang_cc1 -fcuda-is-device -verify -S %s -o /dev/null
+// RUN: %clang_cc1 -verify -DHOST %s -S -o /dev/null
+
+#include "Inputs/cuda.h"
+
+#ifdef HOST
+// expected-no-diagnostics
+#endif
+
+__host__ __device__ void hd(int n) {
+  int x[n];
+#ifndef HOST
+  // expected-error@-2 {{cannot use variable-length arrays in __host__ __device__ functions}}
+#endif
+}
+
+// No error because never codegen'ed for device.
+__host__ __device__ inline void hd_inline(int n) {
+  int x[n];
+}
+void call_hd_inline() { hd_inline(42); }

clang/test/SemaCUDA/vla.cu

+// RUN: %clang_cc1 -fcuda-is-device -fsyntax-only -verify %s
+// RUN: %clang_cc1 -fsyntax-only -verify -DHOST %s
+
+#include "Inputs/cuda.h"
+
+void host(int n) {
+  int x[n];
+}
+
+__device__ void device(int n) {
+  int x[n];  // expected-error {{cannot use variable-length arrays in __device__ functions}}
+}