[CUDA] pass -fno-threadsafe-statics to GPU sub-compilations.

We do not have support for the threadsafe statics on the GPU side.

However, we do sometimes end up with empty local static initializers,
and those happen to trigger calls to `__cxa_guard*`, which breaks compilation.

Partially addresses #117023

Author: Artem-B

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -856,8 +856,9 @@ void CudaToolChain::addClangTargetOptions(
           DeviceOffloadingKind == Action::OFK_Cuda) &&
          "Only OpenMP or CUDA offloading kinds are supported for NVIDIA GPUs.");
 
-  CC1Args.append(
-      {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});
+  CC1Args.append({"-fcuda-is-device", "-mllvm",
+                  "-enable-memcpyopt-without-libcalls",
+                  "-fno-threadsafe-statics"});
 
   // Unsized function arguments used for variadics were introduced in CUDA-9.0
   // We still do not support generating code that actually uses variadic

clang/test/Driver/cuda-no-threadsafe-statics.cu

@@ -0,0 +1,10 @@
+// Check that -fno-thread-safe-statics get passed down to device-side
+// compilation only.
+//
+// RUN: not %clang -### --target=x86_64-linux-gnu -c --cuda-gpu-arch=sm_20 %s 2>&1 \
+// RUN:   | FileCheck %s
+//
+// CHECK: "-fcuda-is-device"
+// CHECK-SAME: "-fno-threadsafe-statics"
+// CHECK: "-triple" "x86_64-unknown-linux-gnu"
+// CHECK-NOT: "-fno-threadsafe-statics"