[SYCL][Driver] Fix regression that enabled Cuda-mode in cc1 and defined __CUDA_ARCH__ (#15441)

The `CudaToolChain` set `-fcuda-is-device` unconditionally which made
`InitializePredefinedMacros` (called from
`clang::InitializePreprocessor`) to define `__CUDA_ARCH__` (default-init
to 1). As such, the driver assumed Cuda mode while in also SYCL mode,
but we don't properly support Cuda device-code compatibility and we want
to avoid having the `__CUDA_ARCH__` macro defined altogether for SYCL
offload.

Author: GeorgeWeb

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -938,8 +938,17 @@ void CudaToolChain::addClangTargetOptions(
           DeviceOffloadingKind == Action::OFK_Cuda) &&
          "Only OpenMP, SYCL or CUDA offloading kinds are supported for NVIDIA GPUs.");
 
-  CC1Args.append(
-      {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});
+  // If we are compiling SYCL kernels for Nvidia GPUs, we do not support Cuda
+  // device code compatability, hence we do not set Cuda mode in that instance.
+  if (DeviceOffloadingKind == Action::OFK_SYCL) {
+    toolchains::SYCLToolChain::AddSYCLIncludeArgs(getDriver(), DriverArgs,
+                                                  CC1Args);
+
+    if (DriverArgs.hasArg(options::OPT_fsycl_fp32_prec_sqrt))
+      CC1Args.push_back("-fcuda-prec-sqrt");
+  } else {
+    CC1Args.append(
+        {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});
 
     // Unsized function arguments used for variadics were introduced in CUDA-9.0
     // We still do not support generating code that actually uses variadic
@@ -948,18 +957,10 @@ void CudaToolChain::addClangTargetOptions(
     if (CudaInstallation.version() >= CudaVersion::CUDA_90)
       CC1Args.push_back("-fcuda-allow-variadic-functions");
 
-    if (DriverArgs.hasArg(options::OPT_fsycl)) {
-      // Add these flags for .cu SYCL compilation.
+    // Add these flags for .cu SYCL compilation.
+    if (DeviceOffloadingKind == Action::OFK_Cuda &&
+        DriverArgs.hasArg(options::OPT_fsycl))
       CC1Args.append({"-std=c++17", "-fsycl-is-host"});
-    }
-
-  if (DeviceOffloadingKind == Action::OFK_SYCL) {
-    toolchains::SYCLToolChain::AddSYCLIncludeArgs(getDriver(), DriverArgs,
-                                                  CC1Args);
-
-    if (DriverArgs.hasArg(options::OPT_fsycl_fp32_prec_sqrt)) {
-      CC1Args.push_back("-fcuda-prec-sqrt");
-    }
   }
 
   auto NoLibSpirv = DriverArgs.hasArg(options::OPT_fno_sycl_libspirv) ||

clang/test/Preprocessor/sycl-macro.cpp

@@ -6,7 +6,10 @@
 // RUNx: %clang_cc1 %s -fsycl-id-queries-fit-in-int -fsycl-is-device -E -dM -fms-compatibility | FileCheck --check-prefix=CHECK-MSVC %s
 // RUN: %clang_cc1 -fno-sycl-id-queries-fit-in-int %s -E -dM | FileCheck \
 // RUN: --check-prefix=CHECK-NO-SYCL_FIT_IN_INT %s
-// RUN: %clang_cc1 %s  -triple nvptx64-nvidia-cuda -target-cpu sm_80 -fsycl-is-device -E -dM | FileCheck --check-prefix=CHECK-CUDA %s
+// RUN: %clang_cc1 %s  -triple nvptx64-nvidia-cuda -target-cpu sm_80 -fsycl-is-device -E -dM | FileCheck \
+// RUN: --check-prefix=CHECK-CUDA %s -DARCH_CODE=800
+// RUN: %clangxx %s -fsycl -nocudalib -fsycl-targets=nvptx64-nvidia-cuda -Xsycl-target-backend --offload-arch=sm_80 -E -dM | FileCheck \
+// RUN: --check-prefix=CHECK-CUDA-SYCL-DRIVER %s
 // RUN: %clang_cc1 %s  -triple amdgcn-amd-amdhsa -target-cpu gfx906 -fsycl-is-device -E -dM | FileCheck --check-prefix=CHECK-HIP %s
 
 // RUN: %clang_cc1 %s  -triple nvptx64-nvidia-cuda -target-cpu sm_90a -fsycl-is-device -E -dM | FileCheck --check-prefix=CHECK-CUDA-FEATURE %s
@@ -32,8 +35,10 @@
 // CHECK-NO-SYCL_FIT_IN_INT-NOT:#define __SYCL_ID_QUERIES_FIT_IN_INT__ 1
 // CHECK-SYCL-ID:#define __SYCL_ID_QUERIES_FIT_IN_INT__ 1
 
-// CHECK-CUDA:#define __SYCL_CUDA_ARCH__ 800
-// CHECK-CUDA-NOT:#define __CUDA_ARCH__ 800
+// CHECK-CUDA:#define __SYCL_CUDA_ARCH__ [[ARCH_CODE]]
+// CHECK-CUDA-NOT:#define __CUDA_ARCH__ {{[0-9]+}}
+
+// CHECK-CUDA-SYCL-DRIVER-NOT: #define __CUDA_ARCH__ {{[0-9]+}}
 
 // CHECK-HIP:#define __CUDA_ARCH__ 0