[NVPTX] Add __CUDA_ARCH__ macro to standalone NVPTX compilations

[jhuber6]

Summary

We can now target the NVPTX architecture directly via
--target=nvptx64-nvidia-cuda. This currently does not define the
__CUDA_ARCH__ macro with is used to allow code to target different
codes based on support. This patch simply adds this support.

[jdoerfert]

LG

[tra]

clang/lib/Basic/Targets/NVPTX.cpp

171

Wouldn't just if(!HostTarget) be sufficient here?

clang/test/Frontend/nvptx-macros.c

1 ↗

(On Diff #508672)

I'd rename the file to make it more obvious that it deals with a standalone compilation.

standalone-nvptx-macros.c ?

[jhuber6]

clang/lib/Basic/Targets/NVPTX.cpp

171

HostTarget is the host toolchain, provided via -aux-triple. So it's set for OpenMP and CUDA but not for a standalone.

clang/test/Frontend/nvptx-macros.c

1 ↗

(On Diff #508672)

I can do that.

clang/lib/Basic/Targets/NVPTX.cpp

return llvm::StringSwitch<bool>(Feature)
       .Cases("ptx", "nvptx", true)
       .Default(false);
 }
 
 void NVPTXTargetInfo::getTargetDefines(const LangOptions &Opts,
                                        MacroBuilder &Builder) const {
   Builder.defineMacro("__PTX__");
   Builder.defineMacro("__NVPTX__");
-  if (Opts.CUDAIsDevice || Opts.OpenMPIsDevice) {
+  if (Opts.CUDAIsDevice || Opts.OpenMPIsDevice || !HostTarget) {
     // Set __CUDA_ARCH__ for the GPU specified.
     std::string CUDAArchCode = [this] {
       switch (GPU) {
       case CudaArch::GFX600:
       case CudaArch::GFX601:
       case CudaArch::GFX602:
       case CudaArch::GFX700:
       case CudaArch::GFX701:

clang/test/Frontend/standalone-nvptx-macros.c

+// REQUIRES: nvptx-registered-target
+
+// RUN: %clang %s -c -E -dM --target=nvptx64-nvidia-cuda -march=sm_70 -o - | \
+// RUN:   FileCheck --check-prefix=CHECK-CUDA-ARCH %s
+// CHECK-CUDA-ARCH: #define __CUDA_ARCH__ 700