[OpenMP] Don't throw cudalib not found error if only front-end is required.

[gtbercea]

Summary

If we only use the compiler front-end, do not throw an error about the cuda device library not being found. This allows the front-end to be run on systems where no Cuda installation is found.

[tra]

lib/Driver/ToolChains/Cuda.cpp

505–508

This could probably be rephrased as:

if (DeviceOffloadingKind == Action::OFK_OpenMP && 
    DriverArgs.hasArg(options::OPT_S, options::OPT_c))) ...

[gtbercea]

Contract check.

[gtbercea]

Fix parantheses.

[tra]

BTW, at least for CUDA compilation, '-c' would still needs libdevice as device-side will compile PTX to SASS and will need all the symbols PTX may refer to.
Would that not be the case for OpenMP's compilation, too?

[gtbercea]

Fix condition.

[tra]

Now we just need a test case to make sure this works as intended.

[gtbercea]

Add test.

[gtbercea]

Fix diff.

[tra]

I'm not particularly familiar with OpenMP internals. Could you elaborate on why libdevice would not be needed with -c for the OpenMP case?
Is that because it would only apply to the host compilation and that nothing will be compiled for the openmp targets?
Does openmp allow separate compilation for the target (i.e. something similar to what --cuda-device-only does?)

test/Driver/openmp-offload-gpu.c

133

I'd have separate runs -- one with -S and another one with -c.

That would cover openMP compilation changes. There should also be similar test to verify new behavior for CUDA compilation.

[gtbercea]

Only check for -S.

[gtbercea]

test/Driver/openmp-offload-gpu.c

133

There is no new behaviour for CUDA actually, I only introduce this for OpenMP offloading.

[hfinkel]

test/Driver/openmp-offload-gpu.c

131

Do you need both -c and -S? I'm under the impression that -c -S is equivalent to -S.

[gtbercea]

Open.

lib/Driver/ToolChains/Cuda.cpp

void CudaToolChain::addClangTargetOptions(
   }
 
   if (DriverArgs.hasArg(options::OPT_nocudalib))
     return;
 
   std::string LibDeviceFile = CudaInstallation.getLibDeviceFile(GpuArch);
 
   if (LibDeviceFile.empty()) {
+    if (DeviceOffloadingKind == Action::OFK_OpenMP &&
+        DriverArgs.hasArg(options::OPT_c) &&
+        DriverArgs.hasArg(options::OPT_S))
+      return;
+
     getDriver().Diag(diag::err_drv_no_cuda_libdevice) << GpuArch;
     return;
   }
 
   CC1Args.push_back("-mlink-cuda-bitcode");
   CC1Args.push_back(DriverArgs.MakeArgString(LibDeviceFile));
 
   // Libdevice in CUDA-7.0 requires PTX version that's more recent

test/Driver/openmp-offload-gpu.c

 /// ###########################################################################
 
 /// PTXAS is passed -c flag by default when offloading to an NVIDIA device using OpenMP
 /// Check that the flag is passed when -fopenmp-relocatable-target is used.
 // RUN:   %clang -### -fopenmp=libomp -fopenmp-targets=nvptx64-nvidia-cuda -fopenmp-relocatable-target -save-temps -no-canonical-prefixes %s 2>&1 \
 // RUN:   | FileCheck -check-prefix=CHK-PTXAS-RELO %s
 
 // CHK-PTXAS-RELO: ptxas{{.*}}" "-c"
+
+/// ###########################################################################
+
+/// Check that error is not thrown by toolchain when no cuda lib device is found when using -c -S.
+/// Check that the flag is passed when -fopenmp-relocatable-target is used.
+// RUN:   %clang -### -c -S -fopenmp=libomp -fopenmp-targets=nvptx64-nvidia-cuda -Xopenmp-target -march=sm_99 -fopenmp-relocatable-target -save-temps -no-canonical-prefixes %s 2>&1 \
+// RUN:   | FileCheck -check-prefix=CHK-NOLIBDEVICE %s
+
+// CHK-NOLIBDEVICE-NOT: error:{{.*}}sm_99