[CMake] No mechanism to set -Xcompiler options with Visual Studio CUDA integration?

[Brad King]

On 06/21/2017 09:08 AM, Hancox, James wrote:
> it almost seems like -Xcompiler flags are being deliberately stripped out.)

Unfortunately we have to strip them out to work around a bug in CUDA's VS
integration.

The -Xcompiler options are parsed for inclusion in AdditionalCompilerOptions
in the .vcxproj file by flag map entries here:

  https://gitlab.kitware.com/cmake/cmake/blob/v3.9.0-rc3/Source/cmVS10CudaFlagTable.h#L3-6

Then a subset of the options that have dedicated elements defined by the
CUDA VS integration are extracted by flag map entries here:

  https://gitlab.kitware.com/cmake/cmake/blob/v3.9.0-rc3/Source/cmVS10CudaHostFlagTable.h

Ideally we should then leave the rest in AdditionalCompilerOptions, but
CUDA's VS integration puts them in the wrong place and breaks builds.
Therefore we have to drop them:

  https://gitlab.kitware.com/cmake/cmake/blob/v3.9.0-rc3/Source/cmVisualStudio10TargetGenerator.cxx#L2508-2510

We've reported this to NVIDIA but can't take the workaround out until
a version of CUDA comes out that fixes it.

-Brad