[PATCH] D28320: [Driver] Driver changes to support CUDA compilation on Windows.

[Justin Lebar via cfe-commits]

That test should be updated to explicitly specify the triple, that
should also fix the problem.  I'll spin that change as soon as I can.

I agree that using the triple to determine the expected directory
layout is kind of bogus.  I have no idea if cross-compiling CUDA is
even going to work...

Do we have prior art for looking at the host (not the host triple, but
the host OS itself) to determine file paths?  If so maybe I should
just do that.

On Fri, Jan 6, 2017 at 2:26 AM, NAKAMURA Takumi via Phabricator
<reviews at reviews.llvm.org> wrote:
> chapuni added a comment.
>
> It fails with LLVM_DEFAULT_TARGET_TRIPLE=i686-pc-win32, on linux.
> http://bb.pgr.jp/builders/ninja-x64-msvc-RA-centos6/builds/33445
> Seems it assumes Windows' layout even if the host is not Windows.
>
>   128838 stat64("/path/to/llvm-project/clang/test/Driver/Inputs/CUDA/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v8.0", 0xffc9bdc0) = -1 ENOENT (No such file or directory)
>
> If you wouldn't consider such a configuration, I could add exclusion to lit.cfg.
>
>
> Repository:
>   rL LLVM
>
> https://reviews.llvm.org/D28320
>
>
>