cc: make CXX_FLAG configurable

[evertedsphere]

This offers an alternative to the solution proposed in #219 to the problem of nixpkgs_cc_configure providing a non-configurable -x c++ in CXX_OPTS when calling a C compiler. We add a new optional cc_lang argument to nixpkgs_cc_configure, and then pass -x <cc_lang> instead.

Fixes #218.

[dmadisetti]

Sure, this looks like the correct way to do things.

Plug and play into my project yields:

cannot find CUDA installation; provide its path via '--cuda-path', or pass '-nocudainc' to build without CUDA includes

While build a regular cc dependency. Not a massive blocker, I reckon this can be solved with setting environmental variables. But as in #219, auto-detection seems to work well enough so setting none seems to be best.

LGMT, but would suggest discussing the none option in the documentation since it is not well documented in clang (I had to refer to https://github.com/llvm-mirror/clang/blob/release_60/include/clang/Driver/Types.def)

[aherrmann]

Thanks for looking into this!

@dmadisetti does this solve the issue for you?

One issue I see with this approach is that it fixes the language for all targets built with that toolchain. Concretely, I'd expect the following type of thing to fail:

# WORKSPACE
nixpkgs_cc_configure(..., cc_lang = "cuda")
# BUILD
cc_binary(
    name = "cxx-bin",
    srcs = ["main.cc"],  # this is C++ code  -  I'd expect this to fail
)
cc_binary(
    name = "cude-bin",
    srcs = ["main.cu"],  # this is Cuda code
)

In that regard, bazelbuild/bazel#15499 seems like a much better solution.

That said, I'm not opposed to making the -x flag configurable, if it solves the issue.

[dmadisetti]

Yes setting cc_lang = none works for me!

[evertedsphere]

@dmadisetti perhaps cc_lang = none together with cc_library(copts = "-x whatever") is one possible way to go about it?

[aherrmann]

Yes setting cc_lang = none works for me!

Nice!

@dmadisetti perhaps cc_lang = none together with cc_library(copts = "-x whatever") is one possible way to go about it?

@evertedsphere if that works, then that would be a good thing to recommend in the docs, instead of setting cuda at the top-level.

[dmadisetti]

It's not needed from what I can see.
However, I did some basic testing to make sure this works. It looks like the none flag is properly overloaded.

[evertedsphere]

@dmadisetti what about without none? Does using copts = "-x cuda" with a vanilla nixpkgs_cc_configure toolchain work?

[dmadisetti]

@evertedsphere no, the conflicting flags fall back to the original

[dmadisetti]

LGTM

[aherrmann]

Nice, thank you!