Hermetic toolchain seems to pick up cc compiler executable but not flags

[calebzulawski]

This probably works in a lot of default cases, but not generally. For example, our compiler (portable_cc_toolchain) uses a sysroot that requires the flags --sysroot=path/to/sysroot --gcc-toolchain=path/to/gcc among a few other required linker flags. I think the correct implementation would use something like get_memory_inefficient_command_line for both compile and link actions, and add those flags via -Xcompiler and -Xlinker.

[udaya2899]

Did you try with the new --repo-env CUDA_COMPILER_USE_CC_TOOLCHAIN=true flag?

Edit: I think this repo_env var works only with clang.

To actually use the cc_toolchain and not just the executable, in your custom cuda_toolchain target, you need to set compiler_use_cc_toolchain: True and make sure to NOT set compiler_executable or compiler_label attrs.