[LLVM] Silence compiler-rt warning in runtimes build

[jhuber6]

Summary:
The compiler-rt project wants LLVM_CMAKE_DIR but the
llvm_ExternalProject_add interface sets the LLVM_CONFIG_PATH. This
patch just makes the utility pass that as well.

[compnerd]

How do LLVM_CONFIG_PATH and LLVM_CMAKE_PATH interact? If they don't then, we should be good.

[jhuber6]

How do LLVM_CONFIG_PATH and LLVM_CMAKE_PATH interact? If they don't then, we should be good.

Not sure myself, they seem to just point to different things and compiler-rt prefers the other one. I think it uses some CMake file operations to get the CMake path from the config path.

[compnerd]

  if (LLVM_CONFIG_PATH AND NOT LLVM_CMAKE_DIR)
    message(WARNING
      "LLVM_CONFIG_PATH is deprecated, please use LLVM_CMAKE_DIR instead")

It seems that they do interact, with LLVM_CMAKE_DIR taking precedence, which sounds good to me.

[jhuber6]

  if (LLVM_CONFIG_PATH AND NOT LLVM_CMAKE_DIR)
    message(WARNING
      "LLVM_CONFIG_PATH is deprecated, please use LLVM_CMAKE_DIR instead")

It seems that they do interact, with LLVM_CMAKE_DIR taking precedence, which sounds good to me.

Alright, thanks for the fast review.

While you're here, I need to get the compiler-rt tests running on the GPU before I can say it's "supported". One problem is that we don't actually create any tests when compiling the builtins through the LLVM_ENABLE_RUNTIMES= approach. This is because we include the compiler-rt/lib/builtins CMake directly. I think we will need to add some logic to configure the builtin tests even when done standalone, which would be step 1.

The next step would be adding logic for executing the tests under an "emulator", i.e. amdhsa-loader a.out in my case, as well as adding some minimal code to provide the "_start" function on the GPU (this is in libc but I don't want the dependency). I think the libcxx project has https://discord.com/channels/636084430946959380/636732535434510338/1263520163290943631 which takes %{exe} as an optional argument. Any thoughts?