-mmacosx-version-min breaks arch detection for compiler-rt

[jfgoog]

I'm filing this bug as requested by @cjappl in #74394.

As noted in rust-lang/rust#122504 (comment), passing -mmacosx-version-min as a command line seems to break arch detection for compiler-rt.

In compiler-rt/cmake/config-ix.cmake we have:

  set(DARWIN_osx_MIN_VER_FLAG "-mmacosx-version-min")

  string(REGEX MATCH "${DARWIN_osx_MIN_VER_FLAG}=([.0-9]+)"
         MACOSX_VERSION_MIN_FLAG "${CMAKE_CXX_FLAGS}")

and then, later:

  if(NOT MACOSX_VERSION_MIN_FLAG)
    darwin_test_archs(osx
      DARWIN_osx_ARCHS
      ${toolchain_arches})
    message(STATUS "OSX supported arches: ${DARWIN_osx_ARCHS}")

So, specifying -mmacosx-version-min on Darwin results in no architectures being detected for compiler-rt, and the build therefore fails.

This seems incorrect, and is currently preventing the Rust compiler from updating a key dependency -- the cc crate for compiling C and C++ code.

@yln @vitalybuka and @petrhosek: can any of you shed some light here on whether this is expected behavior?

[jfgoog]

The if(NOT MACOSX_VERSION_MIN_FLAG) logic appears to have been added in 361231e, https://reviews.llvm.org/D11820.

Relevant comment: "What's the behavior you want if MACOSX_VERSION_MIN_FLAG is set? I'm guessing you're going with 'just assume the user just wants OS X, since they passed that flag in the CXX flags'.
Shouldn't we at least test the OS X archs?"

The response is not inspiring: "The behavior if the user set macos-version-min explicitly is kinda complicated and probably broken. As it is now, I'm pretty sure we just don't build the sanitizers if you specify an explicit version min."

[pirama-arumuga-nainar]

Adding @llvm-beanz for input on how compiler-rt detection should work in this scenario.

[nico]

#87580 looks somewhat related. Do you still see the problem after that?

Edit: Please ignore, I just see that you commented over there that that PR caused this issue.

[jfgoog]

The PR didn't cause the issue, but it doesn't fix it either. It was in response to a bug I filed about the behavior not being consistent between builds due to a cached cmake variable.