[cmake] Allow llvm_ExternalProject_Add to find CMAKE_SYSTEM_NAME among ${proj}_CMAKE_ARGS

[krisb]

Summary

llvm_ExternalProject_Add requires CMAKE_SYSTEM_NAME to be passed to set
correct toolchain tools, for example, for Windows.
But there are no other ways to do this for a default configuration of a
runtimes/builtins project, except using RUNTIMES_CMAKE_ARGS/BUILTINS_CMAKE_ARGS.
So, it is worth to account these variables while searching for CMAKE_SYSTEM_NAME.

[krisb]

Ping

[compnerd]

Hmm, this tries to extract CMAKE_SYSTEM_NAME from ..._CMAKE_ARGS if Im reading this correctly. Why not just say that you must pass CMAKE_SYSTEM_NAME via ..._CMAKE_ARGS rather than CMAKE_ARGS and just drop the clause above it?

[plotfi]

I am a little confused. Can you show an example of what would go in the cache file that would set what would otherwise be set by CMAKE_SYSTEM_NAME ?

[krisb]

@compnerd @plotfi thank you for the comments! Sorry if the description looks confusing. I'll try to be more clear.
I'm trying to build something like

cmake -GNinja -DCMAKE_BUILD_TYPE=Release -DLLVM_TARGETS_TO_BUILD=X86 -DLLVM_ENABLE_PROJECTS=clang -DLLVM_ENABLE_RUNTIMES=compiler-rt ../llvm

for host Windows under MSVC. This gives me the following error:

The Clang compiler tool

  "J:/llvm-project/build/./bin/clang.exe"

targets the MSVC ABI but has a GNU-like command-line interface.  This is
not supported.  Use 'clang-cl' instead, e.g.  by setting 'CC=clang-cl' in
the environment.

llvm_ExternalProject_Add sets to use clang-cl when _cmake_system_name is equal to Windows, but _cmake_system_name is defined only if -DCMAKE_SYSTEM_NAME is coming through CMAKE_ARGS.
This way perfectly works if runtimes/builtins are built for a known target, because we can pass RUNTIMES_${target}_CMAKE_SYSTEM_NAME and it will go to these CMAKE_ARGS.

Building for host (default) target there is no way to pass something to the CMAKE_ARGS.
I'm thinking... Maybe it would be better to do something similar for the default target and respect RUNTIMES_${var} or RUNTIMES_default_${var} inside the runtime_default_target() of llvm/runtimes/CMakeLists.txt (similarly as it is done for runtime_register_target())?
This can bring more flexibility.

I also found another attempt to handle CMAKE_SYSTEM_NAME for default target at https://reviews.llvm.org/D73811, but as far as I can see, it has been reverted.

In D80866#2080119, @compnerd wrote:

Hmm, this tries to extract CMAKE_SYSTEM_NAME from ..._CMAKE_ARGS if Im reading this correctly. Why not just say that you must pass CMAKE_SYSTEM_NAME via ..._CMAKE_ARGS rather than CMAKE_ARGS and just drop the clause above it?

Hmm. I'm not sure this would work for runtimes/builtins if they are being built for a particular target. This would require to set CMAKE_SYSTEM_NAME through ${tagret}_CMAKE_ARGS and not as RUNTIMES_${target}_CMAKE_SYSTEM_NAME which looks as a more preferred way to me.

In D80866#2081482, @plotfi wrote:

I am a little confused. Can you show an example of what would go in the cache file that would set what would otherwise be set by CMAKE_SYSTEM_NAME ?

Sorry, seems I didn't get your question. Did you mean what vars/options would be set depending on CMAKE_SYSTEM_NAME? If so, I'm mostly talking about a compiler (i.e. clang-cl.exe instead of clang.exe).

[krisb]

@plotfi @compnerd
I created another review with an alternative solution at D81877. I'll appreciate if you have a chance to look at it. Thank you.

llvm/cmake/modules/LLVMExternalProjectUtils.cmake

cmake_parse_arguments(ARG
     ${ARGN})
   canonicalize_tool_name(${name} nameCanon)
 
   foreach(arg ${ARG_CMAKE_ARGS})
     if(arg MATCHES "^-DCMAKE_SYSTEM_NAME=")
       string(REGEX REPLACE "^-DCMAKE_SYSTEM_NAME=(.*)$" "\\1" _cmake_system_name "${arg}")
     endif()
   endforeach()
+  if(NOT _cmake_system_name)
+    foreach(arg ${${nameCanon}_CMAKE_ARGS})
+      if(arg MATCHES "^-DCMAKE_SYSTEM_NAME=")
+        string(REGEX REPLACE "^-DCMAKE_SYSTEM_NAME=(.*)$" "\\1" _cmake_system_name "${arg}")
+      endif()
+    endforeach()
+  endif()
 
   if(NOT ARG_TOOLCHAIN_TOOLS)
     set(ARG_TOOLCHAIN_TOOLS clang lld llvm-ar llvm-lipo llvm-ranlib llvm-nm llvm-objdump)
     if(NOT _cmake_system_name STREQUAL Darwin)
       # TODO: These tools don't fully support Mach-O format yet.
       list(APPEND ARG_TOOLCHAIN_TOOLS llvm-objcopy llvm-strip)
     endif()
   endif()