[RFC] Amending `runtimes` build system to be able to build all Apple platforms (not only macOS)

Clang Frontend
,
1

Good morning,
I am working on teachin the build system for the Swift compiler to use LLVM_ENABLE_RUNTIMES instead of LLVM_BUILD_EXTERNAL_COMPILER_RT 1 – in doing so I realized that we may need a change in LLVM to avoid silently dropping build products for some Apple platforms, and I seeking your input to assess if the solution 2 I have in mind goes in the right direction.

The issues seems to stem from the fact that the llvm_ExternalProject_Add invocation used to build the default targets for builtins and runtimes always specify a target triple – when building on macOS that’s set to <arch>-apple-macos<version>.
That is subsequently used by in every compiler invocation generated by CMake, including the ones used to detect valid architectures for other platforms like iOS – the detection fails as a result and we silent skip building for such platforms.

My current solution 2 would be to add a flag to decide whether we set TARGET_TRIPLE for those external project invocation (maintaining by default the current behaviour), and then on Swift side we set that appropriately – however I cannot gauge effectively if this will impact other platforms or other runtime products using this infrastructure.

cc @etcwilde @compnerd

2

cc @petrhosek @beanz

3

Can you share your CMake invocation for building runtimes with LLVM_ENABLE_RUNTIMES? We have been building runtimes for Apple platforms with LLVM_ENABLE_RUNTIMES as part of our toolchain for years and we never encountered this issue so I’m curious how your build differs from ours.

4

We do something similar, except we add x86_64-apple-darwin to our LLVM_BUILTIN_TARGETS and LLVM_RUNTIME_TARGETS explicitly. Clang has special logic for overriding darwin triples with the appropriate OS and architecture, and the Apple runtime builds do so. ⚙ D86313 [runtimes] Allow LLVM_BUILTIN_TARGETS to include Darwin has some related discussion.

5

Answering to aid future searches and troubleshooting: this is the invocation that did not build iOS and iOS Sim – if I drop LLVM_HOST_TRIPLE, those are built

cmake \
 -G Ninja \
 -DLLVM_VERSION_MAJOR:STRING=17 \
 -DLLVM_VERSION_MINOR:STRING=0 \
 -DLLVM_VERSION_PATCH:STRING=0 \
 -DCLANG_VERSION_MAJOR:STRING=17 \
 -DCLANG_VERSION_MINOR:STRING=0 \
 -DCLANG_VERSION_PATCH:STRING=0 \
 -DLLVM_ENABLE_ASSERTIONS:BOOL=TRUE '-DLLVM_TARGETS_TO_BUILD=X86;AArch64' \
 -DCLANG_VENDOR=Apple \
 -DCLANG_VENDOR_UTI=com.apple.compilers.llvm.clang \
 -DPACKAGE_VERSION=17.0.0 \
 -DCMAKE_OSX_DEPLOYMENT_TARGET:STRING=13.0 \
 -DCOMPILER_RT_ENABLE_IOS:BOOL=ON \
 -DCOMPILER_RT_ENABLE_WATCHOS:BOOL=FALSE \
 -DCOMPILER_RT_ENABLE_TVOS:BOOL=FALSE \
 -DSANITIZER_MIN_OSX_VERSION=13.0 \
 -DLLVM_ENABLE_MODULES:BOOL=FALSE \
 -DCMAKE_OSX_ARCHITECTURES=arm64 \
 -DLLVM_TARGET_ARCH=AArch64 \
 -DLLVM_HOST_TRIPLE:STRING=arm64-apple-macosx13.0 \
 -DINTERNAL_INSTALL_PREFIX=local \
 '-DCMAKE_C_FLAGS=-Wno-unknown-warning-option -Werror=unguarded-availability-new -arch arm64 -target arm64-apple-macosx13.0 -fno-stack-protector' \
 '-DCMAKE_CXX_FLAGS=-Wno-unknown-warning-option -Werror=unguarded-availability-new -arch arm64 -target arm64-apple-macosx13.0 -fno-stack-protector' \
 -DCMAKE_BUILD_TYPE:STRING=Release \
 -DLLVM_TOOL_SWIFT_BUILD:BOOL=FALSE \
 -DLLVM_TOOL_LLD_BUILD:BOOL=TRUE \
 -DLLVM_INCLUDE_DOCS:BOOL=TRUE \
 -DCOMPILER_RT_INTERCEPT_LIBDISPATCH=ON '-DLLVM_ENABLE_PROJECTS=clang;clang-tools-extra;lld' \
 -DLLVM_ENABLE_RUNTIMES='compiler-rt' \
 -ULLVM_TOOL_COMPILER_RT_BUILD \
 -ULLVM_BUILD_EXTERNAL_COMPILER_RT \
 -DLLVM_APPEND_VC_REV=OFF '-DCMAKE_IGNORE_PATH=/usr/lib;/usr/local/lib;/lib' \
 -DPKG_CONFIG_EXECUTABLE=/usr/bin/false \
 -S llvm-project-llvm-org-main/llvm \
 -B ${build_folder} \

Using the Fuchsia configuration I am able to build runtimes for the iOS simulator, but not for iOS proper – possibly because we set CMAKE_OSX_ARCHITECTURES in RUNTIMES_CMAKE_ARGS, and that seems to interfere with the SDK detection

cmake -G Ninja \
      -DBUILTIN_TARGETS="default" \
  -DRUNTIME_TARGETS="default" \
  -DCOMPILER_RT_ENABLE_TVOS:BOOL=OFF \
  -DCOMPILER_RT_ENABLE_WATCHOS:BOOL=OFF \
  -DCOMPILER_RT_USE_BUILTINS_LIBRARY:BOOL=ON \
  -DLIBUNWIND_ENABLE_SHARED:BOOL=OFF \
  -DLIBUNWIND_USE_COMPILER_RT:BOOL=ON \
  -DLIBCXXABI_ENABLE_SHARED:BOOL=OFF \
  -DLIBCXXABI_ENABLE_STATIC_UNWINDER:BOOL=ON \
  -DLIBCXXABI_INSTALL_LIBRARY:BOOL=OFF \
  -DLIBCXXABI_USE_COMPILER_RT:BOOL=ON \
  -DLIBCXXABI_USE_LLVM_UNWINDER:BOOL=ON \
  -DLIBCXX_ABI_VERSION=2 \
  -DLIBCXX_ENABLE_SHARED:BOOL=OFF \
  -DLIBCXX_ENABLE_STATIC_ABI_LIBRARY:BOOL=ON \
  -DLIBCXX_HARDENING_MODE="none" \
  -DLIBCXX_USE_COMPILER_RT:BOOL=ON \
  -DRUNTIMES_CMAKE_ARGS="-DCMAKE_OSX_DEPLOYMENT_TARGET=10.13;-DCMAKE_OSX_ARCHITECTURES=arm64|x86_64" \
  -DBUILTINS_CMAKE_ARGS="--trace-expand;--debug-trycompile;--trace-redirect=llvm-main-build-fuchsia-reference.builtins.trace.txt" \
  -DCOMPILER_RT_ENABLE_IOS:BOOL=ON \
    -DCMAKE_BUILD_TYPE:STRING=Release \
    -DLLVM_ENABLE_PROJECTS='clang;clang-tools-extra;lld' \
    -DLLVM_TARGETS_TO_BUILD='X86;AArch64' \
    -DCMAKE_OSX_DEPLOYMENT_TARGET:STRING=10.13 \
    -DLLVM_ENABLE_RUNTIMES='compiler-rt' \
    -DLLVM_APPEND_VC_REV=OFF \
    -DCMAKE_IGNORE_PATH='/usr/lib;/usr/local/lib;/lib' \
    -DPKG_CONFIG_EXECUTABLE=/usr/bin/false \
    -S llvm-project-llvm-org-main/llvm -B ${build_folder}
6

Thanks for the feedback and apologies for my late acknowledgement – the approach described by @smeenai allows us to build all the Apple platforms, so my LLVM change is not needed.

On a slight different note – is there a date for the removal of compiler-rt as a valid value for LLVM_ENABLE_PROJECTS? I cannot see anything when searching in forums and Github (in particular Deprecate the use of `LLVM_ENABLE_PROJECTS` for all runtimes · Issue #124009 · llvm/llvm-project · GitHub)