On Thu, Jun 05, 2025 at 03:45:00PM -0400, Raymond Mao wrote: > Hi Tom, > > On Thu, 5 Jun 2025 at 13:15, Tom Rini <tr...@konsulko.com> wrote: > > > > On Tue, Jun 03, 2025 at 07:42:55AM -0700, Raymond Mao wrote: > > > > > MbedTLS bignum module needs '__udivti3' which is a 128-bit division > > > function provided by the compiler runtime, typically libgcc for GCC or > > > clang_rt.builtins for Clang. > > > Thus 'clang_rt.builtins' library is required when building using Clang. > > > > > > This patch supports the location layout of clang_rt.builtins libraries > > > from both LLVM pre-15 pattern [1] and post-15(up to 20) pattern [2]. > > > > > > As a fact that starts from LLVM 20, no official prebuilt multi-targets > > > are included and users might need to cross-built it from source. [3] is > > > an example of instructions for building. > > > > > > [1] <INST_DIR>/lib/clang/<REV>/lib/linux/libclang_rt.builtins-<ARCH>.a > > > E.g. > > > lib/clang/14.0.0/lib/linux/libclang_rt.builtins-aarch64.a > > > lib/clang/14.0.0/lib/linux/libclang_rt.builtins-x86_64.a > > > > > > [2] > > > <INST_DIR>/lib/clang/<REV_MAJOR>/lib/<CROSSTOOL>/libclang_rt.builtins.a > > > E.g. > > > lib/clang/20/lib/aarch64-none-linux-gnu/libclang_rt.builtins.a > > > lib/clang/20/lib/x86_64-unknown-linux-gnu/libclang_rt.builtins.a > > > > > > [3] > > > git clone https://github.com/llvm/llvm-project.git > > > cd llvm-project > > > git checkout -b llvmorg-20.1.6 llvmorg-20.1.6 > > > mkdir build-builtin && cd build-builtin > > > cmake -G Ninja ../llvm \ > > > -DCMAKE_BUILD_TYPE=Release \ > > > -DCMAKE_C_COMPILER=clang \ > > > -DLLVM_ENABLE_PROJECTS="clang" \ > > > -DLLVM_ENABLE_RUNTIMES="compiler-rt" \ > > > -DLLVM_TARGETS_TO_BUILD="X86;AArch64" \ > > > -DLLVM_BUILTIN_TARGETS=\ > > > "x86_64-unknown-linux-gnu;aarch64-none-linux-gnu" \ > > > -DBUILTINS_aarch64-none-linux-gnu_CMAKE_SYSTEM_NAME=Linux \ > > > -DBUILTINS_aarch64-none-linux-gnu_CMAKE_C_COMPILER=\ > > > <YOUR_TOOLCHAIN_GCC> \ > > > -DBUILTINS_aarch64-none-linux-gnu_CMAKE_ASM_COMPILER=\ > > > <YOUR_TOOLCHAIN_GCC> \ > > > -DBUILTINS_aarch64-none-linux-gnu_CMAKE_SYSROOT=\ > > > <YOUR_TOOLCHAIN_LIBC_DIR> \ > > > -DCOMPILER_RT_BUILD_BUILTINS=ON \ > > > -DCOMPILER_RT_BUILD_SANITIZERS=OFF \ > > > -DCOMPILER_RT_BUILD_XRAY=OFF \ > > > -DCOMPILER_RT_BUILD_LIBFUZZER=OFF \ > > > -DCOMPILER_RT_BUILD_PROFILE=OFF \ > > > -DCMAKE_INSTALL_PREFIX=<INSTALL_DIR> \ > > > -DLLVM_ENABLE_PER_TARGET_RUNTIME_DIR=OFF > > > ninja builtins > > > sudo ninja install > > > > > > Fixes: 13de8483388c ("mbedtls: add mbedtls into the build system") > > > Signed-off-by: Raymond Mao <raymond....@linaro.org> > > > > I think that newer versions don't even ship this moves us further away > > from a solution here as all of the logic. I think we need to look at the > > suggestion of just importing the assembly version of the function. > > But, the assembly versions are architecture-specific, do you want a > portable C code for it?
Is it easily done in C? If not, we can just try and make it clear what needs to be pulled out of LLVM sources if/when another platform encounters the problem. -- Tom
signature.asc
Description: PGP signature
