使用pybind11打包C++库Python绑定遇-lstdc++找不到问题（Mac Mojave）

Let's break down why this issue pops up and walk through actionable fixes to get your package installing smoothly via pip:

Root Cause #

macOS Mojave shifted Apple's default C++ standard library from libstdc++ to libc++, but older Anaconda environments (like your Python 3.6 setup) still ship with toolchains that expect libstdc++. Your CMake config tries to force libc++, but when pip runs the build, the Python environment's built-in build flags can override this setting—creating a linker mismatch that triggers the error.

Solution 1: Harden CMakeLists.txt to Enforce libc++ #

Update your CMake configuration to explicitly lock in libc++ and block any conflicting flags from the Python environment:

cmake_minimum_required(VERSION 3.10) # Bump to a newer CMake version for better pybind11 support
project(xxx)
set(CMAKE_BUILD_TYPE Release)

# Force libc++ as the standard library and disable libstdc++ entirely
set(CMAKE_CXX_STANDARD 11)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CXX_EXTENSIONS OFF)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -stdlib=libc++")
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -stdlib=libc++ -lc++")
set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -stdlib=libc++ -lc++")

# Ensure pybind11 adheres to our standard library settings
find_package(pybind11 REQUIRED)
include_directories(...)
pybind11_add_module(xxx xxx/xxx.cpp)

Solution 2: Pass Environment Variables in setup.py #

Modify your build_cmake method to explicitly set macOS-specific environment variables that force libc++, even when pip initiates the build:

def build_cmake(self, ext):
    build_args = ['--config', 'Release', '--', '-j4']
    cwd = pathlib.Path().absolute()
    build_temp = pathlib.Path(self.build_temp)
    build_temp.mkdir(parents=True, exist_ok=True)
    
    # Override environment variables to enforce libc++
    env = os.environ.copy()
    env['CXXFLAGS'] = '-stdlib=libc++'
    env['LDFLAGS'] = '-stdlib=libc++'
    
    os.chdir(str(build_temp))
    # Pass the modified environment to CMake
    self.spawn(['cmake', str(cwd)], env=env)
    if not self.dry_run:
        self.spawn(['cmake', '--build', '.'] + build_args, env=env)
    os.chdir(str(cwd))

Solution 3: Fix Anaconda Python 3.6 Environment Quirks #

For older Anaconda envs, you may need to point the build system to the Mojave SDK explicitly:

1. First, find your SDK path:

   xcrun --show-sdk-path
   

2. Set the environment variable before running pip install:

   export CONDA_BUILD_SYSROOT=$(xcrun --show-sdk-path)
   pip install my-package
   

You can also add this environment variable directly in your setup.py's build_cmake method to make this automatic for users.

Solution 4: Verify Wheel Compatibility (For Distribution) #

If you're building wheels for public distribution, use delocate to bundle the correct libc++ library into the wheel, so users don't hit linker issues:

pip install delocate
delocate-wheel -w fixed_wheels/ dist/*.whl

Final Notes #

• The reason Python 3.4/3.7 works but 3.6 fails is because newer Python versions on macOS use the system's default libc++ toolchain, while older Anaconda 3.6 environments are still configured for libstdc++.
• Bumping CMake to 3.10+ improves pybind11 integration and standard library handling.

内容的提问来源于stack exchange，提问作者craig zhang