Anaconda compiler tools#

Anaconda 5.0 switched from OS-provided compiler tools to our own toolsets. This allows improved compiler capabilities, including better security and performance. This page describes how to use these tools and enable these benefits.

Compiler packages#

Before Anaconda 5.0, compilers were installed using system tools such as XCode or yum install gcc. Now there are conda packages for Linux and macOS compilers. Unlike the previous GCC 4.8.5 packages that included GCC, g++, and GFortran all in the same package, these conda packages are split into separate compilers:

macOS:

  • clang_osx-64.

  • clangxx_osx-64.

  • gfortran_osx-64.

Linux:

  • gcc_linux-64.

  • gxx_linux-64.

  • gfortran_linux-64.

A compiler's "build platform" is the platform where the compiler runs and builds the code.

A compiler's "host platform" is the platform where the built code will finally be hosted and run.

Notice that all of these package names end in a platform identifier which specifies the host platform. All compiler packages are specific to both the build platform and the host platform.

Using the compiler packages#

The compiler packages can be installed with conda. Because they are designed with (pseudo) cross-compiling in mind, all of the executables in a compiler package are "prefixed." Instead of gcc, the executable name of the compiler you use will be something like x86_64-conda_cos6-linux-gnu-gcc. These full compiler names are shown in the build logs, recording the host platform and helping prevent the common mistake of using the wrong compiler.

Many build tools such as make and CMake search by default for a compiler named simply gcc, so we set environment variables to point these tools to the correct compiler.

We set these variables in conda activate.d scripts, so any environment in which you will use the compilers must first be activated so the scripts will run. Conda-build does this activation for you using activation hooks installed with the compiler packages in CONDA_PREFIX/etc/conda/activate.d, so no additional effort is necessary.

You can activate the root environment with the command conda activate root.

macOS SDK#

The macOS compilers require the macOS 10.9 SDK or above. The SDK license prevents it from being bundled in the conda package. We know of 2 current sources for the macOS SDKs:

We usually install the 10.10 SDK at /opt/MacOSX10.10.sdk but you may install it anywhere. Edit your conda_build_config.yaml file to point to it, like this:

CONDA_BUILD_SYSROOT:
  - /opt/MacOSX10.10.sdk        # [osx]

At Anaconda, we have this configuration setting in a centralized conda_build_config.yaml at the root of our recipe repository. Since we run build commands from that location, the file and the setting are used for all recipes. The conda_build_config.yaml search order is described further at Creating conda-build variant config files.

Build scripts for macOS should make use of the variables MACOSX_DEPLOYMENT_TARGET and CONDA_BUILD_SYSROOT, which are set by conda-build (see Environment variables). These variables should be translated into correct compiler arguments, e.g. for Clang this would be:

clang .. -isysroot ${CONDA_BUILD_SYSROOT} -mmacosx-version-min=${MACOSX_DEPLOYMENT_TARGET} ..

Most build tools, e.g. CMake and distutils (setuptools), will automatically pick up MACOSX_DEPLOYMENT_TARGET but you need to pass CONDA_BUILD_SYSROOT explicitly. For CMake, this can be done with the option -DCMAKE_OSX_SYSROOT=${CONDA_BUILD_SYSROOT}. When building Python extensions with distutils, one should always extend CFLAGS before calling setup.py:

export CFLAGS="${CFLAGS} -i sysroot ${CONDA_BUILD_SYSROOT}"

When building C++ extensions with Cython, CXXFLAGS must be similarly modified.

Backward compatibility#

Some users want to use the latest Anaconda packages but do not yet want to use the Anaconda compilers. To enable this, the latest Python package builds have a default _sysconfigdata file. This file sets the compilers provided by the system, such as gcc and g++, as the default compilers. This way allows legacy recipes to keep working.

Python packages also include an alternative _sysconfigdata file that sets the Anaconda compilers as the default compilers. The Anaconda Python executable itself is made with these Anaconda compilers.

The compiler packages set the environment variable _PYTHON_SYSCONFIGDATA_NAME, which tells Python which _sysconfigdata file to use. This variable is set at activation time using the activation hooks described above.

The new _sysconfigdata customization system is only present in recent versions of the Python package. Conda-build automatically tries to use the latest Python version available in the currently configured channels, which normally gets the latest from the default channel. If you're using something other than conda-build while working with the new compilers, conda does not automatically update Python, so make sure you have the correct _sysconfigdata files by updating your Python package manually.

Anaconda compilers and conda-build 3#

The Anaconda 5.0 compilers and conda-build 3 are designed to work together.

Conda-build 3 defines a special jinja2 function, compiler(), to make it easy to specify compiler packages dynamically on many platforms. The compiler function takes at least 1 argument, the language of the compiler to use:

requirements:
  build:
    - {{ compiler('c') }}

"Cross-capable" recipes can be used to make packages with a host platform different than the build platform where conda-build runs. To write cross-capable recipes, you may also need to use the "host" section in the requirements section. In this example we set "host" to "zlib" to tell conda-build to use the zlib in the conda environment and not the system zlib. This makes sure conda-build uses the zlib for the host platform and not the zlib for the build platform.

requirements:
  build:
    - {{ compiler('c') }}
  host:
    - zlib

Generally, the build section should include compilers and other build tools and the host section should include everything else, including shared libraries, Python, and Python libraries.

An aside on CMake and sysroots#

Anaconda's compilers for Linux are built with something called crosstool-ng. They include not only GCC, but also a "sysroot" with glibc, as well as the rest of the toolchain (binutils). Ordinarily, the sysroot is something that your system provides, and it is what establishes the libc compatibility bound for your compiled code. Any compilation that uses a sysroot other than the system sysroot is said to be "cross-compiling." When the target OS and the build OS are the same, it is called a "pseudo-cross-compiler." This is the case for normal builds with Anaconda's compilers on Linux.

Unfortunately, some software tools do not handle sysroots in intuitive ways. CMake is especially bad for this. Even though the compiler itself understands its own sysroot, CMake insists on ignoring that. We've filed issues at:

Additionally, this Stack Overflow issue has some more information: https://stackoverflow.com/questions/36195791/cmake-missing-sysroot-when-cross-compiling

In order to teach CMake about the sysroot, you must do additional work. As an example, please see our recipe for libnetcdf at AnacondaRecipes/libnetcdf-feedstock

In particular, you'll need to copy the cross-linux.cmake file there, and reference it in your build.sh file:

CMAKE_PLATFORM_FLAGS+=(-DCMAKE_TOOLCHAIN_FILE="${RECIPE_DIR}/cross-linux.cmake")

cmake -DCMAKE_INSTALL_PREFIX=${PREFIX} \
  ${CMAKE_PLATFORM_FLAGS[@]} \
  ${SRC_DIR}

Customizing the compilers#

The compiler packages listed above are small packages that only include the activation scripts and list most of the software they provide as runtime dependencies.

This design is intended to make it easy for you to customize your own compiler packages by copying these recipes and changing the flags. You can then edit the conda_build_config.yaml file to specify your own packages.

We have been careful to select good, general purpose, secure, and fast flags. We have also used them for all packages in Anaconda Distribution 5.0.0, except for some minor customizations in a few recipes. When changing these flags, remember that choosing the wrong flags can reduce security, reduce performance, and cause incompatibilities.

With that warning in mind, let's look at good ways to customize Clang.

  1. Download or fork the code from anacondarecipes/aggregate. The Clang package recipe is in the clang folder. The main material is in the llvm-compilers-feedstock folder.

  2. Edit clang/recipe/meta.yaml:

    package:
      name: clang_{{ target_platform }}
      version: {{ version }}
    

    The name here does not matter but the output names below do. Conda-build expects any compiler to follow the BASENAME_PLATFORMNAME pattern, so it is important to keep the {{target_platform}} part of the name.

    {{ version }} is left as an intentionally undefined jinja2 variable. It is set later in conda_build_config.yaml.

  3. Before any packaging is done, run the build.sh script: AnacondaRecipes/aggregate

    In this recipe, values are changed here. Those values are inserted into the activate scripts that are installed later.

    #!/bin/bash
    
    CHOST=${macos_machine}
    
    FINAL_CPPFLAGS="-D_FORTIFY_SOURCE=2 -mmacosx-version-min=${macos_min_version}"
    FINAL_CFLAGS="-march=core2 -mtune=haswell -mssse3 -ftree-vectorize -fPIC -fPIE -fstack-protector-strong -O2 -pipe"
    FINAL_CXXFLAGS="-march=core2 -mtune=haswell -mssse3 -ftree-vectorize -fPIC -fPIE -fstack-protector-strong -O2 -pipe -stdlib=libc++ -fvisibility-inlines-hidden -std=c++14 -fmessage-length=0"
    # These are the LDFLAGS for when the linker is being called directly, without "-Wl,"
    FINAL_LDFLAGS="-pie -headerpad_max_install_names"
    # These are the LDFLAGS for when the linker is being driven by a compiler, with "-Wl,"
    FINAL_LDFLAGS_CC="-Wl,-pie -Wl,-headerpad_max_install_names"
    FINAL_DEBUG_CFLAGS="-Og -g -Wall -Wextra -fcheck=all -fbacktrace -fimplicit-none -fvar-tracking-assignments"
    FINAL_DEBUG_CXXFLAGS="-Og -g -Wall -Wextra -fcheck=all -fbacktrace -fimplicit-none -fvar-tracking-assignments"
    FINAL_DEBUG_FFLAGS="-Og -g -Wall -Wextra -fcheck=all -fbacktrace -fimplicit-none -fvar-tracking-assignments"
    
    find "${RECIPE_DIR}" -name "*activate*.sh" -exec cp {} . \;
    
    find . -name "*activate*.sh" -exec sed -i.bak "s|@CHOST@|${CHOST}|g" "{}" \;
    find . -name "*activate*.sh" -exec sed -i.bak "s|@CPPFLAGS@|${FINAL_CPPFLAGS}|g"             "{}" \;
    find . -name "*activate*.sh" -exec sed -i.bak "s|@CFLAGS@|${FINAL_CFLAGS}|g"                 "{}" \;
    find . -name "*activate*.sh" -exec sed -i.bak "s|@DEBUG_CFLAGS@|${FINAL_DEBUG_CFLAGS}|g"     "{}" \;
    find . -name "*activate*.sh" -exec sed -i.bak "s|@CXXFLAGS@|${FINAL_CXXFLAGS}|g"             "{}" \;
    find . -name "*activate*.sh" -exec sed -i.bak "s|@DEBUG_CXXFLAGS@|${FINAL_DEBUG_CXXFLAGS}|g" "{}" \;
    find . -name "*activate*.sh" -exec sed -i.bak "s|@DEBUG_CXXFLAGS@|${FINAL_DEBUG_CXXFLAGS}|g" "{}" \;
    # find . -name "*activate*.sh" -exec sed -i.bak "s|@FFLAGS@|${FINAL_FFLAGS}|g"                 "{}" \;
    # find . -name "*activate*.sh" -exec sed -i.bak "s|@DEBUG_FFLAGS@|${FINAL_DEBUG_FFLAGS}|g"     "{}" \;
    find . -name "*activate*.sh" -exec sed -i.bak "s|@LDFLAGS@|${FINAL_LDFLAGS}|g"               "{}" \;
    find . -name "*activate*.sh" -exec sed -i.bak "s|@LDFLAGS_CC@|${FINAL_LDFLAGS_CC}|g"         "{}" \;
    find . -name "*activate*.sh.bak" -exec rm "{}" \;
    
  4. With those changes to the activate scripts in place, it's time to move on to installing things. Look back at the clang folder's meta.yaml. Here's where we change the package name. Notice what comes before the {{ target_platform }}.

    outputs:
      - name: super_duper_clang_{{ target_platform }}
        script: install-clang.sh
        requirements:
          - clang {{ version }}
    

    The script reference here is another place you might add customization. You'll either change the contents of those install scripts or change the scripts that those install scripts are installing.

    Note that we make the package clang in the main material agree in version with our output version. This is implicitly the same as the top-level recipe. The clang package sets no environment variables at all, so it may be difficult to use directly.

  5. Let's examine the script install-clang.sh:

    #!/bin/bash
    
    set -e -x
    
    CHOST=${macos_machine}
    
    mkdir -p "${PREFIX}"/etc/conda/{de,}activate.d/
    cp "${SRC_DIR}"/activate-clang.sh "${PREFIX}"/etc/conda/activate.d/activate_"${PKG_NAME}".sh
    cp "${SRC_DIR}"/deactivate-clang.sh "${PREFIX}"/etc/conda/deactivate.d/deactivate_"${PKG_NAME}".sh
    
    pushd "${PREFIX}"/bin
      ln -s clang ${CHOST}-clang
    popd
    

    Nothing here is too unusual.

    Activate scripts are named according to our package name so they won't conflict with other activate scripts.

    The symlink for Clang is a Clang implementation detail that sets the host platform.

    We define macos_machine in aggregate's conda_build_config.yaml: AnacondaRecipes/aggregate

    The activate scripts that are being installed are where we actually set the environment variables. Remember that these have been modified by build.sh.

  6. With any of your desired changes in place, go ahead and build the recipe.

    You should end up with a super_duper_clang_osx-64 package. Or, if you're not on macOS and are modifying a different recipe, you should end up with an equivalent package for your platform.

Using your customized compiler package with conda-build 3#

Remember the Jinja2 function, {{ compiler('c') }}? Here's where that comes in. Specific keys in conda_build_config.yaml are named for the language argument to that jinja2 function. In your conda_build_config.yaml, add this:

c_compiler:
  - super_duper_clang

Note that we're not adding the target_platform part, which is separate. You can define that key, too:

c_compiler:
  - super_duper_clang
target_platform:
  - win-64

With those two keys defined, conda-build will try to use a compiler package named super_duper_clang_win-64. That package needs to exist for your native platform. For example, if you're on macOS, your native platform is osx-64.

The package subdirectory for your native platform is the build platform. The build platform and the target_platform can be the same, and they are the same by default, but they can also be different. When they are different, you're cross-compiling.

If you ever needed a different compiler key for the same language, remember that the language key is arbitrary. For example, we might want different compilers for Python and for R within one ecosystem. On Windows, the Python ecosystem uses the Microsoft Visual C compilers, while the R ecosystem uses the Mingw compilers.

Let's start in conda_build_config.yaml:

python_c_compiler:
  - vs2015
r_c_compiler:
  - m2w64-gcc
target_platform:
  - win-64

In Python recipes, you'd have:

requirements:
  build:
    - {{ compiler('python_c') }}

In R recipes, you'd have:

requirements:
  build:
    - {{ compiler('r_c') }}

This example is a little contrived, because the m2w64-gcc_win-64 package is not available. You'd need to create a metapackage m2w64-gcc_win-64 to point at the m2w64-gcc package, which does exist on the msys2 channel on repo.anaconda.com.

Expressing the relation between compiler and its standard library#

For most languages, certainly for "c" and for "cxx", compiling any given program may create a run-time dependence on symbols from the respective standard library. For example, the standard library for C on linux is generally glibc, and a core component of your operating system. Conda is not able to change or supersede this library (it would be too risky to try to). A similar situation exists on MacOS and on Windows.

Compiler packages usually have two ways to deal with this dependence:

  • assume the package must be there (like glibc on linux).

  • always add a run-time requirement on the respective stdlib (e.g. libcxx on MacOS).

However, even if we assume the package must be there, the information about the glibc version is still a highly relevant piece of information, which is also why it is reflected in the __glibc virtual package.

For example, newer packages may decide over time to increase the lowest version of glibc that they support. We therefore need a way to express this dependence in a way that conda will be able to understand, so that (in conjunction with the __glibc virtual package) the environment resolver will not consider those packages on machines whose glibc version is too old.

The way to do this is to use the Jinja2 function {{ stdlib('c') }}, which matches {{ compiler('c') }} in as many ways as possible. Let's start again with the conda_build_config.yaml:

c_stdlib:
  - sysroot                     # [linux]
  - macosx_deployment_target    # [osx]
c_stdlib_version:
  - 2.17                        # [linux]
  - 10.13                       # [osx]

In the recipe we would then use:

requirements:
  build:
    - {{ compiler('c') }}
    - {{ stdlib('c') }}

This would then express that the resulting package requires sysroot ==2.17 (corresponds to glibc) on linux and macosx_deployment_target ==10.13 on MacOS in the build environment, respectively. How this translates into a run-time dependence can be defined in the metadata of the respective conda (meta-)package which represents the standard library (i.e. those defined under c_stdlib above).

In this example, sysroot 2.17 would generate a run-export on __glibc >=2.17 and macosx_deployment_target 10.13 would similarly generate __osx >=10.13. This way, we enable packages to define their own expectations about the standard library in a unified way, and without implicitly depending on some global assumption about what the lower version on a given platform must be.

In principle, this facility would make it possible to also express the dependence on separate stdlib implementations (like musl instead of glibc), or to remove the need to assume that a C++ compiler always needs to add a run-export on the C++ stdlib -- it could then be left up to packages themselves whether they need {{ stdlib('cxx') }} or not.

Anaconda compilers implicitly add RPATH pointing to the conda environment#

You might want to use the Anaconda compilers outside of conda-build so that you use the same versions, flags, and configuration, for maximum compatibility with Anaconda packages (but in a case where you want simple tarballs, for example). In this case, there is a gotcha.

Even if Anaconda compilers are used from outside of conda-build, the GCC specs are customized so that, when linking an executable or a shared library, an RPATH pointing to lib/ inside the current enviroment prefix directory ($CONDA_PREFIX/lib) is added. This is done by changing the link_libgcc: section inside GCC specs file, and this change is done so that LD_LIBRARY_PATH isn't required for basic libraries.

conda-build knows how to make this automatically relocatable, so that this RPATH will be changed to point to the environment where the package is being installed (at installation time, by conda). But if you only pack this binary in a tarball, it will continue containing this hardcoded RPATH to an environment in your machine. In this case, it is recommended to manually remove the RPATH.