The latest version of this document is available at https://android.googlesource.com/platform/ndk/+/master/docs/BuildSystemMaintainers.md.
The purpose of this guide is to instruct third-party build system maintainers in adding NDK support to their build systems. This guide will not be useful to most NDK users. NDK users should start with Building Your Project.
Note: This guide is written assuming Linux is the host OS. Mac should be no different, and the only difference on Windows is that file extensions for executables and scripts will differ.
The NDK uses Clang as its C/C++ compiler and Binutils for linking, archiving, and object file manipulation. Binutils provides both BFD and gold for linking. LLVM's LLD is also included for testing. AOSP uses LLD by default for most projects and the NDK is expected to move to it in the future.
Note: In general an architecture may have multiple ABIs. An ABI (application binary interface) is different from an architecture in that it also specifies a calling convention, size and alignment of types, and other implementation details. For Android, each architecture supports only one ABI.
Android supports multiple architectures: ARM32, ARM64, x86, and x86_64. NDK applications must build libraries for every architecture they support. 64-bit devices usually also support the 32-bit variant of their architecture, but this may not always be the case. While in general this means that an app with only 32-bit libraries can run on 64-bit capable devices, the 64-bit ABI will have improved performance.
This document will make use of
<triple> in describing paths and arguments. The values of these variables for each architecture are as follows except where otherwise noted:
Note: Strictly speaking ARMv7 with NEON is a different ABI from ARMv7 without NEON, but it is not a system ABI. Both NEON and non-NEON ARMv7 code uses the ARMv7 system and toolchains.
To programatically determine the list of supported ABIs, their bitness, as well as their deprecation status and whether or not it is recommended to build them by default, use
32-bit ARM can be built using either the Thumb or ARM instruction sets. Thumb code is smaller but may perform worse than ARM. However, smaller code makes more effective use of a processor‘s instruction cache, so benchmarking is necessary to determine which is more effective for a given application. ndk-build and the NDK’s CMake toolchain file generate Thumb code by default.
The ARM or Thumb instruction sets are selected by passing
-mthumb to Clang respectively. By default, Clang will generate ARM code as opposed to Thumb for the
Note: For ARMv7, Thumb-2 is used. Android no longer supports ARMv5, but if your build system mistakenly targets ARMv5 the less efficient Thumb-1 will be used.
Most ARM Android devices support NEON. This is supported by all 64-bit ARM devices and nearly all 32-bit ARM devices running at least Android Marshmallow (API 23). The Android CDD has required NEON support since that version, but it is possible that extant devices that were upgraded to Marshmallow do not include NEON support.
Enabling NEON can significantly improve application performance.
To enable NEON, pass
-mfpu=neon when compiling.
Note: The NDK's supported build systems (ndk-build and the CMake toolchain file) automatically enable NEON for API levels 23 (Marshmallow) and higher.
As users are distributed over a wide variety of Android OS versions (see the Distribution dashboard), applications have a minimum and maximum supported version, as well as a targeted version. These are
targetSdkVersion respectively. See the uses-sdk documentation for more information.
For NDK code, the only relevant value is the minimum supported version. Any time this doc refers to an API level, OS version, or target version, it is referring to the application's
The API level targeted by an NDK application determines which APIs will be exposed for use by the application. APIs that are not present in the targeted API level cannot be linked directly, but may be accessed via
dlsym. An NDK application running on a device with an API level lower than the target will often not load at all. If it does load, it may not behave as expected. This is not a supported configuration.
The major/minor version number given to an Android OS has no meaning when it comes to determining its API level. See the table in the Build numbers document to map Android code names and version numbers to API levels.
Note: Not every API level includes new NDK APIs. If there were no new NDK APIs for the given API level, there is no library directory for that API level. In that case, the build system should select the closest available API that is below the target API level. For example, applications with a
minSdkVersion of 20 should use API 19 for their NDK target.
To programatically determine the list of supported API levels as well as aliases that are accepted by ndk-build and CMake, see
Note: In some contexts the API level may be referred to as a platform. In this document an API level is always an integer, and a platform takes the form of
android-<API level>. The latter format is not specifically used anywhere in the NDK toolchain, but is used to specify target API levels for ndk-build and CMake.
Note: As a new version of the Android OS approaches release, previews and betas of that OS will be released and an NDK will be released that can make use of the new APIs. Targeting a preview API level is no different than targeting a released API level, with the exception that applications built targeting preview releases should not be shipped to production. Consult
<NDK>/meta/platforms.json to determine the API level for a preview release.
Clang is installed to
<NDK>/toolchains/llvm/prebuilt/<host-tag>/bin/clang. The C++ compiler is installed as
clang++ in the same directory.
clang++ will make C++ headers available when compiling and will automatically link the C++ runtime libraries when linking.
clang should be used when compiling C source files, and
clang++ should be used when compiling C++ source files. When linking,
clang should be used if the binary being linked contains no C++ code (i.e. none of the object files being linked were generated from C++ files) and
clang++ should be used otherwise. Using
clang++ ensures that the C++ standard library is linked.
Cross-compilation targets can be selected in one of two ways.
--target flag can be used (see the Clang User Manual for more details on Clang's supported arguments). The value passed is a Clang target triple suffixed with an Android API level. For example, to target API 26 for 32-bit ARM, use
Note: “armv7a” should be used rather than simply “arm” when specifying targets for Clang to generate ARMv7 code rather than the slower ARMv5 code. Specifying ARMv5 and thumb code generation will result in Thumb-1 being generated rather than Thumb-2, which is less efficient.
Second, a target-specific Clang can be used. In addition to the
clang++ binaries, there are also
<triple><API-level>-clang++ scripts. For example, to target API 26 32-bit ARM, invoke
armv7a-linux-androideabi26-clang++ instead of
Note: Target specific Clangs are currently implemented as shell scripts. Linux and Mac NDKs have Bash scripts, Windows includes Bash scripts to support Cygwin and WSL but also batch scripts (with
) for Windows command line support. For large numbers of relatively small source files, the additional overhead caused by these scripts may be noticably slower than using --target
, especially on Windows where CreateProcess
is slower thanfork`.
Gold is used by default for most architectures, but BFD is used for AArch64 as Gold emits broken debug information for that architecture (see Issue 70838247 for more details).
Note: It is usually not necessary to invoke the linkers directly since Clang will do so automatically. Clang will also automatically link CRT objects and default libraries and set up other target-specific options, so it is generally better to use Clang for linking.
The default linkers are installed to
<NDK>/toolchains/llvm/prebuilt/<host-tag>/<triple>/bin/ld. To use BFD or gold explicitly, use
ld.gold from the same locations.
ld.lld is not installed to the triple directory and is not triple-prefixed, but rather is only installed as
<NDK>/toolchains/llvm/prebuilt/<host-tag>/bin/ld.lld because the one binary supports all ABIs.
Warning: Multithreaded linking with LLD may hang on Windows. This feature is on by default. To workaround this,
-Wl,--no-threads must be passed to Clang when using LLD on Windows. See Issue 855.
The Android sysroot is installed to
<NDK>/toolchains/llvm/prebuilt/<host-tag>/sysroot and contains the headers, libraries, and CRT object files for each Android target.
Headers can be found in the
usr/include directory of the sysroot. Target specific include files are installed to
usr/include/<triple>. When using Clang, it is not necessary to include these directories explicitly; Clang will automatically select the sysroot. If using a compiler other than Clang, ensure that the target-specific include directory takes precedence over the target-generic directory.
Libraries are found in the
usr/lib/<triple> directory of the sysroot. Version-specific libraries are installed to
usr/lib/<triple>/<API-level>. As with the header files, when using Clang it is not necessary to include these directories explicitly; the sysroot will be automatically selected. If using a compiler other than Clang, ensure that the version-specific library directory takes precedence over the version-generic directory.
The NDK contains three types of libraries. Static libraries have a .a file extension and are linked directly into app binaries. Shared libraries have a .so file extension and must be included in the app's APK if used. System stub libraries are a special type of shared library that should not be included in the APK. The system stub libraries define the interface of a library that is provided by the Android OS but contain no implementation. They can be identified by their .so file extension and their presence in the system_libs list in ndk-build. The entries in this file are a key/value pair that maps library names to the first API level the library is introduced.
The STL provided by the NDK is libc++. Its headers are installed to
<NDK>/sysroot/usr/include/c++/v1. To use this STL, use the
-stdlib=libc++ flag. This STL comes in both a static and shared variant. The shared variant is used by default. To use the static variant, pass
-static-libstdc++ when linking. If using the shared variant, libc++_shared.so must be included in the APK. This library is installed to
Warning: There are a number of things to consider when selecting between the shared and static STLs. See the Important Considerations section of the C++ Support document for more details.
There are version-specific libc++.so and libc++.a libraries installed to
<NDK>/sysroot/usr/lib/<triple>/<version>. These are not true libraries but implicit linker scripts. They inform the linker how to properly link the STL for the given version. Older OS versions may require that a compatibility library (libandroid_support) be linked with libc++ to provide APIs not available in those versions. These scripts also handle the inclusion of any libc++ dependencies if necessary. Linker scripts should not be included in the APK.
Build systems should prefer to let Clang link the STL. If not using Clang, the version scripts should be used. Linking libc++ and its dependencies manually should only be used as a last resort.
Note: Linking libc++ and its dependencies explicitly may be necessary to defend against exception unwinding bugs caused by improperly built dependencies on ARM32 (see Issue 379). If not dependent on stack unwinding (the usual reason being that the application does not make use of C++ exceptions) or if no dependencies were improperly built, this is not necessary. If needed, link the libraries as listed in the linker script and be sure to follow the instructions in Unwinding.
The legacy “system STL” is also included, but it will be removed in a future NDK release. It is not in fact an STL; it contains only the barest C++ library support: the C++ versions of the C library headers and basic C++ runtime support like
delete. Its headers are installed to
<NDK>/toolchains/llvm/prebuilt/<host-tag>/include/c++/4.9.x and its library is the libstdc++.so system stub library. To use this STL, use the
TODO: Shouldn't it be installed to sysroot like libc++?
Note: The system STL will likely be removed in a future NDK release.
To avoid using the STL at all, pass
-nostdinc++ when compiling and
-nostdlib++ when linking. This is not necessary when using
clang, only when using
The NDK supports Address Sanitizer (ASan). This tool is similar to Valgrind in that it diagnoses memory bugs in a running application, but ASan is much faster than Valgrind (roughly 50% performance compared to an unsanitized application).
To use ASan, pass
-fsanitize=address when both compiling and linking. The sanitizer runtime libraries are installed to
<NDK>/toolchains/llvm/prebuilt/<host-tag>/lib64/clang/<clang-version>/lib/linux. The library is named
libclang_rt.asan-<arch>-android.so. This library must be included in the APK. A wrap.sh file must also be included in the APK. Premade wrap.sh files for ASan are installed to
Note: wrap.sh is only available for debuggable APKs running on Android Oreo (API 26) or higher. ASan can still be used devices prior to Oreo but at least Lollipop (API 21) if the device has been rooted. Direct users to the AddressSanitizerOnAndroid document for instructions on using this method.
Note: It is a bug that any of these need to be specified by the build system. All flags discussed in this section should be automatically selected by Clang, but they are not yet. Check back in a future NDK release to see if any can be removed from your build system.
32-bit ARM targets should use
-mfpu=vfpv3-d16 when compiling unless using NEON. This allows the compiler to make use of the FPU.
C++ builds should use
-stdlib=libc++ when using libc++. This flag is used both when compiling and when linking. This allows the compiler to find the correct C++ standard headers and libraries.
For x86 targets prior to Android Nougat (API 24),
-mstackrealign is needed to properly align stacks for global constructors. See Issue 635.
All code must be linked with
-Wl,-z,relro, which causes relocations to be made read-only after relocation is performed.
Android requires Position-independent executables beginning with API 21. Clang builds PIE executables by default. If invoking the linker directly or not using Clang, use
-pie when linking.
-faddrsig by default, but this produces output that is incompatible with GNU binutils. To workaround this,
-fno-addrsig must be passed to Clang when using GNU binutils. See Issue 884.
Clang does not properly set the ARMv7 architecture for the non-integrated assembler. If using
-fno-integrated-as, you must explicitly pass
-march=armv7-a when compiling for 32-bit ARM. Note that by default Clang will use the integrated assembler, and this flag is not needed in that case. See Issue 906.
It is recommended that
-Wl,--exclude-libs,<library file name> be used for each static library linked. This causes the linker to give symbols imported from a static library hidden visibility. This prevents a binary from unintentionally re-exporting an API other than its own. If the intent is to re-export all the symbols in a static library,
-Wl,--whole-archive <library> -Wl,--no-whole-archive should be used to ensure that the whole archive is preserved. By default, only symbols in used sections will be included in the linked binary.
If this behavior is not desired for your build system, ensure that these flags are at least used for libgcc.a and libunwind.a (libunwind is only used for ARM32). This is necessary to avoid unwinding bugs on ARM32. See Unwinding for more information.
To minimize the size of an APK, it may be desirable to use the
-Oz optimization mode. This will generate somewhat slower code than
-O3, but it will be smaller.
-Os behavior is not the same with Clang as it is with GCC. Clang‘s
-Oz behaves similarly to GCC’s
-Os with Clang is a middle ground between size and speed optimizations.
To aid the linker in removing as much unused code as possible, the compiler flags
-fdata-sections may be used. These flags should only be used in conjunction with the
-Wl,--gc-sections linker flag. Failing to use
-Wl,--gc-sections will cause the former flags to increase output size. The linker is only able to discard unused sections, so it can only discard at per-function or per-variable granularity if each is in its own section.
-Wl,--gc-sections should always be used, whether or not to enable
-fdata-sections depends on how the object file being compiled is expected to be used. If it will be used in a shared library then all of its public symbols will be preserved and the additional overhead of placing each item in its own section may make the shared library larger rather than smaller. If it will be used only in a static library or an executable then it will depend on how much of the resulting object file is expected to be unused.
It is recommended that build systems promote the following warnings to errors. These warnings indicate either a bug or undefined behavior, the latter of which Clang will usually turn into a bug.
-Werror=return-type: A non-void function is missing a return statement. Clang may “optimize” this function to fall through into the next one.
-Werror=pointer-to-int-cast: These indicate bugs that will affect the 64-bit version of the application.
-Werror=implicit-function-declaration: Undeclared functions may be inferred to have a return type of
intin C. For functions that return a pointer, the return type will be silently truncated to a 32-bit
int, resulting in bugs that will affect the 64-bit version of the application.
For more information on Clang's supported arguments, see the Clang User Manual.
For 32-bit ARM the NDK makes use of two unwinders: libgcc and LLVM's libunwind. libunwind is needed to provide C++ exception handling support. libgcc is needed to provide compiler runtime support and as such its unwinder is also seen by the linker.
These two unwinders are not ABI compatible but do use the same names, so caution is required to avoid ODR bugs. For 32-bit ARM, the libgcc.a in the NDK is a linker script that ensures that libunwind is linked before libgcc, causing the linker to prefer symbols from libunwind to those from libgcc.
As these are static libraries, the symbols will be included in the linked binary. By default they will be linked with public visibility. If used in a build system that does not strictly adhere to only linking shared libraries after all objects and static libraries, the binary being linked may instead load these symbols from a shared library. If this library was built with the wrong unwinder, it is possible for one unwinder to call into the other. As they are not compatible, this will likely result in either a crash or a failed unwind. To avoid this problem, libraries should always be built with
-Wl,--exclude-libs,libunwind.a (the latter is only necessary for 32-bit ARM) to ensure that unwind symbols are not re-exported from shared libraries.
Even with the above precautions, it is still possible for an improperly built external dependency to provide an incorrect unwind implementation as described in the above paragraph. The only way to guarantee protection against this for libraries built in your build system is to ensure that objects are linked in the following order:
-nostdlib when linking, crtend and crtbegin will be linked automatically by Clang. Linking libraries in the order above will require
-nostdlib++ when using libc++.
Command line length limits on Windows are short enough that they can pose problems when building large projects. Commands executed via cmd.exe are limited to 8,191 characters and commands executed with
CreateProcess are limited to 32,768 characters.
To work around these issues, Clang, the linkers, and the archiver all accept a response file that specifies the input files in place of specifying each input explicitly on the command line. Response files are identified on the command line with a “@” prefix and are formatted as space separated arguments. For example:
$ ar crsD liba.a @inputs.rsp
If the contents of
a.o b.o c.o then
ar will insert
Windows paths are limited to 260 characters, including the drive letter, colon, backslash, and terminating null. See Microsoft's documentation on path length limits for possible solutions.
Our experience shows that builds on Windows are generally slower than they are on Linux. The cost of
CreateProcess in comparison to
fork accounts for much of the difference, so it is best to minimize process creation in your build system.
File system performance can also make a large difference. This also appears to be the reason that Mac, while it has better build performance than Windows, still underperforms Linux.
Windows and Mac users will see optimum build performance in a Linux VM.