Android Native CPU ABI Management
Introduction:
=============
Every piece of native code generated with the Android NDK matches a given
"Application Binary Interface" (ABI) that defines exactly how your
application's machine code is expected to interact with the system at
runtime.
A typical ABI describes things in *excruciating* details, and will typically
include the following information:
- the CPU instruction set that the machine code should use
- the endianness of memory stores and loads at runtime
- the format of executable binaries (shared libraries, programs, etc...)
and what type of content is allowed/supported in them.
- various conventions used to pass data between your code and
the system (e.g. how registers and/or the stack are used when functions
are called, alignment constraints, etc...)
- alignment and size constraints for enum types, structure fields and
arrays.
- the list of function symbols available to your machine code at runtime,
generally from a very specific selected set of libraries.
This document lists the exact ABIs supported by the Android NDK and the
official Android platform releases.
- - - -
I. Supported ABIs:
==================
Each supported ABI is identified by a unique name.
I.1. 'armeabi'
--------------
This is the name of an ABI for ARM-based CPUs that support *at* *least*
the ARMv5TE instruction set. Please refer to following documentation for
more details:
- ARM Architecture Reference manual (a.k.a ARMARM)
- Procedure Call Standard for the ARM Architecture (a.k.a. AAPCS)
- ELF for the ARM Architecture (a.k.a. ARMELF)
- ABI for the ARM Architecture (a.k.a. BSABI)
- Base Platform ABI for the ARM Architecture (a.k.a. BPABI)
- C Library ABI for the ARM Architecture (a.k.a. CLIABI)
- C++ ABI for the ARM Architecture (a.k.a. CPPABI)
- Runtime ABI for the ARM Architecture (a.k.a. RTABI)
- ELF System V Application Binary Interface
(DRAFT - 24 April 2001)
- Generic C++ ABI (http://mentorembedded.github.com/cxx-abi/abi.html)
Note that the AAPCS standard defines 'EABI' as a moniker used to specify
a _family_ of similar but distinct ABIs. Android follows the little-endian
ARM GNU/Linux ABI as documented in the following document:
> http://sourcery.mentor.com/sgpp/lite/arm/portal/kbattach142/arm_gnu_linux_abi.pdf
With the exception that wchar_t is only one byte. This should not matter
in practice since wchar_t is simply *not* really supported by the Android
platform anyway.
This ABI does *not* support hardware-assisted floating point computations.
Instead, all FP operations are performed through software helper functions
that come from the compiler's libgcc.a static library.
Thumb (a.k.a. Thumb-1) instructions are supported. Note that the NDK
will generate thumb code by default, unless you define LOCAL_ARM_MODE
in your Android.mk (see docs/ANDROID-MK.html for all details).
I.2. 'armeabi-v7a'
------------------
This is the name of another ARM-based CPU ABI that *extends* 'armeabi' to
include a few CPU instruction set extensions as described in the following
document:
- ARM Architecture v7-a Reference Manual
The instruction extensions supported by this Android-specific ABI are:
- The Thumb-2 instruction set extension.
- The VFP hardware FPU instructions.
More specifically, VFPv3-D16 is being used, which corresponds to 16
dedicated 64-bit floating point registers provided by the CPU.
Other extensions described by the v7-a ARM like Advanced SIMD (a.k.a. NEON),
VFPv3-D32 or ThumbEE are optional to this ABI, which means that developers
should check *at* *runtime* whether the extensions are available and provide
alternative code paths if this is not the case.
(Just like one typically does on x86 systems to check/use MMX/SSE2/etc...
specialized instructions).
You can check docs/CPU-FEATURES.html to see how to perform these runtime
checks, and docs/CPU-ARM-NEON.html to learn about the NDK's support for
building NEON-capable machine code too.
IMPORTANT NOTE: This ABI enforces that all double values are passed during
function calls in 'core' register pairs, instead of dedicated FP ones, via
switch -mfloat-abi=softfp. However, all internal computations can be performed
with the FP registers and will be greatly sped up.
This little constraint, while resulting in a slight decrease of
performance, ensures binary compatibility with all existing 'armeabi'
binaries.
Starting from r9b, it's possible to compile code in -mhard-float and still link
with Android native APIs which follow softfp. Please see
tests/device/hard-float/jni/Android.mk for details
IMPORTANT NOTE: The 'armeabi-v7a' machine code will *not* run on ARMv5 or
ARMv6 based devices.
I.3. 'x86'
----------
This is the name of an ABI for CPUs supporting the instruction set
commonly named 'x86' or 'IA-32'. More specifically, this ABI corresponds
to the following:
- instructions normally generated by GCC with the following compiler
flags:
-march=i686 -mtune=intel -mstackrealign -mssse3 -mfpmath=sse -m32
which targets Pentium Pro instruction set, according to the GCC
documentation, plus the MMX, SSE, SSE2, SSE3, SSSE3 instruction set
extensions. Generated code is a balanced optimization across top Intel
32-bit CPUs.
IMPORTANT NOTE: Flags above are not optimization guide. Compiler
optimization options which are used by default and/or recommended for
performance boost on x86 are not included. For performance optimization
hints on x86 GCC please refer to the following article:
> http://software.intel.com/blogs/2012/09/26/gcc-x86-performance-hints
- using the standard Linux x86 32-bit calling convention (e.g. section 6,
"Register Usage" of the "Calling conventions..." document below), not
the SVR4 one.
The ABI does *not* include any other optional IA-32 instruction set
extension, including, but not limited to:
- the MOVBE instruction
- any variant of "SSE4"
You can still use these, as long as you use runtime feature probing to
enable them, and provide fallbacks for devices that do not support them.
Please refer to the following documents for more details:
* http://gcc.gnu.org/onlinedocs/gcc/i386-and-x86_002d64-Options.html
* Calling conventions for different C++ compilers and operating systems:
> http://www.agner.org/optimize/calling_conventions.pdf
* Intel IA-32 Intel Architecture Software Developer's Manual
volume 2: Instruction Set Reference
* Intel IA-32 Intel Architecture Software Developer's Manual
volume 3: System Programming
* Amendment to System V Application Binary Interface
Intel386 Processor Architecture Supplement
I.4. 'mips'
-----------
This is the name of an ABI for MIPS-based CPUs that support *at* *least*
the MIPS32r1 instruction set. The ABI includes the following features:
- MIPS32 revision 1 ISA
- Little-Endian
- O32
- Hard-Float
- no DSP application specific extensions
Please refer to following documentation for more details:
- ELF for the MIPS Architecture (a.k.a. MIPSELF)
- FAQ for MIPS Toolchains (a.k.a. MIPSFAQ)
- Toolchain Specifics (a.k.a. MIPSTOOL)
- SDE Library (a.k.a. MIPSSDE)
- Instruction Set Quick Reference (a.k.a. MIPSISA)
- Architecture for Programmers (a.k.a. MIPSARCH)
- ELF System V Application Binary Interface
(DRAFT - 24 April 2001)
- Generic C++ ABI (http://sourcery.mentor.com/public/cxx-abi/abi.html)
The MIPS specific documentation is available at:
> http://www.mips.com/products/product-materials/processor/mips-architecture/
> https://sourcery.mentor.com/sgpp/lite/mips/portal/target_arch?@action=faq&target_arch=MIPS
Note: This ABI assumes a CPU:FPU clock ratio of 2:1 for maximum
compatibility.
Note: that MIPS16 support is not provided, nor is micromips.
I.5. 'x86-64'
----------
This is the name of an ABI for CPUs supporting the instruction set
commonly named 'x86-64'. More specifically, this ABI corresponds
to the following:
- instructions normally generated by GCC with the following compiler
flags:
-march=x86-64 -msse4.2 -mpopcnt -m64 -mtune=intel
which targets x86-64 instruction set, according to the GCC
documentation, plus the MMX, SSE, SSE2, SSE3, SSSE3, SSE4.1,
SSE4.2 and POPCNT instruction set extensions. Generated code is a balanced
optimization across top Intel 64-bit CPUs.
IMPORTANT NOTE: Flags above are not optimization guide. Compiler
optimization options which are used by default and/or recommended for
performance boost on x86-64 are not included. For performance
optimization hints on x86-64 GCC please refer to the following
article:
> http://software.intel.com/blogs/2012/09/26/gcc-x86-performance-hints
The ABI does *not* include any other optional x86-64 instruction set
extension, including, but not limited to:
- the MOVBE instruction
- the SHA instruction
- the AVX extension
- the AVX2 extension
You can still use these, as long as you use runtime feature probing to
enable them, and provide fallbacks for devices that do not support them.
Please refer to the following documents for more details:
* http://gcc.gnu.org/onlinedocs/gcc/i386-and-x86_002d64-Options.html
* Calling conventions for different C++ compilers and operating systems:
> http://www.agner.org/optimize/calling_conventions.pdf
* Intel64 and IA-32 Intel Architecture Software Developer's Manual
volume 2: Instruction Set Reference
* Intel64 and IA-32 Intel Architecture Software Developer's Manual
volume 3: System Programming
* Amendment to System V Application Binary Interface
AMD64 Processor Architecture Supplement
- - - -
II. Generating code for a specific ABI:
=======================================
By default, the NDK will generate machine code for the 'armeabi' ABI.
You can however add the following line to your Application.mk to generate
ARMv7-a compatible machine code instead:
APP_ABI := armeabi-v7a
It is also possible to build machine code for two or more distinct ABIs,
for example:
APP_ABI := armeabi armeabi-v7a
This will instruct the NDK to build two versions of your machine code: one for
each ABI listed on this line. Both libraries will be copied to your application
project path and will be ultimately packaged into your .apk.
Such a package is called a "fat binary" in Android speak since it contains
machine code for more than one CPU architecture. At installation time, the
package manager will only unpack the most appropriate machine code for the
target device. See below for details.
Also you can use:
APP_ABI := all
which will generate machine code for all supported ABIs with this NDK. Doing so
will ensure that your application package contains libraries for all target ABIs.
Note that this has an impact on package size, since each ABI will correspond to
its own set of native libraries built from the same sources.
- - - -
III. ABI Management on the Android platform:
============================================
This section provides specific details about how the Android platform manages
native code in application packages.
III.1. Native code in Application Packages:
-------------------------------------------
It is expected that shared libraries generated with the NDK are stored in
the final application package (.apk) at locations of the form:
lib/<abi>/lib<name>.so
Where <abi> is one of the ABI names listed in section II above, and <name>
is a name that can be used when loading the shared library from the VM
as in:
System.loadLibrary("<name>");
Since .apk files are just zip files, you can trivially list their content
with a command like:
unzip -l <apk>
to verify that the native shared libraries you want are indeed at the
proper location. You can also place native shared libraries at other
locations within the .apk, but they will be ignored by the system, or more
precisely by the steps described below; you will need to extract/install
them manually in your application.
In the case of a "fat" binary, up to four distinct libraries can be placed
in the .apk, for example at:
lib/armeabi/libfoo.so
lib/armeabi-v7a/libfoo.so
lib/x86/libfoo.so
lib/x86-64/libfoo.so
lib/mips/libfoo.so
IMPORTANT NOTE: ARMv7-based Android device running 4.0.3 or before installs native
library from the 'armeabi' directory instead of 'armeabi-v7a' directory if both
exist and 'lib/armeabi' is listed after 'lib/armeabi-v7a' in apk. This issue is
fixed in 4.0.4 or later.
III.2. Android Platform ABI support:
------------------------------------
The Android system knows at runtime which ABI(s) it supports. More
precisely, up to two build-specific system properties are used to
indicate:
- the 'primary' ABI for the device, corresponding to the machine
code used in the system image itself.
- an optional 'secondary' ABI, corresponding to another ABI that
is also supported by the system image.
To achieve the best performance for your NDK component, you should compile
directly for the primary ABI.
For example, a typical ARMv5TE-based device would only define
the primary ABI as '`armeabi`' and not define a secondary one.
On the other hand, a typical ARMv7-based device would define the
primary ABI to '`armeabi-v7a`' and the secondary one to '`armeabi`'
since it can run application native binaries generated for both
of them.
Many x86-based devices can also run armeabi-v7a and armeabi NDK
binaries and define the primary ABI to '`x86`' and the secondary
one to '`armeabi-v7a`'.
A typical MIPS-based device only defines a primary abi named '`mips`'.
III.3. Automatic extraction of native code at install time:
-----------------------------------------------------------
When installing an application, the package manager service will scan
the .apk and look for any shared library of the form:
lib/<primary-abi>/lib<name>.so
If one is found, then it is copied under `$APPDIR/lib/lib<name>.so`,
where `$APPDIR` corresponds to the application's specific data directory.
If none is found, and a secondary ABI is defined, the service will
then scan for shared libraries of the form:
lib/<secondary-abi>/lib<name>.so
If anything is found, then it is copied under `$APPDIR/lib/lib<name>.so`
This mechanism ensures that the best machine code for the target
device is automatically extracted from the package at installation
time.