The ARM EABI port is now available. See ArmEabiHowto and ArmEabiTodo.

Its progress towards inclusion in the Debian "lenny" release, scheduled for September 2008, is shown at armelLennyReleaseRecertification.

In a nutshell

EABI is the new "Embedded" ABI by ARM ltd. EABI is actually a family of ABI's and one of the "subABIs" is GNU EABI, for Linux. The effective changes for users are:

Terminology

Strictly speaking, both the old and new ARM ABIs are subsets of the ARM EABI specification, but in everyday usage the term "EABI" is used to mean the new one described here and "OABI" or "old-ABI" to mean the old one. However, there are one or two programs that sometimes describe an old ABI binary as "EABI".

To add to the confusion, powerpc has also had an ABI called "EABI" for some, which has nothing to do with this one.

GCC view

New ABI is not only a new ABI field, it is also a new GCC target.

Legacy ABI

private flags = 2: [APCS-32] [FPA float format] [has entry point]

ELF 32-bit LSB executable, ARM, version 1 (ARM), for GNU/Linux 2.2.0, dynamically linked (uses shared libs), for GNU/Linux 2.2.0, stripped

Flags: 0x0

Arm EABI:

private flags = 4000002: [Version4 EABI] [has entry point]

ELF 32-bit LSB executable, ARM, version 1 (SYSV), for GNU/Linux 2.4.17, dynamically linked (uses shared libs), for GNU/Linux 2.4.17, stripped

Flags: 0x4000002, has entry point, Version4 EABI

Furthermore, as well as the usual __arm__ and maybe also __thumb__ symbols, the C preprocessor symbol __ARM_EABI__ is also defined when compiling into EABI, while __ARMEL__ is predefined under both the new and old ABIs.

ARM floating points

The current Debian port creates hardfloat FPA instructions. FPA comes from "Floating Point Accelerator". Since the FPA floating point unit was implemented only in very few ARM cores, these days FPA instructions are emulated in kernel via Illegal instruction faults. This is of course very inefficient: about 10 times slower that -msoft-float for a FIR test program. The FPA unit also has the peculiarity of having mixed-endian doubles, which is usually the biggest grief for ARM porters, along with structure packing issues.

ARM has now introduced a new floating point unit, VFP (Vector Floating Points), which uses a different instruction set than FPA and stores floats in natural-endian IEEE-754 format. VFP is implemented in new some ARM9/10/11 cores, like in the new TI OMAP2 family. It seems likely that ARM cores without VFP will remain popular, as in many places ARM is used floats are unnecessary.

To complicate thing further, ARM processors are being integrated with many other FPUs and DSPs, each of which adds its own set of instructions to the ARM set:

For a generic-purpose distribution like Debian, targeting binary compatibility, EABI lets us have the cake and eat it. We can make a soft-float distribution using IEEE-754 with FPU-specific versions of packages (linux-kernel-2.6.x-vfp, libc6-iwmmxt, mediaplayer-maverick, etc) where needed. This also enables individual packages to do runtime FPU detection and call code compiled for different FP scenarios (in liboil for example).

The major FP variants worth support as alternative versions of FP-critical packages seem to be

GCC preprocessor macros for floating point

When porting code to ARM EABI, the following preprocessor macros are interesting:

__VFP_FP__ and __MAVERICK__ are mutually exclusive. If neither is set, that means the floating point format in use is the old mixed-endian 45670123 format of the FPA unit.

Note that __VFP_FP__ does not mean that VFP code generation has been selected. It only speaks of the floating point data format in use and is normally set when soft-float has been selected. The correct test for VFP code generation, for example around asm fragments containing VFP instructions, is

#if (defined(__VFP_FP__) && !defined(__SOFTFP__))

Paradoxically, the -mfloat-abi=softfp does not set the __SOFTFP___ macro, since it selects real floating point instructions using the soft-float ABI at function-call interfaces.

By default in Debian armel, __VFP_FP__ && __SOFTFP__ are selected.

Struct packing and alignment

With the new ABI, default structure packing changes, as do some default data sizes and alignment (which also have a knock-on effect on structure packing). In particular the minimum size and alignment of a structure was 4 bytes. Under the EABI there is no minimum and the alignment is determined by the types of the components it contains. This will break programs that know too much about the way structures are packed and can break code that writes binary files by dumping and reading structures.

Stack alignment

The ARM EABI requires 8-byte stack alignment at public function entry points, compared to the previous 4-byte alignment.

64-bit data type alignment

"One of the key differences between the traditional GNU/Linux ABI and the EABI is that 64-bit types (like long long) are aligned differently. In the traditional ABI, these types had 4-byte alignment; in the EABI they have 8-byte alignment. As a result, if you use the same structure definitions (in a header file) and include it in code used in both the kernel and in application code, you may find that the structure size and alignment differ."

Enum sizes

The EABI defines an optional system for controlling the size of C enumerated types. For arm-linux it was decided to keep the existing behaviour (enums are at least the same size as an int) for consistency with other Linux systems.

This is also reflected in the -mabi=aapcs or -mabi=aapcs-linux switches to GCC: aapcs defines enums to be a variable sized type, while with aapcs-linux they are always ints (4 bytes).

System call interface

Two things change in the system call interface: alignment of 64-bit parameters passed in registers and the way the system call number itself is passed.

With EABI, 64-bit function parameters passed in registers are aligned to an even-numbered register instead of using the next available pair.

Here's an explanation from Russell King, 12 Jan 2006:

We have r0 to r6 to pass 32-bit or 64-bit arguments. With EABI,
64-bit arguments will be aligned to an _even_ numbered register.
Hence:
long sys_foo(int a, long long b, int c, long long d);
will result in the following register layouts:
        EABI                            Current
r0      a                               a
r1      unused                          \_ b
r2      \_ b                            /
r3      /                               c
r4      c                               \_ d
r5                                      /
r6      ... out of space for 'd'        ... room for one other int.
r7      syscall number

Since this already causes an incompatible change in the system call interface, the opportunity has been taken to slip in a more efficient, totally incompatible way of doing system calls: instead of using the swi __NR_SYSCALL_BASE(==0x900000)+N instruction, where N is the number of the system call, swi 0 is always used with the system call number stashed in register r7. This is more efficient because the kernel no longer has to go and fish N out of the instruction stream(*), which used to have a negative impact on the efficiency of processors with separate text and data caches (i.e. most ARMs).

Fortunately, the two schemes can coexist and EABI kernels have an option to support the old syscall interface (including old structure layout rules) for running old-EABI binaries. However some features (e.g., ALSA, MD (RAID) and system calls from Thumb mode) do not work correctly from old-ABI binaries.

Some third party EABI toolchains (e.g., CodeSourcery 2005q3) use the old kernel interface via userspace shims in glibc. This is now obsolete and no longer supported by glibc.

(*) This is only true if the old-ABI compatibility option is disabled.

See this article for more details.

Choice of minimum CPU

Thumb interworking suggests armv4t

The EABI includes thumb interworking, which means that 16-bit Thumb and normal 32-bit ARM instructions can be mixed at function-level granularity.

Thumb interworking is mandatory according to the ARM EABI spec and requires every return and indirect function call to execute a BX instruction to set the core to the correct state, which is only present in armv4t cores and above. Gcc, too, only supports thumb interworking for armv4t and above.

So Debian armel runs on a minimum CPU of ARMv4t and by default the Debian armel GCC generates code for armv4t (rather than the usual default ARM target of armv5t).

Other scenarios

However a lower entry-level CPU is possible to do using different function return sequences which are of various speeds, and that work and/or allow Thumb interworking on different selections of the ARM architectures.

0.  mov pc,lr 

Is what GCC currently emits for -march=armv4. It works on ARMv4 and above but is only Thumb interworking-safe from ARMv7.

1.  bx lr 

Is what GCC emits for -march=armv4t. It works on ARMv4t and above and Thumb interworking is possible on ARMv4t and above. Excludes armv4, the StrongARM which are very common and some armv5 users, but armv5 with no t seems a rare processor. CC needs modifying to implement any of the following choices.

2.  tst lr, #1; moveq pc, lr; bx lr 

was suggested by Paul Brook as an alternative to BX. It works on ARMv4 and above and Thumb interworking is possible on ARMv4t and above, with the extra cost of two instructions per indirect call/function return, though in line with the run-on-minimum-hardware Debian way.

Here is a patch by Richard K. Pixley which illustrates what is needed, but is not (April 2007) tested: http://lists.debian.org/debian-arm/2007/05/msg00015.html

This is problematic because hand written assembly has to be manually fixed.

3.  ldm/ldr 

Works on ARMv4 and above but Thumb interworking is only possible on ARMv5t and above, excluding ARMv4t users from using Thumb code with Debian. Gcc currently emits this for non-leaf functions on ARMv4 and ARMv5 (but not ARMv4t, where it uses BX, the only way to do interworking on v4t). Although a single instruction, this method may be slower than the three-instruction sequence because of the memory accesses it requires.

4. Drop Thumb interworking

A final option would be simply to compile the standard Debian repo --with-arch=armv4 --with-no-thumb-interwork. This would work on all processors without the dangers inherent in modifying GCC and, according to the GCC manual page, saves a slight size and speed overhead caused by being thumb-interworkable.

There is significant discussion of the technical merits of these various schemes in the debian-arm mailing list thread Re: ARM EABI port: minimum CPU choice of which the above is a partial summary.

5. tat says that simply compile for armv3 would work, though the code would be relatively slower on the most common, later CPUs. armv3 is fairly rare: Psion 5

6. Kernel emulation traps

It may be possible to catch illegal instruction in the kernel generated by the missing "BX" instruction, in the same way as missing hard floating point instructions can be emulated. It wouldn't be that fast on armv4 hardware (causing a context switch per function call/return) but such a kernel hack would allow the current repository to be used unmodified on armv4 hardware.

7. Linker fixups

The EABI provides mechanisms (R_ARM_V4BX relocations) for the linker to fixup bx instructions. Currently the linker only knows how to convert these to mov pc instructions, so you have to choose between armv4 or interworking at static link time. However the linker could be taught how to convert these into branches to a tst;moveq;bx stub. This has the advantage of also working for hand written assembly. It may be desirable to get the compiler to also generate this triplet inline for performance reasons.

This is implemented in recent binutils. Code should be compiled with -march=armv4, and arrange for --fix-v4bx-interworking to be passed to the linker and --fix-v4bx to be passed to the assembler. A gcc patch may also need backporting to avoid and earlier sanity check.

If you pass --fix-v4bx to the linker it will generate plain v4 binaries, which are not interworking safe, so should not be used on later cores (which may have Thumb libraries). --fix-v4bx-interworking generates code that works on armv4, and is also interworking safe on later hardware.

The limitation of this option is that any bx instruction will clobber the condition codes. The ABI specifies that condition codes are normally call clobbered, so normally this is not a problem. AFAIK there is only one exception. The libgcc cfcmp* routines (gcc/config/arm/ieee-{sf,df}.S) need to preserve the C and Z flags.

The linker fixup does introduce some additional overhead, so it may be desirable to also implement (2). Care should be taken to avoid double fixups.

Why a new port

In Debian, we want to assure complete binary compatibility. Since the old ABI is not compatible with the new one, we can't allow packages built with old ABI to link against new-abi libs, or the other direction. So the options are:

0. Not an option!

Under no circumstances distribute EABI binaries as .arm.deb depending on current library package names!!!

1. Rename all library packages

This is an ABI transition that affects all architectures, and it has been done before (aout -> elf, c++ ABI)

2. New arch

For the last point, a statically compiled [ArchTakeover] tool could be created. This would also allow i386->amd64 style migrations.

3. ABI: field in control file

This was suggested as part of Multiarch proposal. It is unknown if it would actually become part of Debian or not

From these choices, we believe a new port is the best compromise.

4. conflicting libc packages

In this scenario, we create a libc6-eabi(-dev) package that has eabi glibc and ld-linux.so.3. This package will conflict with libc6(-dev), and thus you can mix and match eabi and non-eabi binaries and libs.

Let's not make perfect an enemy of good!!

Roadmap

Armel (EABI) will be released with etch+1 as it should be in good shape by then. That release will thus contain arm and armel. Arm will be dropped in etch+2, assuming that the above gcc changes to support armv4 CPUs in armel prove practical. If we cannot support armv4 in armel then arm will remain around until we drop v4/StrongARM support, i.e. the port falls into general disuse.

EABI status

The commercial ARM ?RealView C/C++ compiler was the first to support EABI, and usable EABI support came into GCC from version 4.1.0.

CodeSourcery provide GNU ARM toolchains. The 2005Q3 release is a modified version of gcc-3.4.4 while 2006Q1 is from gcc-4.1.0. These toolchains produce EABI object code and the 2006Q1 release also uses the EABI Linux kernel interface.

EABI is supported in the ARM Linux kernel from version 2.6.16 and there is an optional compatibility feature to allow the running of old-ABI binaries with an EABI kernel. The inverse mechanism, to run EABI binaries in an ABI kernel, is not implemented and is said to be non-trivial to do.

Riku Voipio has built a booting EABI root filesystem up to X as proof of concept, which seems stable, built with codesourcery gcc 3.4 toolchain.

Koen Kooi has used ?OpenEmbedded to build a pure EABI root filesystem including native toolchain, visible under http://dominion.kabel.utwente.nl/koen/cms/working-native-eabi-toolchain The compiler is gcc-4.1.1 with glibc-2.4: the exact versions we need. The system boots and runs fine on armv5t and the C compiler seems to work well. However the C++ compiler is not working because libstdc++ is not installed and perl does not execute because libperl.so.5 is not installed as well. Both problems can be solved by using ipkg to install them.

The Angstrom distro of ?OpenEmbedded has a public repository of ARM EABI ipackages compiled for armv5te and visible under http://angstrom-distribution.org/unstable/feed/

Lennert Buytenhek and ADS have built an EABI root filesystem, and some ten thousand (probably many more by now) packages. For download locations and other details, see the announcement

QEMU 0.8.1 can run ARM EABI systems, though when running with the 2.6.16 kernel it is mind-bogglingly slow on x86 processors of a few hundred megahertz. Using 2.6.17-rc3 or later fixes this anomaly. To run a single ARM EABI executable in qemu-user mode, some patches are required, though these are not complete yet.

Minimum versions of components with the first working EABI support are:

Naming

At the Extremadura emdebian meeting, 12-16 April 2006, the name "armel" was chosen. If a bigendian arm EABI port will be created, it will be called "armeb", and it will replace the previous oldabi-based "armeb" port effort.

Strategy

The ultimate aim is a new standalone architecture, composed of three concrete components:


The chronological steps to bootstrap the new arch seem to be:

1) Make Debian packages of a cross-compiler targeting ARM EABI. This means gcc-4.1, glibc-2.4+glibc-ports-2.4, binutils-2.17 and linux-2.6.16. This can be compiled using crosstool-0.42 and the patches and control files at http://freaknet.org/martin/crosstool, and packaged with the scripts at the same location.

2) Make a package for the existing Debian experimental ARM arch of the Linux kernel compiled to use EABI internally, with run-old-ABI-binaries enabled, and test it with existing old-ABI Debian userland.

3) Cross-build essential and required EABI userland packages using dpkg-cross. A parallel effort is the slind project, which is busy improving dpkg-cross support within the Emdebian framework.

4) Make the Debian installer debootstrap work for the new arch.

5) Populate the new-arch repository(-ies) with the rest of the Debian packages.

Is there a "HOWTO Create a New Debian Arch" document? No.

On the crossgcc list, Michael K. Edwards says of the procedure to bootstrap a new Debian arch by doing all building in a QEMU chroot on a fast host:

References


CategoryPermalink