Arch specific notes — AMD64/EM64T

Position Independent Code issues

Gentoo Policy demands all shared objects to be compiled with -fPIC in CFLAGS. Since this is only a rule, you can break it on some arches. You might never notice it. On AMD64, this is a necessity, if shared objects aren't built with support for position independent code, the build process bails out with an error message like this:

foo.o: relocation R_X86_64_32 can not be used when making a shared
object; recompile with -fPIC

How to fix -fPIC issues

There are several ways to enforce -fPIC on shared objects, each with its pros and cons.

sed'ing the Makefile

Sometimes, a simple sed command is enough to fix it, however, the statements normally aren't very readable and may fail when upstream changes the file. Please verify that you only change the CFLAGS for shared objects, not for the whole package.

Patching Makefile.in/configure

This is more readable, and easier to send upstream.

How not to fix -fPIC issues

Do not patch the Makefile itself, since it is usually generated by the configure script and may vary heavily, so the patch could fail. Also, this doesn't help upstream at all.

Another bad idea is to (ab)use append-flags function from flag-o-matic.eclass. Applying -fPIC on all objects should not be done. It should only be applied to shared objects.

Multilib on AMD64

The current AMD64 processors are able to natively run 32-bit code on a 64-bit kernel. Therefore, you can run programs compiled for x86 in an amd64 environment. However, 32-bit applications need to be linked against 32-bit libraries. Mixing them won't work. For this reason the libraries are sorted, 32-bit libraries normally go to /lib32 respectively /usr/lib32, the 64-bit ones normally to /lib64 or /usr/lib64. In a perfect world, you wouldn't have to read on. Unfortunately, that's not the case, and so it's a bit more complicated.

Multilib toolchain

GCC

To generate 32-bit code, we need a multilib-capable GCC. On other architectures, this functionality is enabled with the USE flag multilib. This is also true for amd64 with the pre-2005.0 profiles. From 2005.0 on, you have to choose whether you want multilib support or not by selecting the profile. Choose 2005.0/no-multilib if you don't want it, all other profiles have the multilib USE flag masked, you're forced to it. With these profiles, GCC will produce x86-code whenever you add -m32 to its command line. Adding -m64 or omitting any bit-width option will default to producing 64-bit code.

glibc

If you've chosen a multilib profile, glibc will be built twice, once 64-bit and once 32-bit. This is because nearly every application links against glibc. To understand how this is done in the ebuild, read The ABI variable.

32-bit compatibility

As you read above, 32-bit applications must be linked against 32-bit libraries. For that, we've made the most common libraries as multilib (via ABI variable and multilib.eclass).

Currently, we provide several profiles, each with its own combination of libdir configurations. Table entries x86, amd64, etc. indicate that the directory contains objects for this ABI; entries with an arrow indicate a symlink to the respective directory.

Profile lib lib32 lib64 libx32
17.0 -> lib64 x86 amd64 non-existent
17.0/no-multilib -> lib64 non-existent amd64 non-existent
17.0/x32 -> libx32 x86 amd64 x32
17.1 x86 non-existent amd64 non-existent
17.1/no-multilib n/a non-existent amd64 non-existent

To always get the right path, you should use $(get_libdir) which is available as a package manager function since EAPI 6. It will always return the correct directory, on all arches. And of course it also takes care of the ABI variable.

The multilib-strict feature

Many Makefiles assume that their libraries should go to /usr/lib, or $(prefix)/lib. This assumption can cause a serious mess if /usr/lib isn't a symlink to /usr/lib64. To find the bad packages, we have a Portage feature called multilib-strict. It will prevent emerge from putting 64-bit libraries into anything other than (/usr)/lib64.

multilib-strict currently doesn't check perl5, gcc, gcc-lib and eclipse-3, this behaviour is controlled by the MULTILIB_STRICT_EXEMPT variable in make.profile.

How to fix ebuilds properly

In most cases, default econf behaviour is sufficient, because it will pass the correct --libdir option to configure.

Some packages provide really bad Makefiles which hard-code /usr/lib. Those should be sed -ed or patched. Don't forget to let upstream know about your modifications!

Headers and multilib

Most C/C++ programs need standard header files like types.h. Some of them depend on architecture specific facts, e.g. types.h on the length of machine words. To ensure that we can compile both 32-bit and 64-bit applications and libraries, we treat /usr/include/asm a bit special.

This is what /usr/include/asm/types.h looks like on an AMD64 box:

/* Common header file autogenerated by create_ml_includes in multilib.eclass */
#ifdef __i386__
#include <asm-i386/types.h>
#endif /* __i386__ */

#ifdef __x86_64__
#include <asm-x86_64/types.h>
#endif /* __x86_64__ */

As you can see, this is just a wrapper that decides which file you need depending on the parameter -D given to gcc. You'll probably run into some troubles if you try to compile something by hand and forget to append -D__x86_64__ to your CFLAGS. Of course, this is not necessary when using Portage. For an explanation, see the The ABI variable section.

The ABI variable

Whenever Portage builds something on amd64, it has to decide whether it should be 32-bit or 64-bit. As stated in Headers and multilib the __i386__ or __x86_64__ respectively, is needed in CDEFINE. Also, gcc has to know what code it should produce, therefore -m32 or -m64 must be appended to CFLAGS. This is done via profile.bashrc. All you need to do if you want to build a package 32-bit is to set ABI=x86.

The details are shown in make.defaults:

MULTILIB_ABIS="x86 amd64"
DEFAULT_ABI="amd64"

CFLAGS_amd64="-m64"
LDFLAGS_amd64="-m elf_x86_64"
CHOST_amd64="x86_64-pc-linux-gnu"
CDEFINE_amd64="__x86_64__"
LIBDIR_amd64="lib64"

CFLAGS_x86="-m32 -L/emul/linux/x86/lib -L/emul/linux/x86/usr/lib"
LDFLAGS_x86="-m elf_i386 -L/emul/linux/x86/lib -L/emul/linux/x86/usr/lib"
CHOST_x86="i686-pc-linux-gnu"
CDEFINE_x86="__i386__"
LIBDIR_x86="lib32"

Porting notes

Machine word sizes

On AMD64, some types differ in size from x86:

Type x86 (ILP32) amd64 (LP64)
char 1 byte 1 byte
short 2 bytes 2 bytes
int 4 bytes 4 bytes
long 4 bytes 8 bytes
long long 8 bytes 8 bytes
pointer 4 bytes 8 bytes
float 4 bytes 4 bytes
double 8 bytes 8 bytes
long double 16 bytes 16 bytes

If you need an exact amount of space, don't use these types but the uXX and sXX ones provided by types.h, where XX is the number of bits you need. Switching to a type that is the same on both arches is not suggested since it's not a clean solution and might cause problems with other arches.

Casts

Many upstream developers assume that the length of a pointer is 4 bytes, which can cause problems when programs do casts from void * to int and vice versa. With GCC 3.4, this causes warnings, the compilation won't abort. If you're lucky, your package works, but it's likely that you encounter segmentation faults or strange behaviour. GCC 4.0 refuses to compile such code.