ii. Toolchain Technical Notes

About Cross-Compilation

Note

The LFS book is not (and does not contain) a general tutorial to build a cross- (or native) toolchain. Don't use the commands in the book for a cross-toolchain for some purpose other than building LFS, unless you really understand what you are doing.

It's known installing GCC pass 2 will break the cross-toolchain. We don't consider it a bug because GCC pass 2 is the last package to be cross-compiled in the book, and we won't “fix” it until we really need to cross-compile some package after GCC pass 2 in the future.

Cross-compilation involves some concepts that deserve a section of their own. Although this section may be omitted on a first reading, coming back to it later will help you gain a fuller understanding of the process.

Let us first define some terms used in this context.

As an example, let us imagine the following scenario (sometimes referred to as “Canadian Cross”). We have a compiler on a slow machine only, let's call it machine A, and the compiler ccA. We also have a fast machine (B), but no compiler for (B), and we want to produce code for a third, slow machine (C). We will build a compiler for machine C in three stages.

Then, all the programs needed by machine C can be compiled using cc2 on the fast machine B. Note that unless B can run programs produced for C, there is no way to test the newly built programs until machine C itself is running. For example, to run a test suite on ccC, we may want to add a fourth stage:

In the example above, only cc1 and cc2 are cross-compilers, that is, they produce code for a machine different from the one they are run on. The other compilers ccA and ccC produce code for the machine they are run on. Such compilers are called native compilers.

Implementation of Cross-Compilation for LFS

Note

All the cross-compiled packages in this book use an autoconf-based building system. The autoconf-based building system accepts system types in the form cpu-vendor-kernel-os, referred to as the system triplet. Since the vendor field is often irrelevant, autoconf lets you omit it.

An astute reader may wonder why a “triplet” refers to a four component name. The kernel field and the os field began as a single “system” field. Such a three-field form is still valid today for some systems, for example, x86_64-unknown-freebsd. But two systems can share the same kernel and still be too different to use the same triplet to describe them. For example, Android running on a mobile phone is completely different from Ubuntu running on an ARM64 server, even though they are both running on the same type of CPU (ARM64) and using the same kernel (Linux).

Without an emulation layer, you cannot run an executable for a server on a mobile phone or vice versa. So the “system” field has been divided into kernel and os fields, to designate these systems unambiguously. In our example, the Android system is designated aarch64-unknown-linux-android, and the Ubuntu system is designated aarch64-unknown-linux-gnu.

The word “triplet” remains embedded in the lexicon. A simple way to determine your system triplet is to run the config.guess script that comes with the source for many packages. Unpack the binutils sources, run the script ./config.guess, and note the output. For example, for a 32-bit Intel processor the output will be i686-pc-linux-gnu. On a 64-bit system it will be x86_64-pc-linux-gnu. On most Linux systems the even simpler gcc -dumpmachine command will give you similar information.

You should also be aware of the name of the platform's dynamic linker, often referred to as the dynamic loader (not to be confused with the standard linker ld that is part of binutils). The dynamic linker provided by package glibc finds and loads the shared libraries needed by a program, prepares the program to run, and then runs it. The name of the dynamic linker for a 32-bit Intel machine is ld-linux.so.2; it's ld-linux-x86-64.so.2 on 64-bit systems. A sure-fire way to determine the name of the dynamic linker is to inspect a random binary from the host system by running: readelf -l <name of binary> | grep interpreter and noting the output. The authoritative reference covering all platforms is in a Glibc wiki page.

There are two key points for a cross-compilation:

In order to cross-compile a package for the LFS temporary system, the name of the system triplet is slightly adjusted by changing the "vendor" field in the LFS_TGT variable so it says "lfs" and LFS_TGT is then specified as “the host” triplet via --host, so the cross-toolchain will be used for generating and processing the machine code running as a part of the LFS temporary system. And, we also need to enable “the cross-compilation mode”: despite “the host” machine code, i.e. the machine code for the LFS temporary system, is able to execute on the current CPU, it may refer to a library not available on the “the build” (the host distro), or some code or data non-exist or defined differently in the library even if it happens to be available. When cross-compiling a package for the LFS temporary system, we cannot rely on the configure script to detect this issue with the dummy program: the dummy only uses a few components in libc that the host distro libc likely provides (unless, maybe the host distro uses a different libc implementation like Musl), so it won't fail like how the really useful programs would likely. Thus we must explicitly specify “the build” triplet to enable “the cross-compilation mode.” The value we use is just the default, i.e. the original system triplet from config.guess output, but “the cross-compilation mode” depends on an explicit specification as we've discussed.

We use the --with-sysroot option when building the cross-linker and cross-compiler, to tell them where to find the needed files for “the host.” This nearly ensures that none of the other programs built in Chapter 6 can link to libraries on “the build.” The word “nearly” is used because libtool, a “compatibility” wrapper of the compiler and the linker for autoconf-based build systems, can try to be too clever and mistakenly pass options allowing the linker to find libraries of “the build.” To prevent this fallout, we need to delete the libtool archive (.la) files and fix up an outdated libtool copy shipped with the Binutils code.

In the preceding table, “on pc” means the commands are run on a machine using the already installed distribution. “On lfs” means the commands are run in a chrooted environment.

This is not yet the end of the story. The C language is not merely a compiler; it also defines a standard library. In this book, the GNU C library, named glibc, is used (there is an alternative, "musl"). This library must be compiled for the LFS machine; that is, using the cross-compiler cc1. But the compiler itself uses an internal library providing complex subroutines for functions not available in the assembler instruction set. This internal library is named libgcc, and it must be linked to the glibc library to be fully functional. Furthermore, the standard library for C++ (libstdc++) must also be linked with glibc. The solution to this chicken and egg problem is first to build a degraded cc1-based libgcc, lacking some functionalities such as threads and exception handling, and then to build glibc using this degraded compiler (glibc itself is not degraded), and also to build libstdc++. This last library will lack some of the functionality of libgcc.

The upshot of the preceding paragraph is that cc1 is unable to build a fully functional libstdc++ with the degraded libgcc, but cc1 is the only compiler available for building the C/C++ libraries during stage 2. As we've discussed, we cannot run cc-lfs on pc (the host distro) because it may require some library, code, or data not available on “the build” (the host distro). So when we build gcc stage 2, we override the library search path to link libstdc++ against the newly rebuilt libgcc instead of the old, degraded build. This makes the rebuilt libstdc++ fully functional.

In Chapter 8 (or “stage 3”), all the packages needed for the LFS system are built. Even if a package has already been installed into the LFS system in a previous chapter, we still rebuild the package. The main reason for rebuilding these packages is to make them stable: if we reinstall an LFS package on a completed LFS system, the reinstalled content of the package should be the same as the content of the same package when first installed in Chapter 8. The temporary packages installed in Chapter 6 or Chapter 7 cannot satisfy this requirement, because some optional features of them are disabled because of either the missing dependencies or the “cross-compilation mode.” Additionally, a minor reason for rebuilding the packages is to run the test suites.

Other Procedural Details

The cross-compiler will be installed in a separate $LFS/tools directory, since it will not be part of the final system.

Binutils is installed first because the configure runs of both gcc and glibc perform various feature tests on the assembler and linker to determine which software features to enable or disable. This is more important than one might realize at first. An incorrectly configured gcc or glibc can result in a subtly broken toolchain, where the impact of such breakage might not show up until near the end of the build of an entire distribution. A test suite failure will usually highlight this error before too much additional work is performed.

Binutils installs its assembler and linker in two locations, $LFS/tools/bin and $LFS/tools/$LFS_TGT/bin. The tools in one location are hard linked to the other. An important facet of the linker is its library search order. Detailed information can be obtained from ld by passing it the --verbose flag. For example, $LFS_TGT-ld --verbose | grep SEARCH will illustrate the current search paths and their order. (Note that this example can be run as shown only while logged in as user lfs. If you come back to this page later, replace $LFS_TGT-ld with ld).

The next package installed is gcc. An example of what can be seen during its run of configure is:

checking what assembler to use... /mnt/lfs/tools/i686-lfs-linux-gnu/bin/as
checking what linker to use... /mnt/lfs/tools/i686-lfs-linux-gnu/bin/ld

This is important for the reasons mentioned above. It also demonstrates that gcc's configure script does not search the PATH directories to find which tools to use. However, during the actual operation of gcc itself, the same search paths are not necessarily used. To find out which standard linker gcc will use, run: $LFS_TGT-gcc -print-prog-name=ld. (Again, remove the $LFS_TGT- prefix if coming back to this later.)

Detailed information can be obtained from gcc by passing it the -v command line option while compiling a program. For example, $LFS_TGT-gcc -v example.c (or without $LFS_TGT- if coming back later) will show detailed information about the preprocessor, compilation, and assembly stages, including gcc's search paths for included headers and their order.

Next up: sanitized Linux API headers. These allow the standard C library (glibc) to interface with features that the Linux kernel will provide.

Next comes glibc. This is the first package that we cross-compile. We use the --host=$LFS_TGT option to make the build system to use those tools prefixed with $LFS_TGT-, and the --build=$(../scripts/config.guess) option to enable “the cross-compilation mode” as we've discussed. The DESTDIR variable is used to force installation into the LFS file system.

As mentioned above, the standard C++ library is compiled next, followed in Chapter 6 by other programs that must be cross-compiled to break circular dependencies at build time. The steps for those packages are similar to the steps for glibc.

At the end of Chapter 6 the native LFS compiler is installed. First binutils-pass2 is built, in the same DESTDIR directory as the other programs, then the second pass of gcc is constructed, omitting some non-critical libraries.

Upon entering the chroot environment in Chapter 7, the temporary installations of programs needed for the proper operation of the toolchain are performed. From this point onwards, the core toolchain is self-contained and self-hosted. In Chapter 8, final versions of all the packages needed for a fully functional system are built, tested, and installed.