This is where all the magic begins. We start with our hex0 and kaem seeds and bootstrap our way up to M2-Planet, a subset of C, and mes-m2, an independent port of GNU Mes to M2-Planet. The following steps are taken here:
- hex0 (seed)
- hex0 compiles hex1
- hex0 compiles catm
- hex1 compiles hex2 (v1)
- hex2 (v1) compiles M0
- M0 compiles cc_x86
- cc_x86 compiles M2-Planet (v1)
- M2-Planet (v1) compiles blood-elf (v1)
- M2-Planet (v1) compiles hex2 (final)
- M2-Planet (v1) compiles M1
- M2-Planet (v1) compiles kaem
- M2-Planet (v1) compiles blood-elf (final)
- M2-Planet (v1) compiles get_machine
- M2-Planet (v1) compiles M2-Planet (final)
- M2-Planet (final) compiles mescc-tools-extra (see next step)
This seems very intimidating, but becomes clearer when reading
From here, we can move on from the lowest level stuff.
mescc-tools-extra contains some additional programs, namely filesystem
chown. This allows us to have one unified
directory for our binaries. Futhermore, we also build
checksumming tool, that we use to ensure reproducibility and authenticity
of generated binaries. We also build initial
utilities to deal with compressed archives.
We now move into the
/after directory. As stage0-posix has no
chdir() (not added until very late in stage0-posix),
we have to copy a lot of files into the root of the initramfs, making it
very messy. We get into the move ordered directory
copying over all of the required binaries from
mes is a scheme interpreter. It runs the sister project
which is a C compiler written in scheme, which links against the Mes C
Library. All 3 are included in this same repository. Note that we are
using the experimental
wip-m2 branch to jump over the gap between
mes. There are two stages to this part:
- Compiling an initial mes using
M2-Planet. Note that this is *only* the Mes interpreter, not the libc or anything else.
- We then use this to compile the Mes libc. We need the libc to compile all the programs until we get musl.
tinycc is a minimal C compiler that aims to be small and fast. It
complies with all C89 and most of C99 standards.
First, we compile janneke’s fork of tcc 0.9.26 using
containing 27 patches to make it operate well in the bootstrap
environment and make it compilable using
mescc. This is a
non-trivial process and as seen within tcc.kaem has many different parts
within it: a. tcc 0.9.26 is first compiled using
mescc. b. The mes
libc is recompiled using tcc (
mescc has a non-standard
format), including some additions for later programs. c. tcc 0.9.26 is
recompiled 5(!) times to add new features that are required for other
long long and
float. Each time, the libc is
Note that now we begin to delve into the realm of old GNU software, using older versions compilable by tinycc. Prior to this point, all tools have been adapted significantly for the bootstrap; now, we will be using old tooling instead.
gzip is the most common compression format used for software source
code. It is luckily distributed as a barebones uncompressed
which we extract and then build.
Going forward, we can now use
.tar.gz for source code.
We build GNU Tar 1.12, the last version compilable with mes libc.
You are most likely aware of GNU
sed, a line editor.
patch is a very useful tool at this stage, allowing us to make
significantly more complex edits, including just changes to lines.
Luckily, we are able to patch
make is now built so we have a more robust building system.
make allows us to do things like define rules for files rather than
writing complex kaem scripts.
bzip2 is a compression format that compresses more than
is preferred where we can use it, and makes source code sizes smaller.
Now, we compile upstream tcc 0.9.27, the latest release of tinycc, using the final version of tcc 0.9.26.
From this point onwards, until further notice, all programs are compiled using tinycc 0.9.27.
We patch tinycc to force static linking.
GNU Coreutils is a collection of widely used utilities such as
A few of the utilities cannot be easily compiled with Mes C library, so we skip them.
cp in this stage replaces the
yacc from the Heirloom project. The Heirloom project is
a collection of standard UNIX utilities derived from code by Caldera and
Sun. Differently from the analogous utilities from the GNU project, they
can be compiled with a simple
bash is the most well known shell and the most complex piece of
software so far. However, it comes with a number of great benefits over
kaem, including proper POSIX sh support, globbing, etc.
Bash ships with a bison pre-generated file here which we delete. Unfortunately, we have not bootstrapped bison but fortunately for us, heirloom yacc is able to cope here.
flex is a tool for generating lexers or scanners: programs that
recognize lexical patters.
flex also depends on itself for compiling its own
scanner, so first flex 2.5.11 is compiled, with its scanner definition
manually modified so that it can be processed by lex for the Heirloom
project (the required modifications are mostly syntactical, plus a few
workarounds to avoid some flex advanced features).
mes 0.23 (libc)Edit
In order to build musl we need tcc to support more than 255 command line arguments. Due to the bug in older mes libc this was not possible. Hence, we patch mes libc.
Rebuild tcc 0.9.27 to get the fix from patched mes libc. We have to do
it using older
tcc because tcc-0.9.27 is not self hosting.
musl is a C standard library that is lightweight, fast, simple,
free, and strives to be correct in the sense of standards-conformance
musl is used by some distributions of GNU/Linux as their
C library. Our previous Mes C library was incomplete which prevented us
from building many newer or more complex programs.
tcc has slight problems when building and linking
musl, so we
apply a few patches. In particular, we replace all weak symbols with
strong symbols and will patch
tcc in the next step to ignore
tcc 0.9.27 (musl)Edit
tcc against musl. This is a two stage process. First we
build tcc-0.9.27 using tcc-0.9.26 that itself links to Mes C library but produces
binaries linked to musl. Then we recompile newly produced tcc with
itself. Interestingly, tcc-0.9.27 linked against musl is self hosting.
musl 1.1.24 (tcc-musl)Edit
We now rebuild
musl with the just built
tcc-musl, which fixes a
number of bugs, particularly regarding floats, in the first
tcc 0.9.27 (musl v2)Edit
Now that we have a ‘fixed’
musl, we now recompile
uses floats extensively.
sed is rebuilt against musl.
bzip2 is rebuilt unpatched with the new tcc and musl fixing issues
with reading files from stdin that existed in the previous build.
m4 is the first piece of software we need in the autotools suite,
flex 2.6.4 and bison. It allows macros to be defined and files to be
generated from those macros.
We recompile unpatched GNU
flex using older flex 2.5.11. This is
again a two stage process, first compiling flex using
scan.l) created by old flex, then recompile
scan.c using the new
version of flex to remove any buggy artifacts from the old flex.
bison is a parser generator. With
flex we can now
bootstrap it following https://gitlab.com/giomasce/bison-bootstrap. It’s
a 3 stage process:
1. Build bison using a handwritten grammar parser in C. 2. Use bison from previous stage on a simplified bison grammar file. 3. Build bison using original grammar file.
Finally we have a fully functional
grep is a pattern matching utility. Is is not immediately needed
but will be useful later for autotools.
diffutils is useful for comparing two files. It is not immediately
needed but is required later for autotools.
coreutils is rebuilt against musl. Additional utilities are built
uniq. This fixes a variety of issues with existing
sha256sum from coreutils 6.10 which are
either missing or don't build correctly in 5.0. Other utils are not built at
gawk is the GNU implementation of
awk, yet another pattern
matching and data extraction utility. It is also required for autotools.
Perl is a general purpose programming language that is especially suitable for text processing. It is essential for autotools build system because automake and some other tools are written in Perl.
Perl itself is written in C but ships with some pre-generated files that
need perl for processing, namely
bootstrap Perl we will start with the oldest Perl 5 version which has
the fewest number of pregenerated files. We reimplement two remaining
perl scripts in awk and use our custom makefile instead of Perl’s
pre-generated Configure script.
At this first step we build
miniperl which is
support for loading modules.
We now use
perl from the previous stage to recreate pre-generated
files that are shipped in perl 5.003. But for now we still need to use
handwritten makefile instead of
Yet another version of perl; the last version buildable with 5.003.
More perl! This is the last version buildable with 5.004. It also
introduces the new pregenerated files
Even more perl. 5.6.2 is the last version buildable with 5.005.
GNU Autoconf is a tool for producing
configure scripts for building, installing and
packaging software on computer systems where a Bourne shell is available.
At this stage we still do not have a working autotools system, so we manually install
autoconf script and replace a few placeholder variables with
Autoconf 2.52 is the newest version of
autoconf that does not need
perl, and hence
a bit easier to install.
GNU Automake is a tool for automatically generating Makefile.in files. It is another major
part of GNU Autotools build system and consists of
We bootstrap it using a 3 stage process:
sed to replace a few placeholder variables in
Then we manually install
aclocalscript and its dependencies.
configure.in to create
automake file but skip
Again we manually install
automakescript and its dependencies.
3. We now use
automake to do a proper build and install.
This is an older version of GNU Automake. Various versions of GNU Autotools are not fully
compatible, and we will need older
automake to build some older software.
We now properly rebuild
automake-1.4 and manually installed
autoconf will be necessary to build GNU Binutils.
Yet another old autoconf version that we will need for some parts of GNU Binutils.
GNU Libtool is the final part of GNU Autotools. It is a script used to hide away differences when compiling shared libraries on different platforms.
The GNU Binary Utilities, or binutils, are a set of programming tools for creating and managing binary programs, object files, libraries, profile data, and assembly source code.
In particular we can now use full featured
ar instead of
the GNU linker
ld, which allows us building shared libraries,
and the GNU assembler
musl 1.1.24 (v3)Edit
We rebuild musl for the third time. This time we use GNU ar rather than
so we can drop weak symbols patch. Also, we can use GNU as to build assembly source files,
so those assembly files that tcc failed to compile no longer have to be patched.
tcc 0.9.27 (musl v3)Edit
We rebuild tcc against new musl and without a patch to ignore duplicate symbols.
We now start bootstrapping newer versions of autoconf. Version 2.53 now uses
In order to build it with
autoconf-2.52 we have to patch it a bit and then do
a second unpatched build with newly built
Automake 1.7 and Autoconf 2.54 depend on each other, so we patch out two offending
autoconf macros to make it build with
autoconf. This is the last version of
autoconf that is buildable
automake. This is the latest
automake that is buildable with
autoconf. This time we were able to skip version 2.56.
Again, we managed to skip one version.
We need newer
automake to proceed to newer
autoconf versions. This is the latest
automake version from 1.8 release series.
help2man automatically generates manpages from programs
outputs. This is not strictly required for bootstrapping but will help us to avoid patching
build process to skip generation of manpages. This is the newest version of
does not require Perl 5.8.
Yet another version of
Latest GNU Automake from 1.9 series. Slightly annoyingly depends itself but it is easy to patch
to make it buildable with 1.8.5. Then we rebuild
automake-1.9 with itself.
GNU Find Utilities can be used to search for files. We are mainly interested
xargs that are often used in scripts.
Newer version of libtool which is more compatible with modern Autotools.
GNU Automake from 1.10 series.
aclocal is slightly patched to work
Slightly newer version of GNU Autoconf. At this stage Autoconf is mostly
backwards compatible but newer versions need newer
The GNU Compiler Collection (GCC) is an optimizing compiler produced by the GNU Project. GCC is a key component of the GNU toolchain and the standard compiler for most projects related to GNU and the Linux kernel.
Only the C frontend is built at this stage.
At this stage we are not yet able to regenerate top-level
which needs GNU Autogen and hence Guile. Luckily, building GCC without
Makefile is fairly easy.
GCC can build the latest as of the time of writing musl version.
We also don't need any of the TCC patches that we used before.
Rebuild GCC with GCC and also against the latest musl.
util-linux contains a number of general system administration utilities.
Most pressingly, we need these for being able to mount disks (for non-chroot
mode, but it is built it in chroot mode anyway because it will likely be useful
later). The latest version is not used because of autotools/GCC
loadkeys which is required for building the Linux kernel.
The 2.x series is not used because it requires particular features of autotools
that we do not have available.
GNU Make is updated by .02. The most notable thing is this is now built properly using the build system and GCC, which means that it does not randomly segfault while building the Linux kernel.
kexec is a utility for the Linux kernel that allows the re-execution of the
Linux kernel without a manual restart from within a running system. It is a
kind of soft-restart. It is only built for non-chroot mode, as we only use it
in non-chroot mode. It is used to go into sysb/sysc.
The next step is not a package, but the creation of the sysb rootfs, containing all of the scripts for sysb (which merely move to sysc). Again, this is only done in non-chroot mode, because sysb does not exist in chroot mode.
Linux kernel 4.9.10Edit
A lot going on here. This is the first (and currently only) time the Linux kernel
is built. Firstly, Linux kernel version 4.9.x is used because newer versions
require much more stringent requirements on the make, GCC, binutils versions.
However, the docs are also wrong, as the latest of the 4.9.x series does not
work with our version of binutils. However, a much earlier 4.9.10 does
(selected arbitarily, could go newer but did not test), with a small amount
of patching. This is also modern enough for most hardware and to cause few
problems with software built in sysc. Secondly, the linux-libre scripts are used
to deblob the kernel. Unauditable, unbootstrappable binary blobs within our
kernel are unacceptable. Our gawk is too buggy/old so we use sed instead for
this operation. Every other pregenerated file is appended with
we use a
find command to remove those, which are automatically regenerated.
The kernel config was originally taken from Void Linux, and was then modified
for the requirements of live-bootstrap, including compiler features, drivers,
and removing modules. Speaking of which, modules cannot be used. These cannot
be transferred to subsequent systems, and we do not have
the initramfs of sysb is generated in this stage, using
the Linux kernel tree. This avoids the compilation of
cpio as well.
This is the last step of sysa, run for non-chroot mode. It uses kexec to load the new Linux kernel into RAM and execute it, moving into sysb.
sysb is purely a transition to sysc, allowing binaries from sysa to get onto a disk (as sysa does not nessecarily have hard disk support in the kernel). It populates device nodes, mounts sysc, copies over data, and executes sysc.
Up to this point, our build of
bash could run scripts but could not be used
interactively. This new version of
bash compiles without any patches,
provides new features, and is built with GNU readline support so it can be used
as an interactive shell.
autoconf-2.61 is used to regenerate the configure
bison is used to recreate some included generated files.
XZ Utils is a set of free software command-line lossless data compressors, including lzma and xz. In most cases, xz achieves higher compression rates than alternatives like gzip and bzip2.
GNU Automake from 1.11 series. This is not the latest point release as newer ones need Autoconf 2.68.
This is a much newer version of GNU Autoconf.
GNU Automake from 1.15 series. This is the last version that runs on Perl 5.6.
Newer tar has better support for decompressing .tar.bz2 and .tar.xz archives. It also deals better with modern tar archives with extra metadata.
We build the latest available coreutils 8.32 which adds needed options to make
results of build metadata reproducible. For example, timestamps are changed with
pkg-config is a helper tool that helps to insert compile and link time flags.
A newer version of make built using autotools is much more reliable and is compiled using a modern C compiler and C library. This removes a couple of segfaults encountered later in the process and allows more modern make features to be used. We do not go for the latest because of the use of automake 1.16 which we do not have yet.
GNU Multiple Precision Arithmetic Library (GMP) is a free library for arbitrary-precision arithmetic, operating on signed integers, rational numbers, and floating-point numbers.
GMP is required by newer versions of GCC and Guile.
The GNU Autoconf Archive is a collection of Autoconf macros that are used by various projects and in particular GNU MPFR.
The GNU Multiple Precision Floating-Point Reliable Library (GNU MPFR) is a library for arbitrary-precision binary floating-point computation with correct rounding, based on GNU Multi-Precision Library.
GNU MPC is a library for multiprecision complex arithmetic with exact rounding based on GNU MPFR.
An older version of flex is required for bison 2.3. We cannot use 2.5.11 that was compiled much earlier, as it does not produce reproducible output when building bison 2.3.
This is an older version of bison required for the bison files in perl 5.10.1. We backwards-bootstrap this from 3.4.1, using 3.4.1 to compile the bison files in 2.3. This parser works sufficiently well for perl 5.10.5.
Bison 3.4.1 is buggy and segfaults when perl 5.32.1 is built. This is probably because it was built with a hand-written makefile. We do not build the latest bison because perl 5.32.1 requires bison <= 3.4.2.
Perl 5.10.1 is an intermediate version used before Perl 5.32. We require this version as it adds a couple of modules into lib/ required to regenerate files in Perl 5.32. We still use the Makefile instead of the metaconfig strategy, as metaconfig history becomes poor more than a few years back.
dist is perl's package used for generating Perl's Configure (which is written in Perl itself). We 'compile' (aka generate) metaconfig and manifake only from dist. We do not use dist's build system because it itself uses dist.
We finally compile a full version of Perl using Configure. This includes all base extensions required and is the latest version of Perl. We are now basically able to run any Perl application we want.
GNU Automake from 1.16 series that required newer Perl.
GNU Autoconf 2.71 is even newer version of autoconf. It does not build with miniperl, so we postponed it until full perl was built.
Our old patch was built with manual makefile and used mes libc. This is a newer version which we need in order to import gnulib into gettext.
GNU Gettext is an internationalization and localization system used for writing multilingual programs.
Texinfo is a typesetting syntax used for generating documentation. We can now use
makeinfo script to convert
.texi files into
.info documentation format.
zlib is a software library used for data compression and implements an abstraction of
DEFLATE algorithm that is also used in
GCC 4.7.4 is the last version written in C. This time we build both C and C++ backends.
C++ backend has some dependency on
gperf which is written in C++. Fortunately, it is
easy to patch it out and resulting
g++ compiler is capable of building
gperf is a perfect hash function generator (hash function is injective).
Library for manipulating Unicode and C strings according to Unicode standard. This is a dependency of GNU Guile.
The libffi library provides a portable, high level programming interface to various calling conventions.
libatomic_ops provides semi-portable access to hardware-provided atomic memory
update operations on a number of architectures.
The Boehm-Demers-Weiser conservative garbage collector can be used as a garbage collecting replacement for C malloc or C++ new.
GNU Ubiquitous Intelligent Language for Extensions (GNU Guile) is the preferred extension language system for the GNU Project and features an implementation of the programming language Scheme.
guile-psyntax-bootstrapping project to bootstrap Guile's
without relying on pre-expanded code.