Le mardi 10 novembre 2026, 20:59:26 heure normale d’Europe centrale Edmund 
Lodewijks a écrit :
> Hi all,
> 
> 
> Apologies for the lengthy email. This is based on a good two weeks or so 
> of work on making Zig packageable for Debian.
> 
> 
> Short version: Zig's bootstrap blob can now be produced from source, and 
> the result is bit-for-bit identical to the blob that upstream ships in 
> the signed 0.17.1 release. I have a working `zig` source package built 
> on that basis. Before I take it further I would like to check two things 
> with the list, and find a sponsor.
> 
> This follows on the January 2025 "getting ziggy with it" thread 
> <https://lists.debian.org/debian-devel/2025/01/msg00098.html>, which 
> ended without a bootstrap plan.
> 
> 
> ## The problem in short
> 
> Zig is written in Zig, and therefore needs Zig to compile itself. 
> Upstream breaks the build circularity by committing `stage1/zig1.wasm` 
> to the source tree: `wasm2c` converts it to C, the system C compiler 
> builds that, and the result builds the compiler. That artifact is a 
> build product, not source in the sense of DFSG §2, so Debian cannot ship it.
> 
> A first attempt: Nick Black's proposal in msg00153 (i.e.: bootstrap with 
> the blob, then regenerate it) proves the blob is honest but does not 
> remove it from what the source package needs. The first upload is the 
> part that was unsolved, and which I am solving here.
> 
> ## What is new
> 
> GNU Guix solved the hard part (issue 74217, merged as 
> 40b9f33066..42e32d52c3): A chain starting from the C++ compiler upstream 
> used before the WebAssembly seed existed, walking forward through Zig's 
> history and regenerating the seed at each step. At the time of writing 
> this email, it reaches Zig 0.16.0.
> 
> I extended that chain from 0.16.0 to 0.17.1 in nine further builds 
> across six seed-regeneration boundaries. The seed regenerated at the 
> last of those is identical to upstream's:
> 
>    $ minisign -Vm zig-0.17.1.tar.xz \
>        -P RWSGOq2NVecA2UPNdBUZykf1CCb147pkmdtYxgb3Ti+JO/wCYvhbAb/U
>    Signature and comment signature verified
> 
>    $ tar xf zig-0.17.1.tar.xz zig-0.17.1/stage1/zig1.wasm -O | sha256sum
>    77c02ad2496aab66fb408f831a49ee8f7ed6746326388c972735707881cac770
> 
>    regenerated from source, never executing upstream's copy:
>    77c02ad2496aab66fb408f831a49ee8f7ed6746326388c972735707881cac770
>    3,246,758 bytes
> 
> This is upstream's minisign signature rather than a git clone, so the 
> claim does not rest on my checkout being authentic. The same seed ships 
> in 0.17.0, since the last regeneration commit precedes both releases.
> 
> For those who are interested: The chain, the hops, the hashes, the 
> driver script, the logs... they are all documented and uploaded to Salsa:
> 
>    https://salsa.debian.org/ejjl/zig-bootstrap

I am interest to have a full documentation for your system (on wiki.d.o)
I have a similar problem with gwt a java toolkit needed for building js

Particularly how to use guix to solve bootstrap problem

rouca
> 
> ## One Zig package for the archive
> 
> One source package:
> 
>    https://salsa.debian.org/ejjl/zig
> 
>    * `Files-Excluded: stage1/zig1.wasm` in debian/copyright, producing 
> zig_0.17.1+dfsg.orig.tar.xz.
>    * `Build-Depends: zig`: each zig built by the previous zig in the 
> archive, as gcc and rustc are.
>    * `debian/README.source` documents the bootstrap and how to reproduce 
> the seed.
>    * The bootstrap chain itself is *not* uploaded. It lives on Salsa.
> 
> What the +dfsg repack removes is stage1/zig1.wasm, which cannot be 
> produced without an existing zig.
> 
> I am deliberately keeping the four lib/libc/*/abilists symbol tables. 
> These can be regenerated at any time from glibc's and the BSDs' own 
> LGPL/BSD sources by an MIT tool. That tool is itself written in Zig; 
> however, regenerating them needs a zig, but they are not needed to 
> *build* zig, only to cross-compile to glibc and BSD targets. A zig built 
> without them works, and can then regenerate them.
> 
> So the claim I want to make is not that the archive contains no 
> binaries. It is that it contains no binary whose source cannot be 
> rebuilt, and nothing circular at all.
> 
> The package builds in a clean unstable chroot, and lintian reports no 
> errors or warnings beyond the expected unreleased-changes. What is left 
> at info and pedantic level is in Open Issues below.
> 
> There is, of course, the question of how to solve the packaging problem 
> of the circular dependency on Zig for building Zig in Debian. I see two 
> ways around it for the package as it stands, plus a third that may 
> already have been proposed by Ivan (see Coordination below). I would be 
> happy to hear your thoughts on which should be preferred.
> 
> (a) Ship the seed in the source package, upload source-only. This is 
> what the rust package does.
>      Based on what rust does, I would suggest we copy two out of their 
> three options:
> 
>       - no stage0/ present?  --> build with the distro's zig.
>       - stage0/ present?     --> bootstrap from that shipped seed.
> 
>      Following Rust, the detection would be by file presence rather than 
> by a build profile, so one debian/rules serves both cases and a normal 
> build cannot accidentally take the bootstrap path. Rust has a third 
> path, the pkg.rustc.dlstage0 build profile, which downloads a stage0 
> during the build. That one won't work here: Debian buildds have no 
> network access during a build.
> 
>      For zig, that stage0 file would be 
> zig_0.17.1+dfsg.orig-stage0.tar.xz, carrying bin/zig and lib/zig from 
> the compiler my chain produced. dpkg-source unpacks a supplementary orig 
> tarball named orig-stage0 into stage0/ in the build tree, and 
> debian/rules would use that compiler instead of the system one. For that 
> upload only, Build-Depends: zig is disabled ([!amd64]). The seed is used 
> only to compile; it is never installed into a binary package. The next 
> upload leaves the tarball out and restores Build-Depends: zig (this 
> needs to be done manually, of course).
> 
>      Worth imitating from Rust as well: rust's d/rules refuses to build 
> if a stage0 is present while rustc is still listed as a Build-Depends 
> for the architecture being built (or vice-versa). A bootstrap upload has 
> two halves (ship the seed, *and* drop the build dependency), and we 
> might as well have that protection in place.
> 
>      Rather than only describe this, I have implemented it on a branch, 
> so that it can be reviewed and tried:
> 
>        https://salsa.debian.org/ejjl/zig/-/tree/debian/stage0
> 
>      It differs from debian/latest in three files, and includes a script 
> that builds the seed tarball reproducibly and refuses to package a seed 
> carrying a WebAssembly artifact, a /gnu/store ELF interpreter or an 
> embedded build path. The source package comes to about 50 MB rather than 
> 24, for one upload cycle.
> 
> (b) A maintainer binary upload. I (or the sponsor) build(s) the binaries 
> locally against the bootstrap compiler, and a sponsor uploads those 
> rather than just signing a .dsc.
> 
> Between those options, I lean towards option (a). Against option (b), 
> rust's README.source puts it quite nicely into words:
> 
>    "[The traditional Debian way is that] a developer locally installs 
> some upstream
>    bootstrapping binaries [and] builds a Debian package, using these 
> binaries as
>    undeclared build dependencies. [...] The problem with this is, Debian 
> does not
>    have any policy nor infrastructure that can try to reproduce what this
>    developer supposedly did."
> 
> The objection to (a) is that for one upload cycle the source package 
> would contain a compiled binary; rust's README concedes the mechanism 
> "stretches the definition of source package". My answer is below, under 
> question 1.
> 
> After all this text: Two questions
> 
> 1. Is a Guix-produced, hash-recorded seed an acceptable one-time 
> bootstrap for the first upload?
> 
>     I would argue it is stronger than the rustc precedent, which seeded 
> from an upstream prebuilt binary with no verification at all. Guix's 
> seed is free software built from source, its builder is in Debian main, 
> and the seed it produces is now shown to match upstream's bit for bit.
> 
>     Just for reference: Debian already accommodates this class of 
> problem, in several different ways. ghc and rustc self-host from the 
> archive, with rustc's pkg.rustc.dlstage0 profile falling back to 
> downloading upstream's prebuilt stage0 unverified. sbcl bootstraps from 
> clisp, an independent implementation. gcc self-hosts and its original 
> seed predates DFSG. And ocaml 5.4.1-1 ships boot/ocamlc, a bytecode 
> image, in its orig tarball in main, with no ocaml in Build-Depends. 
> debian/rules executes it on every build.
> 
>     I think that the solution here proposed for Zig is a good attempt at 
> being DFSG compliant for such a situation: A seed derived entirely from 
> source, and that seed shown to be identical to what upstream ships.
> 
> 2. Is an amd64-only first upload acceptable?
> 
>     Guix's chain is x86_64-linux only. Efraim Flashner reported in 74217 
> that he could not get the early steps building on riscv64 or ppc64el. 
> Other architectures need either that work done upstream or a 
> cross-bootstrap. I just wanted to mention this upfront. I would be happy 
> to figure out if and how more architectures can be added. For a first 
> upload, I hope this will be enough, so that we can get this into Forky 
> still, together with other Zig-software.
> 
> ## Current Status and Open Issues
> 
> 1. The seed I used for hop 00 came from `guix build [email protected]` **with** 
> substitutes. A `--no-substitutes` run is in progress as I write this; it 
> rebuilds the entire toolchain from Guix's stage0-posix seed upward, 
> including seven LLVM builds, so it will probably take days. I will 
> report the result when they come in.
> 
> 2. debian/copyright is not finished (barely started). The bundled 
> mingw-w64, wasi-libc, BSD libc and aro trees still need auditing. I 
> would appreciate guidance from anyone who has done this at scale (Rust, 
> Go, npm?). There are ~30,000 files in the tarball, and from what I can 
> see, there will be a variety of licenses. Perhaps there are tools 
> available to extract this information? Thank you in advance!
> 
> 3. Should the abilists tables (above) stay? Guix removes them; I have 
> kept them, because removing them costs cross-compilation for no gain I 
> can see, and the tables shipped in 0.17.1 are current -- glibc 2.0 
> through 2.44, so both unstable's 2.43 and 2.44 are covered. Advice welcome.
> 
> 4. One file needs a second opinion. 
> lib/libc/include/generic-freebsd/netlink/netlink.h carries 
> SPDX-License-Identifier: BSD-2-Clause with FreeBSD copyright, and then a 
> separate "Copyright (C) The Internet Society (2003)" notice for the 
> material quoted from RFC 3549, including the no-derivatives clause that 
> lintian's license-problem-non-free-RFC notices. I read that as a 
> dual-licensed file and overrode the tag, which is what lintian's own 
> description of the tag suggests for multi-licensed files. It is the only 
> file in the tree carrying RFC terms. If that is not the correct 
> approach, I can also remove that.
> 
> ## Coordination
> 
> There are overlapping ITPs and I do not want to cause any trouble here 
> (which is why I haven't filed an ITP myself yet):
> 
>    #995670   ITP: zig (Nick Black), stalled
>    #1092895  ITP: zig 0.13.0 (Nick Black, Jan 2025), stalled
>    #1150314  ITP: zig-wasm-bootstrap (Ivan Carvalho, Oct 2026): 
> described as "a port of Guix's Zig bootstrap to Debian", which is close 
> to what I have done
>    Jonathan Carter created debian/zig on Salsa in Sep 2024, which didn't 
> see much action
> 
> What would be the best way forward? Jonathan, Nick, Ivan: Happy to work 
> together with you, and anyone else, on getting Zig into Debian! This is 
> my attempt of trying to get Zig and other Zig programmes into Debian, in 
> time for Forky, if possible.
> 
> (c) Besides options (a) and (b), there is the option of adding a second 
> source package to the archive: zig-bootstrap. That would run the chain I 
> ran in the VM, producing a from-source seed that can compile Zig, with 
> zig build-depending on it. That is the cleanest outcome of the three, 
> since no compiled binary would enter the archive at any point. My chain 
> takes around 1.5 to 2 hours on 8 cores, which may or may not be out of 
> reach for a buildd. I think this may also be what Ivan's ITP is about, 
> though arrived at in a slightly different way. Whatever works for Debian!
> 
> ## What I am asking for
> 
> Views on the two questions above, and on (a) versus (b) versus (c). If 
> (a) is acceptable, then this is an ordinary source-only sponsorship 
> request rather than anything unusual, which I hope makes it easier to 
> help with.
> 
> My own interest is, besides Zig, also ghostty, which needs zig >= 
> 0.17.0; I have built it end-to-end with the blob-free compiler on my 
> builder VM, and it builds and works. But this one is about Zig, which 
> would unblock other Zig packages.
> 
> Finally, if I should file an ITP for this, or not?
> 
> ## A final note
> 
> I have also written to the Guix people who work on Zig there, and have 
> had favourable replies, though nobody has yet walked through the steps I 
> took. I will report back here either way.
> 
> Thank you very much for reading this, for your time and advice, and for 
> any assistance.
> 
> Kind regards,
> 
> Edmund Lodewijks
> 
> 
> P.S.: Again, for good measure, the links:
> 
>    https://salsa.debian.org/ejjl/zig/                         the package
>    https://salsa.debian.org/ejjl/zig/-/tree/debian/stage0     option 
> (a), implemented
>    https://salsa.debian.org/ejjl/zig-bootstrap                the 
> bootstrap chain
> 
> 
> -- 
> Edmund Lodewijks <[email protected]>
> TZ: UTC+2 / GMT+2
> 
> 

Attachment: signature.asc
Description: This is a digitally signed message part.

Reply via email to