ba34cbad41
moved the location of the handling for _lldb.so in the python bindings,
which we delete in gnu-install-dirs.patch. to allow the patch to apply
to lldb 23, we make the patch version dependent and move the deletion to
the new location in the source for the llvm 23 version.
Building both egg sets on aarch64-darwin for the first time turns up four
CHICKEN 6 failures, three of them cases where a Darwin bundle has to resolve
every symbol at link time while a Linux shared object may leave them to the
loader.
bvconv binds to iconv, which is part of glibc but a library of its own on
Darwin.
constructive-solid-geometry names every OpenCASCADE library it uses except
TKBO, which defines the BRepAlgoAPI_* boolean operations it calls. On Linux
that library arrives anyway, as a transitive dependency of TKBool, so the egg
both links and loads; the link option is added unconditionally, as the
dependency is real on either platform.
nutils calls mempcpy, a GNU extension that Darwin's libc does not provide, so
it is marked broken there, as is tkgui, which depends on it.
The three CHICKEN 5 eggs that fail on aarch64-darwin -- fusion-arrays,
hypergiant and libyaml -- fail identically at the base of this branch and are
left alone.
Also records why neither release runs its test suite on Darwin, which was
until now an unexplained condition inherited from the CHICKEN 5 expression:
CHICKEN 5's runtests.sh drives the compiler through /usr/bin/env, which the
sandbox denies, and CHICKEN 6's csc tests run binaries whose install name
install_name_tool has already rewritten to a $out that does not exist until
the install phase.
Assisted-by: Claude Code (claude-opus-5[1m])
Regenerated with update.sh. Upstream has ported 61 more eggs to CHICKEN 6
since the set was first written, taking it from 216 to 277, and released new
versions of 19 of the ones already in it. Nothing was dropped.
srfi-4 rejoins the dependencies that are stripped because they are part of
the compiler: it is core in CHICKEN 6, like r7rs, and npdiff is the first
egg in the set to declare a dependency on it.
Two of the three eggs marked broken now build. transducers is unchanged at
0.10.0, but srfi-253, the only dependency of it to have moved, went from
0.2.0 to 0.3.1, and its module now resolves; generalized-arrays follows, as
it only failed because it depends on transducers.
feathers stays broken, but for a corrected reason: its tarball does have
sources. The egg is implemented in Tcl and its .egg declares no source for
the program it builds, so chicken-install falls back to a feathers.scm that
is not there, which only fails because the egg is built and installed by two
separate chicken-install runs. Diagnosed by @DerGuteMoritz in review, who
reports that upstream intends to deprecate the egg, so it is left alone.
Of the newly ported eggs, only ezxdisp needs an override, for the X11
headers it binds to. Unlike the CHICKEN 5 egg, this version builds without
any extra compiler flags.
276 of the 277 eggs build.
Assisted-by: Claude Code (claude-opus-5[1m])
CHICKEN_REPOSITORY_PATH is a search path of egg repository directories, which
hold compiled modules and .egg-info files, not executables, so prefixing PATH
with it never made a program findable. It also bakes the build time
repository search path, i.e. the egg's own build inputs, into the wrapper of
every installed program.
Reported by @DerGuteMoritz in review, who verified it with the csm egg; csm
still runs with the prefix removed, on both CHICKEN 5 and 6.
Assisted-by: Claude Code (claude-opus-5[1m])
The CHICKEN 6 set was added as a copy of the CHICKEN 5 one, following the
existing 4-vs-5 split, but the two releases only really differ in the
compiler's version, its egg binary version, and the egg channel their eggs
come from. Everything else -- the compiler derivation, the egg builder, the
scope that ties them together, the egg metadata reader and the script that
generates the egg set -- now lives in common/, and each release directory
keeps only what is genuinely per release: deps.toml, overrides.nix, and the
arguments in default.nix.
CHICKEN 4 is left alone. It predates deps.toml and update.sh, it patches and
regenerates the compiler's own sources, and it is no longer supported
upstream, so it has little to share.
Aligning the two expressions changes the CHICKEN 5 compiler derivation:
- The bootstrap stage is gone. Building a throwaway compiler and rebuilding
with it is a leftover of the CHICKEN 4 expression, which patches Scheme
sources and therefore needs a working CHICKEN to regenerate the C files.
CHICKEN 5 needs no patch and its tarball ships the generated C files.
- checkTarget is set, so the test suite now actually runs. The generic
check phase probes for a target with a bare `make`, which fails here
because PLATFORM is only given in makeFlags, so the suite was silently
skipped. It passes: 76 sections, in 12 minutes, with the two tests the
expression disables in postPatch still disabled.
- dontConfigure is set. CHICKEN 5 has no configure script at all, so this
is a no-op for it; it is CHICKEN 6, whose hand-written configure rejects
the options stdenv passes, that needs it.
- The setup hook is generated from the binary version with replaceVars,
rather than kept as one copy per release.
The egg builder gains the CHICKEN 6 egg cache export, which is harmless for
CHICKEN 5, and loses an unused binding. As @DerGuteMoritz notes in review,
the export only needs doing once: the install phase runs in the same shell as
the build phase.
Every CHICKEN 5 egg therefore rebuilds, with no expected change to what they
contain.
Sharing the expression also shares its meta, so Xophmeister, who added the
CHICKEN 6 set, is now listed as a maintainer of CHICKEN 5 as well.
Assisted-by: Claude Code (claude-opus-5[1m])
Derived from building the whole set: 213 of the 216 eggs now build, and the
three that do not are marked broken with a reason.
Only 10 of the 96 entries in the CHICKEN 5 overrides file are needed, as most
of the eggs they refer to have not been ported to CHICKEN 6 yet, and one of
them needs a different fix: svn-client's build script no longer hardcodes
/usr/include/subversion-1, so it just needs pkg-config and the libraries,
rather than the substitution the CHICKEN 5 override performs.
The helpers are carried over unchanged, minus the ones nothing uses yet, so
the two files stay recognisably the same.
Assisted-by: Claude Code (claude-opus-5[1m])
Generated from upstream's eggs-6-latest tree with update.sh, the same way
as the CHICKEN 5 set; 216 eggs have been ported to CHICKEN 6 so far,
against 621 for CHICKEN 5.
r7rs is no longer an egg in CHICKEN 6, so it joins the dependencies that
are stripped because they are part of the compiler.
The license translation table gains "bsd-2", which four eggs in the
CHICKEN 6 channel use as a spelling of BSD-2-Clause. Only the abbrev egg
is left with a license Nixpkgs has no attribute for, the Open Works
License, which stays a plain string.
Assisted-by: Claude Code (claude-opus-5[1m])
Same builder as the CHICKEN 5 one: chicken-install can build and install
an egg offline from an unpacked tarball, and the workaround for
https://bugs.call-cc.org/ticket/1855 is still needed, as 6.0.0 installs
eggs whose .egg carries no version property without one.
Unlike CHICKEN 5, chicken-install now takes a lock in its egg cache even
when building an egg from the current directory, so the cache has to be
pointed at a writable location for the build to work in the sandbox.
overrides.nix is empty for now; entries follow once the whole egg set has
been built.
Assisted-by: Claude Code (claude-opus-5[1m])
CHICKEN 6.0.0 is the first major release since 5.x. It bumps the egg
binary version from 11 to 12, so it gets its own package set rather than
replacing chickenPackages_5; the unversioned attributes still point at
CHICKEN 5.
Compared to the 5 expression:
- 6.0.0 needs no patches, so the two-stage bootstrap (building a
throwaway compiler just to regenerate the C sources) is gone and the
setup hook is unconditional.
- The tarball now ships a configure script, but it is hand-written rather
than generated by autoconf: it only writes make variables into
config.make and rejects options it does not know, including the ones
stdenv would pass. Since command line variables take precedence over
config.make, keep passing them to make and skip configure.
- checkTarget is set explicitly. The generic check phase probes for a
target with a bare `make check`, which fails because PLATFORM is only
given in makeFlags, so the test suite would silently not run. With it
set, the whole suite passes, so the test disables carried by the 5
expression are not needed here.
Assisted-by: Claude Code (claude-opus-5[1m])
This pattern is error prone, especially with __structuredAttrs, where
nested lists can leave important attributes such as `patches` null.
The treewide rewrite was produced with ast-grep, followed by `nix fmt`
and manual fixes for two `sema-undefined-variable` violations.
Future occurrences will be rejected by nixpkgs-vet once this change is
merged:
https://github.com/NixOS/nixpkgs-vet/pull/301
Crystal runs CPU detection to figure out worker count if not explicitly
set. This detection code has a bug and errors on machines with >= 128
CPUs, since it uses an 8 bit integer for counting CPUs. Large CPU counts
do work if the CPU detection code is skipped. This sets CRYSTAL_WORKERS
to NIX_BUILD_CORES to avoid the CPU detection code.
Assisted-by: Codex (GPT-5.6 Sol)
The patch failed to apply on 18 and 19. We have to version it.
LLVM 18 has no MLIR_SRC_SHARDER_TABLEGEN_EXE. LLVM 19 and 20 fail to
apply the 21+ patch due to mechanical hunk failures, but otherwise use
an identical patch to 21+.
