Skip to content

fix(gap): choose the bump source by toolchain family and generation - #40

Merged
HaoZeke merged 11 commits into
mainfrom
fix/gap-bump-source
Nov 16, 2026
Merged

HaoZeke merged 11 commits into
mainfrom
fix/gap-bump-source

Conversation

@HaoZeke

@HaoZeke HaoZeke commented Nov 16, 2026

Copy link
Copy Markdown
Collaborator

stack gap picks the bump source of a no-recipe row among builds of the newest version by toolchain family and generation, not by file name.

  • Between builds of the newest version the file name decided, so an upper-case toolchain name sorted first and an older toolchain before a newer one: QuantumESPRESSO 7.5 was bumped from NVHPC-25.3-CUDA-12.8.0 rather than foss-2025b, and Bandage 0.9.0 from GCCcore-11.3.0 rather than GCCcore-14.3.0.
  • A build on a toolchain family of the generation now comes first, then the newest toolchain generation.

an_equal_version_prefers_the_generation_toolchain_family and an_equal_version_prefers_the_newest_toolchain_generation cover those. This branch and fix/date-versions both add tests at the same spot in src/gap.rs; keep both. Merged with the other thirteen in this batch, the suite is 1940 passed, 0 failed, 10 ignored (1912 on main). The commit has a towncrier fragment. Main has no towncrier yet; #26 (fix/towncrier) adds it.

For a no-recipe row, stack gap names the newest easyconfig of the name
on any toolchain as the bump source. Between builds of the newest
version the file name decided, so an upper-case toolchain name sorted
first and an older toolchain before a newer one: QuantumESPRESSO 7.5 was
bumped from NVHPC-25.3-CUDA-12.8.0 rather than foss-2025b, and Bandage
0.9.0 from GCCcore-11.3.0 rather than GCCcore-14.3.0.

Between builds of one version the bump source is now one on a toolchain
family the generation uses, then the one on the newest toolchain
generation, as data/eessi-generations.toml and the shipped hierarchies
place it, then the newest toolchain of one name. Versions of two
toolchain names are not compared. The recipe route keeps its order.

On the EESSI 2026.06 site list from the deck demo, 120 of 249 no-recipe
rows change bump source, each to a build from a newer toolchain
generation; recipe.tsv and roots.txt are unchanged. Unit tests in gap.rs
cover the family, the generation, and the family match; the first two
fail without the fix.
The eessi target execs eb-local with EB_LOCAL_EESSI_EXTEND=1.
scripts/eessi-extend-eb.sh is removed, and with it
tests/eessi_extend_eb_script.rs, which ran the script, and its CI entry.
eb-stack does not build eb-local; the container bind-mounts the binary
the operator installed.
The Cargo prelude reads the host triple only when rustc -vV succeeds,
but it opened that step with `_triple="";`. The semicolon ended the &&
chain, so when an earlier step of the prelude failed, EasyBuild's
install command still ran. The step is now `_triple= && if ...; fi`,
and the prelude stays one && chain.

The doc comment says where the compat layer comes from: EESSI_EPREFIX,
else EESSI_CVMFS_REPO and EESSI_VERSION, default 2026.06.

A unit test runs the prelude under pipefail with a rustc that exits
127, once with EESSI_EPREFIX and once with only EESSI_CVMFS_REPO.
eb-local 2.4.0 has no extend mode. Asked for --version, it prints its
own version and exits 0, so target doctor passed an EESSI target that
cannot build: outside a Slurm job its builds stop with "not inside a
Slurm job".

When the target environment turns EB_LOCAL_EESSI_EXTEND on, read as
eb-local reads it (set, not empty, not 0), the EasyBuild probe now also
needs the "This is EasyBuild" line that only eb prints. Without it the
check fails, and a new reason field says that eb-local before 2.5.0 has
no extend mode. Other targets are judged by exit status alone, as
before.

reason is left out when there is none, so the JSON of a passing check
is unchanged and a stored report without the field still reads.
The EESSI target needs eb-local 2.5.0 or later, the first release with
EB_LOCAL_EESSI_EXTEND. The target reference and
examples/targets/eessi-extend.toml now say so, and that 2.4.0 and older
ignore the variable and answer --version themselves, which target
doctor fails.
…eessi

When none of the four install variables was set, the removed
scripts/eessi-extend-eb.sh exported EESSI_USER_INSTALL=$HOME/EESSI and
created it. eb-local exports nothing in that case, so EESSI-extend
installs under its own default, $HOME/eessi. The removal entry now says
so, and that setting EESSI_USER_INSTALL in the target environment keeps
the old prefix.
The rebased #31, "fix(cargo): quote CC and LIBRARY_PATH in the EESSI
prelude", on main.
The rebased #34, "feat(target): run the EESSI target through eb-local",
on top of #31.

The Cargo prelude in src/cargo.rs keeps both: #31's CC default, quoted
LINKER, filtered triple and quoted eval, inside #34's chained triple read
(`_triple= && if ...; fi &&`). The inline comment that cited
scripts/eessi-extend-eb.sh goes, as in #34, which deletes that script and
moves the best-effort note to the function's doc comment.
The rebased #40, "fix(gap): choose the bump source by toolchain family and
generation", on top of #31 and #34.
@HaoZeke
HaoZeke force-pushed the fix/gap-bump-source branch from 3576740 to fbca34e Compare November 16, 2026 07:13
@HaoZeke
HaoZeke merged commit 639fde7 into main Nov 16, 2026
22 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant