The identity step hashed the caller's source hash into the cache prefix and
restore-keys only matched that prefix, so any revision that touched Cargo.toml,
Cargo.lock, build.rs, .cargo, rust-toolchain or packaging/ started from an
empty cache. Releases bump a version in the manifests, and PRs that change a
dependency do too. Issue #33 records the consequence on the aarch64 leg:
compiling ~260 dependency crates takes 4h32m under emulation, and 93 PR runs
were cancelled at the 300-minute limit with 0 passing.
Keep the exact key as the primary match and add a second restore key that pins
only what decides artifact compatibility: the guest toolchain and packages, the
serialized configuration and the profile. Cargo's fingerprints still reject
stale artifacts, so restoring a previous revision only saves work.
Bump the cache namespace to v3 because the key layout changed. Existing v2
entries are not readable through the new prefixes, so the first run after this
merge starts cold, as for any namespace bump.
Refs #33, #39
* ci(release): cross-compile FreeBSD aarch64 in the x86_64 VM
The FreeBSD release job built aarch64 in an emulated aarch64 guest, which
ran into the 300-minute limit for most apps, so aarch64 was dropped from
releases (#39). #49 fixed the same problem in freebsd.yml by cross-compiling
in the KVM-accelerated x86_64 guest; this ports that setup to release.yml.
The release VM is now always x86_64. For aarch64 it pins the quarterly
package branch, installs lang/rust and runs the same "Set up the aarch64
cross toolchain" step as freebsd.yml; a test keeps the two copies identical.
The build step sets the version before loading the cross environment, so
xtask still builds for the guest. The built-in packager reads binaries from
target/<triple>/release, and the job fails if an archive does not match the
matrix architecture.
aarch64 stays opt-in through the new freebsd-architectures input (default
'["x86_64"]'): each app's package.sh still reads target/release and names
its archive after uname -m, so it must honor CARGO_BUILD_TARGET first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* ci(release): reject unsupported FreeBSD architectures before the VM starts
The guest is always x86_64 and only aarch64 is cross-compiled, so any other
freebsd-architectures value would build x86_64 binaries and the fallback
packager would label them with that value. Check matrix.arch on the host
first and fail for anything but x86_64 and aarch64.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The aarch64 leg booted an emulated aarch64 guest, which compiles about 23
times slower than the KVM-accelerated x86_64 one, so it ran into the
300-minute limit in every caller (#33). GitHub's Arm runners cannot help:
none of them exposes nested virtualization.
Both legs now boot an x86_64 guest. The aarch64 leg pins the quarterly
package branch, installs lang/rust, copies the aarch64 standard library from
the aarch64 lang/rust package of the same version, builds an arm64 sysroot
from base.txz and the GUI library packages, and builds the workspace with
CARGO_BUILD_TARGET=aarch64-unknown-freebsd through base clang and lld. Stable
Rust, no -Zbuild-std.
The cross leg compiles every target, including the tests, but cannot run them
or package.sh, which builds for the guest's own architecture.
The emulated aarch64 VM takes over 5 hours for most apps, so the job is
cancelled at the limit. GitHub's ARM64 runners have no KVM, so they do not
help. Releases ship FreeBSD x86_64 only until #39 is fixed.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Saadi
ZuujisandClaude Opus 5.5
Brandon ThomasandClaude Opus 5.5
skylar taylor-barrick