The unique-temp-name fix was treating symptoms. The actual defect is
structural: run_parallel_cook documents the invariant
'each recipe builds in its own target/stage/sysroot'
and that invariant is false. dep_levels() keys on the package NAME, so a
recipe's optional-package variants -- llvm-native, llvm-native.dev,
llvm-native.runtime -- are distinct units with no dependency between
them. They therefore land in the same level and cook concurrently while
sharing one target/<triple>/ directory, because build/, sysroot/ and the
stage* dirs are all derived from the recipe directory, not the package
name.
create_dir_clean() on the shared build/ then wipes a sibling's tree
mid-compile:
ninja: error: failed recompaction: No such file or directory
which is the error the earlier parallel builds kept producing and which
unique temp names could never have fixed -- build/ is a real, shared,
destructively re-created directory, not a staging path. It is also why
llvm-native cooks clean in isolation but failed in every parallel build,
and why that failure was misread as a porting gap serious enough to
exclude the package from the config.
Make the recipe DIRECTORY the unit of mutual exclusion. Distinct recipes
still cook fully in parallel; only siblings serialise, and that costs
almost nothing because one cook already populates every stage dir for
the recipe -- the siblings that follow find their work done and return
cached. It also stops LLVM being built three times over.
Residual, not addressed here: recipes joined by [source] same_as (e.g.
llvm-native -> llvm21) share a source tree. That is read-only during
cook, but a concurrent fetch/patch of the shared source would be a
separate race.
vasilito
3f8ebd8631