amol- / consolidatewheels / 37805243776

100%

push

github

Support versioned Linux shared libraries (#5)

* Support versioned Linux shared libraries

Co-authored-by: Jon Shimwell <174589458+jon-proximafusion@users.noreply.github.com>

* Simplify Linux shared library handling per review

* Tidy tests and document the versioning limitation

Drop the parametrize rows that re-ran the demangle table through
buildlibmap, since those names are already covered directly by
test_demangle_libname, and remove the contrived libfoo.so.helper.so.1
case that no real wheel would contain.

Move the note about versioned names being mapped exactly, rather than
satisfying a shorter dependency such as libfoo.so.1, into the
buildlibmap docstring so it is not lost with the pull request
description.

* Add a failing test for the double mangling defect

auditwheel can graft a library that already carries a hash, so a wheel
holds both libfoo-<hash>.so.X.Y.Z and libfoo-<hash>-<hash>.so.X.Y.Z.
Demangling the second yields the first, which is a real bundled file
and the name its consumers already record in DT_NEEDED, so buildlibmap
emits a mapping that repoints working references at a different copy
of the library.

Reproduced against the scipy 1.15.3 manylinux wheel, which ships both
libgfortran-040039e1.so.5.0.0 and libgfortran-040039e1-0352e75f.so.5.0.0.
The test is left failing on purpose to record the defect until the
guard in buildlibmap is written.

* more tests for edge cases

* Do not mangle names that are themselves embedded libraries

auditwheel can graft a library that already carries a hash, so a wheel
holds both libfoo-aaaaaaaa.so and libfoo-aaaaaaaa-bbbbbbbb.so. The
second demangles onto the first, which is a real embedded file and the
name its users already record in DT_NEEDED, so buildlibmap emitted a
mapping that repointed working references at the other copy. On the
scipy 1.15.3 wheel that was 240 of 600 patchelf calls, with no error.

Collect the name of every embedded library across all wheel directories,
since a sing... (continued)

14 of 14 new or added lines in 1 file covered. (100.0%)

330 of 330 relevant lines covered (100.0%)

1.0 hits per line