[gdevenyi]

Summary

libminc's installed CMake package describes its dependencies the way they look inside the minc-toolkit-v2 superbuild. As a result, make install-ing libminc and building a consumer (bicpl, minc-tools, mni_autoreg, …) against it does not work today. The superbuild hides every defect below — it builds its own HDF5, stages it into libminc's own install prefix, and compiles consumers from the source tree.

Each fix corresponds to a build that fails on current develop-1.9.18 and passes with this branch.

No CI changes here — the workflow that found all of this is split out into a separate PR that stacks on this one (see Context).

1. BuildNIFTI: install bundled nifti_clib into CMAKE_INSTALL_LIBDIR

nifti_clib defaults its install library dir to plain lib, but BuildNIFTI.cmake expects the staged libraries under ${CMAKE_INSTALL_LIBDIR} — lib64 on RPM distros. So LIBMINC_USE_NIFTI=ON on Fedora fails:

No rule to make target 'external//usr/local/lib64/libniftiio.a', needed by 'testdir/icv'

One line: pass NIFTI_INSTALL_LIBRARY_DIR=${CMAKE_INSTALL_LIBDIR} to the external project.

2. Installed config: use LIBMINC_INSTALL_LIB_DIR, not hard-coded lib/

The generated config pointed LIBMINC_USE_FILE and LIBMINC_LIBRARY_DIRS at <prefix>/lib/..., while GNUInstallDirs had installed those files to <prefix>/lib64. Consumers on Fedora:

INCLUDE could not find requested file: <prefix>/lib/cmake/UseLIBMINC.cmake

3. Installed config: make the exported package usable standalone

Four related defects, all invisible inside the superbuild:

a. Dependency libraries were exported as bare file names. Consumers link those names with link_directories(${LIBMINC_LIBRARY_DIRS}), which only covers libminc's own prefix. Correct for a superbuild-staged HDF5 — and what keeps such an install relocatable — but a system HDF5 lives elsewhere:

ld: library 'hdf5' not found                              # macOS / Homebrew
undefined reference to symbol 'H5Pset_chunk@@HDF5_SERIAL_1.8.7'
/lib/x86_64-linux-gnu/libhdf5_serial.so.103: error adding symbols: DSO missing from command line   # Debian/Ubuntu

(Debian ships HDF5 as libhdf5_serial under /usr/lib/*/hdf5/serial.)

The rule now applied: a dependency outside the install prefix is exported as an absolute path; one staged inside the prefix keeps its plain name. This leaves the superbuild's relocatable install untouched while making system dependencies resolvable. The mapping handles multi-entry lists (module-mode FindHDF5 reports libhdf5.so;libsz.so;libz.so;libdl.a;libm.so) and entries that are already bare -l names (h5cc/Homebrew), whose directories are added to the exported LIBMINC_LIBRARY_DIRS.

b. Dependency include dirs were not exported. minc.h includes <hdf5.h> and volume_io needs netcdf.h with MINC1 support, but LIBMINC_INCLUDE_DIRS held only <prefix>/include:

/opt/bic-mni/include/minc.h:170:10: fatal error: hdf5.h: No such file or directory

Same in/out-of-prefix rule as (a).

c. internal_volume_io.h was never installed. It is absent from every header list, yet consumers (bicpl, conglomerate) #include <internal_volume_io.h> directly — inside the superbuild it resolves from the source tree:

bicpl/Testing/test_obj_io.c:1:11: fatal error: internal_volume_io.h: No such file or directory

d. UseLIBMINC.cmake injected NOTFOUND into the include path. It did include_directories(${HDF5_INCLUDE_DIR}) unconditionally; when the consumer has not resolved HDF5 itself (it no longer needs to, given (b)), the literal NOTFOUND fails the configure:

CMake Error: The following variables are used in this project, but they are set to NOTFOUND.
HDF5_INCLUDE_DIR (ADVANCED)

Now guarded on both HDF5_INCLUDE_DIR and NETCDF_INCLUDE_DIR.

Review notes

• The load-bearing decision is in (3a): outside the prefix → absolute path, inside → bare name. Deliberately conservative, so the superbuild's relocatable install is unaffected. The alternative — always export absolute paths — is simpler but bakes build-machine paths into a redistributable install. Happy to switch if you prefer.
• An earlier revision of (3a) tried to re-resolve HDF5 on the consumer side instead. A debug print in CI showed HDF5_FOUND=0 there — it dropped hdf5 entirely and was worse than the original bug. Discarded in favour of fixing this at the origin; noting it since it looks like the obvious approach.
• No change to the imported-target path (LIBMINC::minc2); this only repairs the legacy variables existing consumers use.

Validation

These commits are the base of the CI PR (#151), where libminc's own matrix runs green on ubuntu 22.04/24.04, macOS 14/15 and fedora:44 across shared/static × MINC1 × NIFTI: https://github.com/gdevenyi/libminc/actions/runs/29535929762

Every tier-1 consumer also builds and tests against these fixes on ubuntu:24.04 + fedora:44 + macOS, all green:
EBTKS ·
bicpl ·
minc_gco ·
glim_image ·
mrisim ·
postf ·
xdisp ·
minc-tools ·
BEaST ·
mni_autoreg

(Those consumer runs temporarily build libminc from this branch; their own PRs point back at BIC-MNI once this merges.)

Context

Found while adding standalone build+test CI to the BIC-MNI submodule repos, mirroring the superbuild CI in minc-toolkit-v2. This PR is the keystone — 9 of the 10 tier-1 repos cannot build against an installed libminc without it.

• CI: run on all branches, add a Fedora job #151 — the CI workflow, stacked on this PR (merge this one first).
• Related: CI: add standalone build+test workflow EBTKS#9, CI: add GitHub Actions build workflow arguments#2.

🤖 Generated with Claude Code

https://claude.ai/code/session_011nEnygr1q87PepYiJobEGW

[Copilot]

Pull request overview

Modernizes the GitHub Actions CI workflow to run on all branches/PRs and expands coverage with a Fedora container matrix, while also addressing two CI-discovered issues: NIFTI staging on lib64 distros and macOS leak-test max RSS unit differences.

Changes:

• Update .github/workflows/cmake.yml to run on all branch pushes, add concurrency + least-privilege defaults, and add a Fedora 44 container job mirroring the existing option matrix.
• Fix NIFTI external build staging on RPM/lib64 platforms by forcing nifti_clib’s install library directory to match CMAKE_INSTALL_LIBDIR.
• Normalize ru_maxrss units on macOS in minc2-leak-test so leak thresholds behave consistently.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
.github/workflows/cmake.yml	CI modernization (triggers/concurrency/permissions), matrix cleanup, and Fedora container job added.
cmake-modules/BuildNIFTI.cmake	Align nifti_clib install libdir with CMAKE_INSTALL_LIBDIR for lib64 staging correctness.
testdir/minc2-leak-test.c	Normalize macOS RSS units to KB to prevent leak-test false positives.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.