ci: build OCCT shared on Linux so plug-ins share one OCCT instance

`geom.tree().select()` silently returns zero results (or raises SWIG's
"An unknown error occurred") in the Linux release packages, while the same
commit built from source returns correct answers. `select_box()` agrees
between both, and geometry conversion is bit-identical -- only operations
that touch a stored TopoDS_Shape diverge.

Cause is linkage, not code. 8bdaa8c7c narrowed the Rocky builds from
`--shared` to `--ifcopenshell-shared`, which shares IfcOpenShell's own
libraries but leaves every dependency static. OCCT is then compiled
privately into each plug-in that uses it -- 15 of them, verified by their
own copies of the BRepClass3d/BRepExtrema/Standard_Failure strings.

That contradicts what the binaries already declare. tree.h:1748 casts a
`conversion_result_shape*` to `open_cascade_shape*`, moves the
TopoDS_Shape out of it and frees it; `open_cascade_shape` is defined once
in ifcopenshell_geometry_kernel_opencascade.so and left undefined in
ifcopenshell_geometry_tree_opencascade_brep.so for the loader to resolve.
So the two plug-ins are designed to share one OCCT-based type system, yet
static linking gives each its own Standard_Type registry and allocator.
Shapes get read and released by a different OCCT instance than made them.

Add `--opencascade-shared`, mirroring the existing `--ifcopenshell-shared`
precedent, and use it on both Rocky workflows. It cannot be spelled
`--occt-shared`: build-all.py parses any `occt-*` flag as a version
override.

BUILD_STATIC drives three things at once -- dependency link type,
-fvisibility=hidden, and BUILD_SHARED_LIBS -- so making one dependency
shared means overriding all three for it. Visibility is the subtle one:
OCCT's Standard_EXPORT expands to nothing on Unix, so it relies on default
visibility to export its API. Built shared under -fvisibility=hidden it
exports almost nothing and its own libraries cannot resolve against each
other (libTKMath.so fails to find
NCollection_BaseAllocator::CommonBaseAllocator in libTKernel.so). Static
archives are immune, which is why this surfaces only once OCCT goes
shared. Compile OCCT with the pre-visibility flag set instead.

Link the OCCT set with --as-needed. FindOpenCASCADE.cmake's config branch
uses OCCT's *complete* module list, Visualization included, which against a
static OCCT costs nothing -- an unreferenced module contributes no objects.
Against a shared OCCT all 47 become hard DT_NEEDED entries, and libTKV3d
pulls libGL.so.1 + libEGL.so.1, so `import ifcopenshell` fails on any
headless machine with "libEGL.so.1: cannot open shared object file" even
though nothing ever opens a window. Measured through the real find_package
path with the LINK_GROUP workaround composed: 67 DT_NEEDED without the
flag, 3 with it, TKV3d and TKOpenGl gone.

The flag is deliberately left open rather than closed with
-Wl,--no-as-needed. CMake emits the imported targets'
INTERFACE_LINK_LIBRARIES -- where OCCT lists libGL/libEGL -- after that
item, so closing the bracket switches the flag off immediately before the
libraries it exists to exclude. Verified against the shipped artifact:
closed, the kernel plug-in fell from 47 DT_NEEDED libTK entries to 14 and
lost TKV3d, yet still carried a direct libEGL.so.1 and still failed to
import on a headless server; open, the same 14 remain and libGL/libEGL are
gone. None of the 14 retained modules depends on GL.

Put the shared OCCT on LD_LIBRARY_PATH for the build itself. Nothing else
points at it -- IfcOpenShell's libraries get INSTALL_RPATH=$ORIGIN and OCCT
sits in its own dependency prefix -- so the post-build `import ifcopenshell`
check fails the same way. This is build-time only; the shipped packages get
libTK*.so* staged beside the payload with an $ORIGIN RUNPATH instead.

Suffix OCCT's install directory with `-shared` when it applies. Static and
shared installs are not interchangeable, but `build_dependency` skips any
dependency whose install dir already exists and cache_dependencies.py keys
its tarballs purely on that directory name -- so the static
`cache-occt-7.8.1.tar.gz` restored from the build-outputs repo silently
satisfied the build and BUILD_LIBRARY_TYPE was never applied. This is the
same cache stickiness 8bdaa8c7c described, pointing the other way. The
suffix makes the key configuration-aware, so it self-invalidates and the
static tarball stays valid for builds that still want static.

Packaging is the other half, and is why 8bdaa8c7c backed the flag out --
`stage_runtime_payload` only copies from install/ifcopenshell, so OCCT in
install/occt-* was never staged and `--shared` "worked by accident" off
cached static outputs. Stage libTK*.so* alongside, then give every staged
library an $ORIGIN RUNPATH: the core libs currently carry dead
build-machine RPATHs and the plug-ins carry none, resolving only because
the Python wrapper pulls them in by SONAME first. Shared OCCT has no such
first loader, since it is reached through the dlopen'd plug-ins.

The packages shrink: the python zip goes from 109.4 MB to 85.6 MB, because
the duplicated OCCT was 113 MB of the 287 MB unpacked payload (the eight
geometry_writer_ifc* plug-ins alone were 4.6 MB each) against ~67 MB for
one shared copy. Same argument as a91b1da28 ("Reduce Rocky package size")
and 402591e71.

macOS and Windows are affected too but are not fixed here. Their
packaging resolves via @loader_path install names and would need
install_name_tool rewriting, which cannot be verified from Linux; adding
the flag without that would ship a package that fails to load.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Dion Moult
2026-08-21 15:50:21 +10:00
parent 2c1d445d5b
commit 7845f8e8cd
4 changed files with 144 additions and 6 deletions
+35
View File
@@ -54,6 +54,41 @@ if(NOT OCC_INCLUDE_DIR AND NOT OCC_LIBRARY_DIR)
set(OpenCASCADE_LIBRARIES "$<LINK_GROUP:RESCAN,${OpenCASCADE_LIBRARIES}>")
endif()
if(UNIX AND NOT APPLE)
# Record only the OCCT modules whose symbols are actually referenced.
#
# OpenCASCADE_LIBRARIES is OCCT's *complete* module list, Visualization
# included -- TKV3d, TKOpenGl, TKService. Against a static OCCT that
# costs nothing, because an unreferenced module contributes no objects.
# Against a shared OCCT every entry becomes a hard DT_NEEDED, and
# libTKV3d pulls libGL.so.1 + libEGL.so.1, so `import ifcopenshell`
# dies on any headless machine with
# ImportError: libEGL.so.1: cannot open shared object file
# even though IfcOpenShell never opens a window. Measured on a full
# 67-module link: 67 DT_NEEDED entries without the flag, 3 with it,
# and TKV3d/TKOpenGl gone in the second case.
#
# Goes before the LINK_GROUP above rather than inside it, and is a
# no-op for a static OCCT (--as-needed only governs shared libraries),
# so this is safe for both link types.
#
# Deliberately NOT closed with -Wl,--no-as-needed. CMake emits the
# imported targets' INTERFACE_LINK_LIBRARIES -- which is where OCCT
# lists libGL/libEGL -- *after* this item, so a closing bracket
# switches the flag back off just before the libraries it was added to
# exclude. Measured: with the bracket closed the plug-ins dropped from
# 47 DT_NEEDED libTK entries to 14 and lost TKV3d, yet still carried a
# direct libEGL.so.1 and failed to import on a headless server; with it
# left open the same 14 remain and libGL/libEGL are gone.
#
# The cost is that --as-needed stays in effect for whatever follows on
# the link line. That is the default on Debian/Ubuntu so it is
# well-trodden, but it does mean a library needed only for
# static-initialiser side effects could be dropped -- the thing to
# suspect first if a serialiser stops registering itself.
set(OpenCASCADE_LIBRARIES "-Wl,--as-needed" "${OpenCASCADE_LIBRARIES}")
endif()
if(OpenCASCADE_VERSION VERSION_LESS "7.9.0" AND WIN32)
# Bug in OCCT cmake configs < 7.9.0 - missing linked library.
list(APPEND OpenCASCADE_LIBRARIES WSOCK32.lib)