We get:
/usr/bin/ld: cannot find -lomptarget.rtl.host: No such file or directory
While trying to link libomptarget.so.19.1
NextGen host plugin is not enabled on riscv64:
01cbbc52dc/offload/plugins-nextgen/host/CMakeLists.txt
-- Not building AMDGPU plugin: only support AMDGPU in Linux x86_64, ppc64le, or aarch64 hosts
-- Building the offload library with support for the "cuda;host" plugins
-- OMPT target enabled
-- OpenMP tools dir in libomptarget: /builddir/build/BUILD/llvm-19.1.0-build/llvm-project-19.1.0.src/llvm/redhat-linux-build/runtimes/runtimes-bins/openmp/runtime/src
-- Building CUDA plugin for dlopened libcuda
-- Not generating NVIDIA tests, no supported devices detected. Use 'LIBOMPTARGET_FORCE_NVIDIA_TESTS' to override.
-- Not building NextGen offloading plugin
-- Building DeviceRTL. Using clang: /builddir/build/BUILD/llvm-19.1.0-build/llvm-project-19.1.0.src/llvm/redhat-linux-build/bin/clang, llvm-link: /builddir/build/BUILD/llvm-19.1.0-build/llvm-project-19.1.0.src/llvm/redhat-linux-build/bin/llvm-link and opt: /builddir/build/BUILD/llvm-19.1.0-build/llvm-project-19.1.0.src/llvm/redhat-linux-build/bin/opt
-- Building the llvm-omp-device-info tool
-- Building the llvm-omp-kernel-replay tool
-- Building offloading runtime library libomptarget.
Signed-off-by: David Abdurachmanov <davidlt@rivosinc.com>
To see whether this fixes the OOMs. This also drops the change on
i686 for other OSs, to see whether that works or not.
On s390x this was already not effective due to a mismatch with
the condition in the cmake invocation.
Also drop the unnecessary python version variables.
If there are multiple python versions installed, the wrong one may be
picked. PYTHON_EXECUTABLE tries to avoid that problem, but the correct
way to spell this nowadays is Python3_EXECUTABLE.
Remove version.spec.inc from git
================================
We no longer track `version.spec.inc` in git. This is useful for
development. If we build a snapshot, the file is required but if we
build a `release` build (see below), then we don't include it.
Make improvements
=================
We distinguish between `snapshot-` and `release-` builds now in the
`Makefile`. These old targets are mapped to their new counterparts and a
deprecation warning is shown:
* `setup` -> `snapshot-setup`
* `local-rpm` -> `snapshot-rpm`
* `local-srpm` -> `snapshot-srpm`
* `local-prep` -> `snapshot-prep`
* `local-clean` -> `snapshot-clean`
We also have these new Make targets that run the build process with a
release tarball instead:
* `release-setup`
* `release-rpm`
* `release-srpm`
* `release-prep`
* `release-clean`
The targets `local-list-check` and `local-tmt-vm` have been commented
out because I think they were not needed. I just kept them for future
reference of research done in those areas.
All builds described by the `Makefile` are still local and no `mock` is
involved. I find this the easiest to debug.
Toggle default bcond state for snapshot_build
=============================================
We want to be able to build non-snapshot releases with this repo. That's
why be default the build condition `snapshot_build` is off. So
only when you explicitly enable `--with=snapshot_build` or define
`--define "_with_snapshot_build 1"` you'll be able to build a snapshot.
If however the build happens in Copr with a namespace that begins with
`fedora-llvm-team/llvm-snapshots-`, then the build condition is ON by
default and there's no need to enable it explicitly with `--with=snapshot_build`.
Things related to release update
================================
* Add llvm-project tarball and signature to sources
* Fix: error: source 1001 defined multiple times
* Fix: error: source 1006 defined multiple times
* Fix missing newline
* Conditionally apply 0001-Always-build-shared-libs-for-LLD.patch
* Only enable offload runtime in snapshot mode
Misc.
=====
* Add missing prep dependency
Have a --with lto_build option, where LTO is enabled by default
for everything except i686 and riscv. Use that flag to condition
whether `-DLLVM_UNITTEST_LINK_FLAGS` is passed.
Additionally, make use of the new Fat LTO functionality in rawhide
to use `-fno-lto` instead of `-Wl,-plugin-opt=O0` to save more
time linking unit tests.
We were defining lto_cflags in three places. It looks like on
aarch64 nil won (LTO disabled), while on other architectures
`-flto=thin` won.
I think on rawhide, this means that we're currently shipping
bitcode in static archives, because the required
`-ffat-lto-objects` option from redhat-rpm-config is missing.
Resolve this by leaving the option at its default value --
if enabling it on aarch64 causes issues, we can disable it just
there again.
Also reenable the LLVM_UNITTEST_LINK_FLAGS option, which should
reduce the amount of time LTO takes.
This already happens implicitly because we don't build from a git
checkout. However, currently this also breaks the build due to
https://github.com/llvm/llvm-project/pull/88164. Avoid this by
explicitly disabling the option.
* %check: export LD_LIBRARY_PATH because runtimes use external cmake projects that wouldn't find libclang++.so
* %check: use LLVM's LIT_XFAIL to mark tests that are expected to fail and don't just delete them
* clang: bring over *.patch files
* clang: remove build requires tags from clang package that are already defined by main llvm package
* clang: remove duplicates found in llvm globals and use %{pkg_datadir} instead of %{install_datadir}
* clang: rename and use pkg_name -> pkg_name_clang
* clang: use %{pkg_name_clang} instead of %{name}
* compiler-rt and libomp: wrap in %if %{without compat_build}
* libomp: also build libomp for s390x (See https://github.com/llvm/llvm-project/pull/66081)
* libomp: exclude libomp-devel on s390x just like libomp itself
* libomp: introduce pkg_name_libomp and cleanup the globals
* libomp: remove gdb-plugin (NEED TO CONFIRM IF THIS IS REALLY TRUE)
* libomp: filter out flaky libomp test: ompt/teams/distribute_dispatch.c
* llvm: add BuildRequires: graphviz to llvm-doc
* llvm: use %{pkg_name_llvm} instead of %{name}
* make: added temporary Makefile so I can more easily run srpm or rpm builds locally
* make: Added local-tmt-vm target (not really useful yet)
* rpm: for non-RHEL add prefix like "Jan 05 16:17:06" to every log line This should help in finding out how long things take to build/install/...
* rpm: fully specify %files for top-level packages
* rpm: make all packages fully qualified
* rpm: obsolete llvm-snapshot-builder
* rpm: remove BuildRequires: python3-lit because we have it in-tree
* rpm: remove BuildRequires: tags for clang and llvm specifics as we're building in-tree
* rpm: rename %llvm_srcdir -> %srcdir_llvm
* rpm: rename %pkg_name -> %pkg_name_llvm
* rpm: use full qualified name in %description
* rpm: use region comments in spec file to group by package. This allows editors to fold text to get a better overview
* rpmlint: update rc file to reflect clang and compiler-rt (libomp has no rpmlint rc file)
* use python-lit from within tree
