Backport patch from https://github.com/llvm/llvm-project/pull/111831
to fix openmp affinity tests on some brew ppc runners.
Also add one more openmp test to the ignore list for s390x.
Also add --time-tests to the lit args -- I had one instance where
openmp tests were running for 4h on s390x. If this happends again,
this should help determine which tests take so much time.
A large number of openmp tests using tsan fail when we hit certain
machines on the rhel8-beefy channel in brew, because they appear
to the use 5-level page tables. This results in memory being
mapped in places where tsan does not expect it.
See https://github.com/llvm/llvm-project/issues/111492 for more
context.
Work around this by disabling the openmp tests that use tsan if
the cpu has the la57 feature.
We currently disable rpath during both build and install. Instead,
we use LD_LIBRARY_PATH to allow the built clang to find the
libLLVM.so etc objects.
However, this does not work well if the system clang and the
clang being built have the same version. During the build, we
use both the system clang and the just-built clang, and they
need to use the system and just-built shared objects respectively.
However, use of LD_LIBRARY_PATH causes us to always use the
just-built objects as long as the versions match.
This is a problem in two scenarios: When building compat packages
for the current system LLVM version, we mix system clang with
compat libraries, which assume different paths. And when building
release candidates, a build using a previous rc of a newer rc
may use ABI-incompatible objects, because we don't version sonames
on rc versions.
Fix this by keeping the rpath during the build and only stripping
it on installation using the CMAKE_SKIP_INSTALL_RPATH option.
For manually installed binaries, we need to also manually strip
the rpath using chrpath.
This way system clang will use system libraries, and just-built
clang will use just-built libraries.
Before we had a very long list of CMake arguments.
There was no room for annotation through comments. Now we have a new
global called `%cmake_config_args` that one can append to.
I've created these `#regions` to group the options for each sub-project:
```
#region clang options
#region compiler-rt options
#region docs options
#region lldb options
#region llvm options
#region openmp options
#region test options
#region misc options
```
This spurious failure has been observed both on aarch64 in fedora
rawhide koji and on ppc64le in c10s koji.
I've filed an upstream issue to track this here:
https://github.com/llvm/llvm-project/issues/111140
Each `#region XY` now has an `#endregion XY` instead of just
`#endregion`. This is optional and `XY` does not have to be repeated for
the "system" to work but it makes working with regions a bit more
convenient.
```
#region globals...
#region packages...
#region prep...
#region build...
#region install...
#region check...
#region misc...
#region files...
#region changelog...
```
The test is run by CI only in rawhide. Requiring it in
all fedora branches will cause that all bodhi updates get
stuck due to the required test being absent.
