This is the newest version currently available in the Ubuntu 24.04
repos.
This fixes building; previously it failed to build libc++, which only
supports being built with one of the two preceding versions of Clang.
This reverts commit 1d8c107d49.
After 49e14021be, when using F(-NAN)
instead of -F(NAN) to hide the negation, this test passes correctly.
I hadn't realized that GCC and Clang optimized cos(-x) = cos(x), when
I made 1d8c107d49 (and at the time, the
F() macro didn't work properly with -NAN on all toolchains).
In the former form, the negation of NAN is hidden inside the
optimization blocking macro/wrapper. After
9be66e9d22, F(-NAN) works just as
well as -F(NAN).
This avoids test issues with recent optimizations in LLVM.
https://github.com/llvm/llvm-project/pull/193586 seems to make
the compiler consider known symmetry for hyperbolic math functions,
exploiting the fact that cosh(-x) is equal to cosh(x) (for values
other than NANs). By hiding the negation inside the F() wrapper,
the compiler can't reason about is existence.
checkout@v5 should be the same as checkout@v4, just running on
Node 24. checkout@v6 has got a few other minor changes that
shouldn't make any difference; opt in to newer behaviours now.
This fixes importing the _ctypes module, if a shared libffi doesn't
happen to exist on the host system with the right so version.
Since Python 3.12, such a build error ends up failing the build,
while it previously passed silently even if some parts of the build
had failed.
This makes sure the openmp tests actually are executed; since
switching to the runtimes build of openmp for the tests (in
98684ab807), the tests weren't
executed, as it no longer found the tools necessary for executing
the tests - as they are expected to be specified differently
in the runtimes build mode.
This requires the commit 48a5119d8e7d7236a28c14d06ec215ef3366ef90
(backported to 22.x in 9794a7932465bdc244b327388905d66df38485e2,
part of the 22.1.1 release) in llvm-project, to be able to run
tests without actually building clang in the same build as the
test tools.
Since 6295b8e763503644961f9087382f80965b98d466 in llvm-project,
one no longer can configure the openmp project entirely standalone,
one has to do it through the runtimes directory.
However, in practice, this currently leads to no tests being
executed at all, as the runtimes build setup doesn't support
pointing to an external FileCheck executable - it currently has
to be part of a full bootstrapping build.
Since 6295b8e763503644961f9087382f80965b98d466 in llvm-project,
it's no longer possible to build openmp entirely standalone;
now one has to build it through the runtimes directory.
This version of mingw-w64 includes the function "mkdtemp", which
breaks building the previously pinned version of busybox-w32 on i686.
Update both projects in sync, so that builds with newer nightly
versions of mingw-w64 also succeed.
Some notable changes since LLVM 21.x related to mingw:
- Handle an uncommon error code in libcxx's std::filesystem,
avoiding spurious test failures if running the libcxx test suite
(noted at least occasionally on Github Actions)
- Reinstate x86 stack probe functions in compiler-rt, that we earlier
beleived weren't used - but turned out to be used by GCC in i386
mode
- Fix interpretation of a few corner case aarch64 SEH unwind opcodes
- Support for FAT LTO (compile with -ffat-lto-objects, link with
-Wl,--fat-lto-objects)
- llvm-objcopy: Support the -p option for COFF targets
- llvm-readobj: A new option --coff-pseudoreloc for dumping information
and symbols for mingw runtime pseudo relocations
- Lots of various fixes for the cygwin target
- Improve linking a custom CRT with binutils ld (by passing the custom
CRT -l option at the same place as the original default -lmsvcrt)
- Make lld use the right mingw specific DLL entry point implicitly,
improving some Rust use cases
- Fix running the libcxx tests with optimization for mingw targets,
allowing more test coverage of the code generation paths
The built-in make is a POSIX make tool, which doesn't support
GNU make features.
As we do have GNU make, insert that as "make.exe" in the busybox
directory. Don't do that in the main llvm-mingw bin directory,
which can be used with other environments (e.g. msys2), where the
native make is to be preferred.
This allows running the built busybox on older versions of
Windows that don't support the utf8 manifest. (The utf8 manifest
is supported since Windows 10 update 1903.)
When building with mingw32-make, it prefers executing commands
through "sh.exe" if it is available.
When busybox sh.exe executes "ar", it prefers executing its own
internal tool over a potential external tool found in $PATH.
(This can be avoided by setting 'export BB_OVERRIDE_APPLETS="ar"',
but as this build of busybox is expected to be used with llvm-mingw,
disable the built-in "ar" tool outright.)
According to https://github.com/mstorsjo/llvm-mingw/issues/533,
if OCaml support is detected and the LLVM OCaml bindings are built,
they are attempted to be installed in a path that require root access.
This version fixes catching pointer data types by reference in C++.
This version also fixes throwing C++ exceptions on aarch64 with
a debugger attached to the process.
Finally, this version fixes delayimport calling of C++ methods that
return structs on aarch64.
These builds fail regularly, under mysterious circumstances.
("The hosted runner lost communication with the server. Anything
in your workflow that terminates the runner process, starves it
for CPU/Memory, or blocks its network access can cause this error.")
This usually seems to happen around the build-libcxx.sh step.
This allows overriding which linker to use.
Currently, building LLVM on Ubuntu 25.10 with these scripts fails,
if LLD is available. This, because the GCC on that distribution
produces object files that LLD (and gold) rejects during linking - see
https://github.com/llvm/llvm-project/issues/156667 for details.
By building with --use-linker=bfd, linking does succeed in that
configuration.
Also pass the same option through from build-all.sh.
In this case it's usually an option that wasn't recognized
(although the misrecognized option may have been an earlier
option and the current one may be the intended prefix).
These variable expansions/modifications work the same if the
option "--pgo" is specified - so either we can simplify the
stripping of the "--pgo=" prefix, or we can just skip the check,
similar to the case above with --instrumented.
For some unknown reason, macos-15 runners (which macos-latest
resolves to currently) seem to be consistently 25-40% slower than
macos-14 runners; the "macos" job (doing an universal binary build
for 2 architectures, with LTO) often take over 5 hours, and have
reached the 6 hour timeout once.
Switch this back to macos-14 to speed it up and reduce the risk of
hitting the timeout.
In the future, if/when the macos-14 runners are phased out, and if
the performance reduction in macos-15 runners remains, we'd need to
switch to building each architecture in a separate job, and merge
them together separately if we still want to provide universal
binaries.
If ARCHS itself was set by the caller, it would already have been
implicitly passed to the makefile - but if the caller set
TOOLCHAIN_ARCHS instead, that wasn't acted upon by the makefile.
As the shell script already has the logic for initializing this
(in the same fashion as all other shell scripts), just use that
and pass the value on to the makefile.
This matches what is done in build-all.sh for the PGO step; this
removes rather large (~50 MB) libraries that mostly are needed
for compiling for the host unix platforms. (On Linux it's needed
if building with profiling, on macOS it's needed for the builtin
functions for platform version availability too.)
Martin Storsjö