IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
The ClingMMapper must remain available until all ClingMemoryManagers
are destructed, which is typically during shutdown of IncrementalJIT.
This was not the case for the global object MMapperInstance that was
introduced during the upgrade to LLVM 13 because libCling variables
are destructed before running TROOT atexit handlers that shut down
the JIT. In practice, it happened to work but this will change with
the upgrade to LLVM 18 where we see crashes in some tests, potentially
because of upstream commit
47f5c54f99
See also commits e0f6c04660 ("Prevent static destruction from ending
DefaultMMapper too early") and 80c14bb948 ("Extend lifetime of
SectionMemoryManager::DefaultMMapper, again") for the same problem
that we previously patched in our copy of LLVM.
Commit 785c9df34d ("Restore symbol lookup in child interpreters")
added a DefinitionGenerator to allow symbol lookup in the parent
IncrementalJIT after the upgrade to LLVM 13. It appears that instead
of the unmangled name, we need to lookup already linker mangled names.
This fixes the tests CodeUnloading/AtExit.C and
MultipleInterpreters/MultipleInterpreters.C on macOS, which adds an
underscore during linker mangling. No change on Linux because there
is no additional linker name mangling.
We are relying on this since a while, for example for reemission of
template symbols. At the moment, we get the incremental extensions
because Preprocessor::enableIncrementalProcessing() turns them on
internally, but this will change with LLVM 18 where this method only
controls incremental processing of a single Preprocessor object.
Upstream Clang keeps TemplateIdAnnotations around "if they might
still be in the token stream." See upstream commit for more details:
6163aa9679
(included in Clang 11, in ROOT since the upgrade to LLVM 13)
This reasoning doesn't apply when we fully reset the Parser state in
ParserStateRAII's destructor, and we expect the swapped out vector of
TemplateIdAnnotations to be empty in order to not leak.
Fixes#16121
This will allow us to simplify our ASTConsumer model instead of creating many
multiplexers that are not needed. That should simplify adoption of latest
versions of clad.
Upstream moved away from manually declaring `*def` and `*inc` files. These
are now auto-generated with tablegen. This patch does the same for cling,
making it easier to rebase and maintain.
As noted in #14593, the build fails during a rootcling execution when
setting CLING_DEBUG=1 in the environment with
error: invalid integral value '0 -fno-omit-frame-pointer' in '-O0 -fno-omit-frame-pointer'
Upon investigation, it only works in the ROOT prompt because TCling
turns on basic -O1 unless in rootcling. This overrides the (misformed)
"-O0 -fno-omit-frame-pointer". Split the argument in two entries and
move it after inserting the user-provided arguments to properly apply
also on the prompt by now taking precedence over -O1.
Closes#14593
Before the passmanager update, the function `populateModulePassManager`
was called instead of `buildPerModuleDefaultPipeline` to set the default
passes. But the former did not have `RelLookupTableConverterPass` turned on
by default.
It is recommended to use a new instance of analysis managers every time
we perform codegen as the analyses might not be cleared. Upstream does
not have any instance where populating pass/analysis managers and
running passes are separated.
According to https://github.com/llvm/llvm-project/commit/af27466c50,
it is used to represent types pulled in with a using declaration, for
example `using std::error_code; error_code x;`.
This fixes the build of the ROOTTMVASofie module on macOS where this
was causing `std::size_t` not be fully desugared to `unsigned long`
at first to later end up with two identical classes in a list that
must not have duplicates.
GenericLLVMIRPlatformSupport now declares atexit, which leads to a
duplicate symbol error during startup and fails the transaction.
Completely switch to the provided infrastructure and just override
the helpers to register the functions with our Interpreter.
In LLVM 16, we now get ElaboratedType far more often, but they don't
have all namespace qualifiers we need. This fixes the build of the
RooFitCore module, but many other places are likely affected as well.
adjustPassManager was only used for a few backends, such as AMDGPU
and NVPTX, but not on the CPU architectures we care about.
The DiscriminatorsPass was also removed upstream Clang, see commit
https://github.com/llvm/llvm-project/commit/2978d02681.
