[cfe-dev] [llvm-dev] 12.0.0-rc1 Release has been tagged

[Tom Stellard via cfe-dev]

On 2/13/21 1:44 PM, Louis Dionne wrote:
> 
> 
>> On Feb 13, 2021, at 12:04, Dimitry Andric <dimitry at andric.com> wrote:
>>
>> On 28 Jan 2021, at 05:05, Tom Stellard via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>>>
>>> I've tagged the 12.0.0-rc1 release.  Testers can begin testing and upload binaries.
>>
>> During 12.0.0-rc1 test builds on FreeBSD, I encountered this unittest failure, which turned out to be caused by recent libc++ changes:
>>
>>   FAILED: tools/clang/bindings/python/tests/CMakeFiles/check-clang-python
>>   cd /home/dim/llvm/12.0.0/llvm-project/clang/bindings/python && /usr/local/bin/cmake -E env CLANG_LIBRARY_PATH=/home/dim/obj/llvmorg-12.0.0-rc1-2-gdf959fc2eb7e-1/lib /usr/local/bin/python3.7 -m unittest discover
>>   EEEEEEEE
>>
>> The 'EEEEEEEE' signifies 8 errors, which become more clear if "unittest discover" is run with the -f (fail immediately) and -v (verbose) options:
>>
>>   $ cd /home/dim/llvm/12.0.0/llvm-project/clang/bindings/python && /usr/local/bin/cmake -E env CLANG_LIBRARY_PATH=/home/dim/obj/llvmorg-12.0.0-rc1-2-gdf959fc2eb7e-1/lib /usr/local/bin/python3.7 -m unittest discover -f -v
>>   test_access_specifiers (tests.cindex.test_access_specifiers.TestAccessSpecifiers)
>>   Ensure that C++ access specifiers are available on cursors ... ERROR
>>
>>   ======================================================================
>>   ERROR: test_access_specifiers (tests.cindex.test_access_specifiers.TestAccessSpecifiers)
>>   Ensure that C++ access specifiers are available on cursors
>>   ----------------------------------------------------------------------
>>   Traceback (most recent call last):
>>     File "/home/dim/llvm/12.0.0/llvm-project/clang/bindings/python/clang/cindex.py", line 4173, in get_cindex_library
>>       library = cdll.LoadLibrary(self.get_filename())
>>     File "/usr/local/lib/python3.7/ctypes/__init__.py", line 442, in LoadLibrary
>>       return self._dlltype(name)
>>     File "/usr/local/lib/python3.7/ctypes/__init__.py", line 364, in __init__
>>       self._handle = _dlopen(self._name, mode)
>>   OSError: /home/dim/obj/llvmorg-12.0.0-rc1-2-gdf959fc2eb7e-1/lib/../lib/libc++.so.1: Undefined symbol "_ZnamSt11align_val_t"
>>
>>   During handling of the above exception, another exception occurred:
>>
>>   Traceback (most recent call last):
>>     File "/home/dim/llvm/12.0.0/llvm-project/clang/bindings/python/tests/cindex/test_access_specifiers.py", line 29, in test_access_specifiers
>>       """, lang = 'cpp')
>>     File "/home/dim/llvm/12.0.0/llvm-project/clang/bindings/python/tests/cindex/util.py", line 41, in get_tu
>>       source)])
>>     File "/home/dim/llvm/12.0.0/llvm-project/clang/bindings/python/clang/cindex.py", line 2812, in from_source
>>       index = Index.create()
>>     File "/home/dim/llvm/12.0.0/llvm-project/clang/bindings/python/clang/cindex.py", line 2699, in create
>>       return Index(conf.lib.clang_createIndex(excludeDecls, 0))
>>     File "/home/dim/llvm/12.0.0/llvm-project/clang/bindings/python/clang/cindex.py", line 212, in __get__
>>       value = self.wrapped(instance)
>>     File "/home/dim/llvm/12.0.0/llvm-project/clang/bindings/python/clang/cindex.py", line 4147, in lib
>>       lib = self.get_cindex_library()
>>     File "/home/dim/llvm/12.0.0/llvm-project/clang/bindings/python/clang/cindex.py", line 4178, in get_cindex_library
>>       raise LibclangError(msg)
>>   clang.cindex.LibclangError: /home/dim/obj/llvmorg-12.0.0-rc1-2-gdf959fc2eb7e-1/lib/../lib/libc++.so.1: Undefined symbol "_ZnamSt11align_val_t". To provide a path to libclang use Config.set_library_path() or Config.set_library_file().
>>
>>   ----------------------------------------------------------------------
>>   Ran 1 test in 0.027s
>>
>>   FAILED (errors=1)
>>
>> It turns out that libc++.so.1 now depends on the symbol '_ZnamSt11align_val_t', which is 'operator new [](unsigned long, std::align_val_t)'. Embarassingly, libc++.so.1 has always depended on this C++17 specific variant of operator new. But FreeBSD's C++ ABI library, libcxxrt, has never supported this variant!
>>
>> Previously, this never was a problem, because libc++ provided its own weak definitions of these operators. But recently the define _LIBCPP_DISABLE_NEW_DELETE_DEFINITIONS was toggled to on by default, which disables those definitions. This happened in 9b40ee8eb0c194f4b2787801ac6f9ef8fc1b8f46 ("[libc++] Define new/delete in libc++abi only by default") by Louis Dionne.
>>
>> While this is reasonable for the libc++abi situation, where it might lead to circular dependencies, I would like to turn the definitions on again in case libcxxrt is used for the C++ ABI, such as on FreeBSD.
>>
>> If you think this is a reasonable approach, I will create a review for it. Otherwise, I will have to patch the test-release.sh script to pass -DLIBCXX_ENABLE_NEW_DELETE_DEFINITIONS=ON to the CMake command line on FreeBSD.
> 
> The correct approach would be to have a CMake cache file under libcxx/cmake/cache/FreeBSD.cmake that sets the right configuration for building libcxx as a system library on freebsd. That way, the cmake code itself can stay free of platform specific logic.
> 
> Also note that I remember asking vendors before making that change, but I guess I either forgot to ask Freebsd or you told me OK without thinking about this. I can’t verify which one happened because I’m on my phone, but I’ll check it out to make sure I have your contact information if I need to ask that sort of questions in the future. It could be useful to have some sort of libcxx-vendors group in Phabricator that I could tag whenever I want to ping you folks.
> 
> Please send a Phab and I’ll revise it quickly (probably on Monday morning).
> 

Has this been fixed in the branch yet?

-Tom

> Louis
> 
>>
>> -Dimitry
>>
>