[IMPALA-11102] LLVM codegen can't resolve boost symbols after bumping boost version to 1.74

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: Backend
• Epic Color: ghx-label-8
• Assignee: Unassigned
• Reporter: Quanlong Huang
• Created: 02/Feb/22 01:34
• Updated: 07/Feb/22 00:05

Description

We currently don't have any codegen functions depending on boost non-header-only libraries. This is found in a WIP patch: https://gerrit.cloudera.org/c/17785/3 . Patch set 3 depends on boost::locale to implement UTF8 support for upper() and lower() functions. It crashs when the UTF8 implementation is codegen:

F0128 15:17:45.989331 13575 llvm-codegen.cc:152] 64458eac382abfb9:85324b2c00000000] LLVM hit fatal error: Program used external function '_ZN5boost6locale9converterIcE2idE' which could not be resolved!
*** Check failure stack trace: ***
    @          0x56d70ec  google::LogMessage::Fail()
    @          0x56d899c  google::LogMessage::SendToLog()
    @          0x56d6a4a  google::LogMessage::Flush()
    @          0x56da608  google::LogMessageFatal::~LogMessageFatal()
    @          0x2ad61a5  impala::LlvmCodegenHandleError()
    @          0x55ddccd  llvm::report_fatal_error()
    @          0x52fdba4  llvm::RuntimeDyldImpl::resolveExternalSymbols()
    @          0x52fdcf4  llvm::RuntimeDyldImpl::resolveRelocations()
    @          0x50c1358  llvm::MCJIT::finalizeLoadedModules()
    @          0x50c1a31  llvm::MCJIT::finalizeObject()
    @          0x2ae0b0a  impala::LlvmCodeGen::FinalizeModule()
    @          0x24cd76b  impala::FragmentState::CodegenHelper()
    @          0x24cceb4  impala::FragmentState::InvokeCodegen()
    @          0x24d620e  impala::FragmentInstanceState::Open()
    @          0x24d2a74  impala::FragmentInstanceState::Exec()
    @          0x240f291  impala::QueryState::ExecFInstance()
    @          0x240d68f  _ZZN6impala10QueryState15StartFInstancesEvENKUlvE_clEv
    @          0x241231b  _ZN5boost6detail8function26void_function_obj_invoker0IZN6impala10QueryState15StartFInstancesEvEUlvE_vE6invokeERNS1_15function_bufferE
    @          0x22af13b  boost::function0<>::operator()()
    @          0x2a79976  impala::Thread::SuperviseThread()
    @          0x2a822c6  boost::_bi::list5<>::operator()<>()
    @          0x2a821ea  boost::_bi::bind_t<>::operator()()
    @          0x2a821ab  boost::detail::thread_data<>::run()
    @          0x4380a90  thread_proxy
    @     0x7f552827e6b9  start_thread
    @     0x7f5524cb951c  clone

The previous boost version is 1.61 which doesn't have this issue. We might encounter the same issue if any codegened function depends on boost libraries.

One difference between boost 1.61 and 1.74 is that symbols are now compiled with hidden visibility by default:

$ readelf -s --wide toolchain/toolchain-packages-gcc7.5.0/boost-1.74.0-p1/lib/libboost_locale.a | grep _ZN5boost6locale9converterIcE2idE
    25: 0000000000000028     8 OBJECT  GLOBAL HIDDEN     4 _ZN5boost6locale9converterIcE2idE
   142: 0000000000000000     0 NOTYPE  GLOBAL DEFAULT  UND _ZN5boost6locale9converterIcE2idE
   140: 0000000000000000     0 NOTYPE  GLOBAL DEFAULT  UND _ZN5boost6locale9converterIcE2idE
$ readelf -s --wide toolchain/toolchain-packages-gcc7.5.0/boost-1.61.0-p2/lib/libboost_locale.a | grep _ZN5boost6locale9converterIcE2idE
    25: 0000000000000028     8 OBJECT  GLOBAL DEFAULT    4 _ZN5boost6locale9converterIcE2idE
   141: 0000000000000000     0 NOTYPE  GLOBAL DEFAULT  UND _ZN5boost6locale9converterIcE2idE
   140: 0000000000000000     0 NOTYPE  GLOBAL DEFAULT  UND _ZN5boost6locale9converterIcE2idE

In boost-1.74, the visibility is HIDDEN. In boost-1.61, it's DEFAULT (global).

Our current LLVM version is 5.0.1. We might need to investigate if it's a LLVM issue and fixed in higher versions, or compile boost with GLOBAL visibility for symbols.

Issue Links

• is caused by: IMPALA-11005 Upgrade Boost library for Impala