Bug#848368: llvm-toolchain-3.9: Please add ELF symbols versions to the libraries

Rebecca N. Palmer rebecca_palmer at zoho.com
Tue Jan 17 22:21:51 UTC 2017 

Control: tags -1 patch

One way to reproduce this bug:
# apt-get install pocl-opencl-icd blender
$ gdb blender
open User Preferences -> System
This promptly crashes; beignet-opencl-icd 1.2.1-1 (LLVM 3.8, like pocl) 
also triggers it, but beignet-opencl-icd 1.2.1-2 and mesa-opencl-icd 
(LLVM 3.9, like the graphics part of mesa) don't.

#0  llvm::cl::Option::setArgStr(llvm::StringRef) ()
     at /build/llvm-toolchain-3.9-3.9.1/include/llvm/ADT/SmallPtrSet.h:224
No locals.
#1  0x00007fffbbe9aacb in llvm::cl::opt<(anonymous 
namespace)::HelpPrinter, true, llvm::cl::parser<bool> >::opt<char [17], 
llvm::cl::desc, llvm::cl::LocationClass<(anonymous 
namespace)::HelpPrinter>, llvm::cl::OptionHidden, 
llvm::cl::ValueExpected, llvm::cl::cat>(char const (&) [17], 
llvm::cl::desc const&, llvm::cl::LocationClass<(anonymous 
namespace)::HelpPrinter> const&, llvm::cl::OptionHidden const&, 
llvm::cl::ValueExpected const&, llvm::cl::cat const&) [clone 
.constprop.300] ()
     at 
/build/llvm-toolchain-3.8-3.8.1/include/llvm/Support/CommandLine.h:1041
No locals.
#2  0x00007fffbbe9ad08 in _GLOBAL__sub_I_CommandLine.cpp ()
     at /build/llvm-toolchain-3.8-3.8.1/lib/Support/CommandLine.cpp:1661
No locals.
#3  0x00007ffff7de95da in ?? () from /lib64/ld-linux-x86-64.so.2
No symbol table info available.
#4  0x00007ffff7de96eb in ?? () from /lib64/ld-linux-x86-64.so.2
No symbol table info available.
#5  0x00007ffff7dedc68 in ?? () from /lib64/ld-linux-x86-64.so.2
No symbol table info available.
#6  0x00007ffff7de9484 in ?? () from /lib64/ld-linux-x86-64.so.2
No symbol table info available.
#7  0x00007ffff7ded419 in ?? () from /lib64/ld-linux-x86-64.so.2
No symbol table info available.
#8  0x00007fffee0f0ee9 in ?? () from /lib/x86_64-linux-gnu/libdl.so.2
No symbol table info available.
#9  0x00007ffff7de9484 in ?? () from /lib64/ld-linux-x86-64.so.2
No symbol table info available.
#10 0x00007fffee0f1521 in ?? () from /lib/x86_64-linux-gnu/libdl.so.2
No symbol table info available.
#11 0x00007fffee0f0f82 in dlopen () from /lib/x86_64-linux-gnu/libdl.so.2
No symbol table info available.
#12 0x00007fffbf9f9212 in ?? () from 
/usr/lib/x86_64-linux-gnu/libOpenCL.so.1
No symbol table info available.
#13 0x00007fffbf9f9360 in ?? () from 
/usr/lib/x86_64-linux-gnu/libOpenCL.so.1
No symbol table info available.
#14 0x00007fffbf9f98d0 in ?? () from 
/usr/lib/x86_64-linux-gnu/libOpenCL.so.1
No symbol table info available.
#15 0x00007fffbf9fa0d3 in clGetPlatformIDs ()
    from /usr/lib/x86_64-linux-gnu/libOpenCL.so.1
No symbol table info available.

The relevant lines of the LLVM build log appear to be

Scanning dependencies of target LLVM
make[4]: Leaving directory '/«PKGBUILDDIR»/build-llvm'
/usr/bin/make -f tools/llvm-shlib/CMakeFiles/LLVM.dir/build.make 
tools/llvm-shlib/CMakeFiles/LLVM.dir/build
make[4]: Entering directory '/«PKGBUILDDIR»/build-llvm'
[ 76%] Building CXX object 
tools/llvm-shlib/CMakeFiles/LLVM.dir/libllvm.cpp.o
cd /«PKGBUILDDIR»/build-llvm/tools/llvm-shlib && /usr/bin/g++-6 
-D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS 
-D__STDC_LIMIT_MACROS -I/«PKGBUILDDIR»/build-llvm/tools/llvm-shlib 
-I/«PKGBUILDDIR»/tools/llvm-shlib -I/«PKGBUILDDIR»/build-llvm/include 
-I/«PKGBUILDDIR»/include  -std=c++0x -gsplit-dwarf -Wl,-fuse-ld=gold 
-fPIC -fvisibility-inlines-hidden -Wall -W -Wno-unused-parameter 
-Wwrite-strings -Wcast-qual -Wno-missing-field-initializers -pedantic 
-Wno-long-long -Wno-maybe-uninitialized -Wdelete-non-virtual-dtor 
-Wno-comment -Werror=date-time -std=c++11 -ffunction-sections 
-fdata-sections -O2 -g -DNDEBUG -fPIC    -fno-exceptions -o 
CMakeFiles/LLVM.dir/libllvm.cpp.o -c 
/«PKGBUILDDIR»/tools/llvm-shlib/libllvm.cpp
[ 76%] Linking CXX shared library ../../lib/libLLVM-3.9.so
cd /«PKGBUILDDIR»/build-llvm/tools/llvm-shlib && /usr/bin/cmake -E 
cmake_link_script CMakeFiles/LLVM.dir/link.txt --verbose=1
/usr/bin/g++-6  -fPIC -std=c++0x -gsplit-dwarf -Wl,-fuse-ld=gold -fPIC 
-fvisibility-inlines-hidden -Wall -W -Wno-unused-parameter 
-Wwrite-strings -Wcast-qual -Wno-missing-field-initializers -pedantic 
-Wno-long-long -Wno-maybe-uninitialized -Wdelete-non-virtual-dtor 
-Wno-comment -Werror=date-time -std=c++11 -ffunction-sections 
-fdata-sections -O2 -g -DNDEBUG  -Wl,-O3 -Wl,--gc-sections -Wl,-z,relro 
-Wl,-z,defs -shared -Wl,-soname,libLLVM-3.9.so.1 -o 
../../lib/libLLVM-3.9.so.1 CMakeFiles/LLVM.dir/libllvm.cpp.o 
-Wl,-rpath,"\$ORIGIN/../lib" -Wl,--whole-archive [big list of 
../../lib/libLLVM(something).a libraries] -lrt -ldl -ltinfo -lpthread 
-lz -lm
cd /«PKGBUILDDIR»/build-llvm/tools/llvm-shlib && /usr/bin/cmake -E 
cmake_symlink_library ../../lib/libLLVM-3.9.so.1 
../../lib/libLLVM-3.9.so.1 ../../lib/libLLVM-3.9.so
make[4]: Leaving directory '/«PKGBUILDDIR»/build-llvm'
[ 76%] Built target LLVM

i.e. the main libLLVM shared library doesn't use a version script at 
all.  (Some of its other libraries do have "version scripts", but these 
are used for their other function of limiting which symbols are public, 
and don't have versions.)

This suggests the fix (warning: untested and not my area of expertise - 
and if it is using that line, why is there no -Wl,--no-whole-archive in 
the build log?):

--- a/tools/llvm-shlib/CMakeLists.txt
+++ b/tools/llvm-shlib/CMakeLists.txt
@@ -42,7 +42,7 @@
  list(REMOVE_DUPLICATES LIB_NAMES)
  if("${CMAKE_SYSTEM_NAME}" STREQUAL "Linux" OR "${CMAKE_SYSTEM_NAME}" 
STREQUAL "GNU" OR "${CMAKE_SYSTEM_NAME}" STREQUAL "kFreeBSD") # FIXME: 
It should be "GNU ld for elf"
    # GNU ld doesn't resolve symbols in the version script.
-  set(LIB_NAMES -Wl,--whole-archive ${LIB_NAMES} -Wl,--no-whole-archive)
+  set(LIB_NAMES 
-Wl,--version-script,../../../tools/llvm-shlib/simple_version_script.map 
-Wl,--whole-archive ${LIB_NAMES} -Wl,--no-whole-archive)
  elseif("${CMAKE_SYSTEM_NAME}" STREQUAL "Darwin")
    set(LIB_NAMES -Wl,-all_load ${LIB_NAMES})
  endif()
--- a/dev/null
+++ b/simple_version_script.map
@@ -0,0 +1,1 @@
+LLVM_3.9 { global: *; };

(Should also work with the obvious change for 3.8; I haven't checked 
3.7.  Deliberately not making 3.9 "depend" on 3.8, as the whole point is 
to make the linker treat them as separate libraries)

Some LLVM-using libraries use -Bsymbolic to avoid similar problems (e.g. 
#768185), but I don't know whether enabling that on LLVM itself would 
help: it may well be papering over a problem that symbol versioning 
would really solve.

In case my patch doesn't work, should we (the OpenCL team, and 
potentially other LLVM rdeps) treat this bug's RC status as "this will 
be fixed, however long it takes", or as "fix this reasonably soon or all 
but one of the versions gets removed"?  Moving pocl-opencl-icd to LLVM 
3.9 probably requires upgrading it to an upstream git snapshot (Fedora 
did), so we need to decide soon whether to do so.

More information about the Pkg-llvm-team mailing list