Proposed: testsuite segfaults on ARM64

[Alexandre Ghiti]

A recent rebuild of this package in proposed led to multiple failures on ARM/ARM64. The only differences between release (where it builds fine) and proposed are:

- llvm upgrade from 1:9.0.1-16.1ubuntu1 to 1:9.0.1-20
- glibc upgrade from 2.34-0ubuntu2 to 2.34-0ubuntu3

After debugging a test that segfaults, it seems to be an LLVM issue though, below the backtrace:

Thread 1 "test_shuffle" received signal SIGSEGV, Segmentation fault.

0x0000fffff296d868 in llvm::StringMapImpl::FindKey(llvm::StringRef) const () from /usr/lib/llvm-9/lib/libLLVM-9.so.1
(gdb) bt
#0 0x0000fffff296d868 in llvm::StringMapImpl::FindKey(llvm::StringRef) const
() from /usr/lib/llvm-9/lib/libLLVM-9.so.1
#1 0x0000fffff2acc148 in llvm::Module::getNamedMetadata(llvm::Twine const&) const () from /usr/lib/llvm-9/lib/libLLV
M-9.so.1
#2 0x0000fffff2acc3e0 in llvm::Module::getModuleFlagsMetadata(llvm::SmallVectorImplllvm::Module::ModuleFlagEntry&)
const ()
from /usr/lib/llvm-9/lib/libLLVM-9.so.1
#3 0x0000fffff2acc584 in llvm::Module::getModuleFlag(llvm::StringRef) const ()
from /usr/lib/llvm-9/lib/libLLVM-9.so.1
#4 0x0000fffff7ac5600 in pocl_llvm_build_program (
program=program@entry=0xaaaaaaad92d0, device_i=device_i@entry=0,
user_options_cstr=<optimized out>,
program_bc_path=program_bc_path@entry=0xfffffffff038 "_global const char16 *in2, __global const uchar16 *mask, __
global char16 out) {\nout = shuffle2( *in1, *in2, *mask);\n}\n", num_input_headers=num_input_headers@entry=65535,
input_headers=input_headers@entry=0xffffffffee40,
header_include_names=header_include_names@entry=0x0,
linking_program=linking_program@entry=-135488132)
at /usr/lib/llvm-9/include/llvm/ADT/StringRef.h:82
#5 0x0000fffff7a52864 in compile_and_link_program (
compile_program=compile_program@entry=1,
link_program=link_program@entry=1, program=<optimized out>,
num_devices=<optimized out>, device_list=<optimized out>,
--Type <RET> for more, q to quit, c to continue without paging--
options=<optimized out>, num_input_headers=num_input_headers@entry=0,
input_headers=input_headers@entry=0x0,
header_include_names=<optimized out>, header_include_names@entry=0x0,
num_input_programs=<optimized out>, num_input_programs@entry=0,
input_programs=<optimized out>, input_programs@entry=0x0,
pfn_notify=<optimized out>, user_data=<optimized out>)
at ./lib/CL/pocl_build.c:651
#6 0x0000fffff7a51c50 in POclBuildProgram (program=<optimized out>,
num_devices=<optimized out>, device_list=<optimized out>,
options=<optimized out>, pfn_notify=<optimized out>,
user_data=<optimized out>) at ./lib/CL/clBuildProgram.c:37
#7 0x0000aaaaaaaa3810 in TestShuffle<signed char, unsigned char>::run (
this=this@entry=0xfffffffff480) at ./tests/kernel/test_shuffle.cc:303
#8 0x0000aaaaaaaa21ec in main (argc=<optimized out>, argv=0xfffffffff9e8)
at ./tests/kernel/test_shuffle.cc:360

[Alexandre Ghiti (alexghiti)]

I downgraded llvm to 1:9.0.1-16.1ubuntu1 and the test passes without any segfault.

[Alexandre Ghiti (alexghiti)]

Pocl 1.8-1 is still blocked in proposed because it also fails to build (on arm64 at least) when upgrading from 1:9.0.1-16.1ubuntu1 to 1:9.0.1-20ubuntu2.

[Alexandre Ghiti (alexghiti)]

Pocl 1.8-3 depends on llvm-toolchain-11, so this bug is not relevant anymore.